Enterprise Risk Management in Finance 1137466286, 9781137466280

Citation preview

Enterprise Risk Management in Finance

Enterprise Risk Management in Finance Desheng Dash Wu Stockholm Business School, Stockholm University, Sweden RiskLab, University of Toronto, Canada and

David L. Olson College of Business Administration, University of Nebraska, USA

© Desheng Dash Wu and David L. Olson 2015 All rights reserved. No reproduction, copy or transmission of this publication may be made without written permission. No portion of this publication may be reproduced, copied or transmitted save with written permission or in accordance with the provisions of the Copyright, Designs and Patents Act 1988, or under the terms of any licence permitting limited copying issued by the Copyright Licensing Agency, Saffron House, 6–10 Kirby Street, London EC1N 8TS. Any person who does any unauthorized act in relation to this publication may be liable to criminal prosecution and civil claims for damages. The authors have asserted their rights to be identified as the authors of this work in accordance with the Copyright, Designs and Patents Act 1988. First published 2015 by PALGRAVE MACMILLAN Palgrave Macmillan in the UK is an imprint of Macmillan Publishers Limited, registered in England, company number 785998, of Houndmills, Basingstoke, Hampshire RG21 6XS. Palgrave Macmillan in the US is a division of St Martin’s Press LLC, 175 Fifth Avenue, New York, NY 10010. Palgrave Macmillan is the global academic imprint of the above companies and has companies and representatives throughout the world. Palgrave® and Macmillan® are registered trademarks in the United States, the United Kingdom, Europe and other countries. ISBN: 978–1–137–46628–0 This book is printed on paper suitable for recycling and made from fully managed and sustained forest sources. Logging, pulping and manufacturing processes are expected to conform to the environmental regulations of the country of origin. A catalogue record for this book is available from the British Library. Library of Congress Cataloging-in-Publication Data Wu, Desheng Dash. Enterprise risk management in finance / Desheng Dash Wu, David L. Olson. pages cm ISBN 978–1–137–46628–0 (hardback) 1. Risk management. 2. Financial risk management. I. Olson, David L. II. Title. HD61.W796 2015 332.1068’1—dc23

2014049736

Contents List of Figures

x

List of Tables

xii

Preface

xv xvii

Acknowledgements 1

Enterprise Risk Management Introduction Definition Accounting perspective The COSO framework Categories Activities Risk appetite ERM process Implementation issues Risk management modeling Book outline

1 1 2 3 3 4 4 6 7 7 9 9

2

Enron Risk management California electricity Accounting impact SOX Conclusions

11 11 12 13 13 14

3

Financial Risk Management Introduction Investment collars VaR Copulas Tranches Conclusions

15 15 17 17 20 21 22

4

The Real Estate Crash of 2008 Introduction Real estate in 2008 Mortgage system

23 23 24 26 v

vi

Contents

Northern Rock AIG Risk management 5

Financial Risk Forecast Using Machine Learning and Sentiment Analysis Introduction Information volume and volatility Information sentiment and volatility Volatility model and modified non-linear GARCH model Daily volatility model Using GARCH-based SVM to associate information sentiment with asset price volatility Sentiment analysis of financial news Using GARCH-based SVM to associate information sentiment and volatility Empirical results and analysis Trading volume volatility forecasting Volatility forecasting with sentiment analysis Conclusions

26 29 30 32 32 32 33 34 38 39 39 41 42 43 44 48

6

Online Stock Forum Sentiment Analysis Introduction Architectural design of GARCH-SVM based on sentiment index Sentiment analysis Data Methodology comparison Sentiment and stock price volatility Conclusions

49 49 49 50 51 53 54 55

7

DEA Risk Scoring Model of Internet Stocks Introduction Different methods of performance evaluation Multivariate statistical analysis Data envelopment analysis Analytic hierarchy process Fuzzy set theory Grey relation analysis Balanced scorecard Financial statement analysis Basics of data envelopment analysis The proposed approach Variable selection Empirical study

57 57 57 58 58 59 59 60 60 60 61 63 64 66

Contents vii

The DEA result Conclusions

66 70

8 Bank Credit Scoring Introduction Risk modeling Performance validation in credit rating Case study: credit scorecard validation Statistical results and discussion Population distributions and stability Conclusions Appendix A8.1: Informal Definitions

72 72 72 74 75 77 80 85 86

9 Credit Scoring using Multiobjective Data Mining Introduction TOPSIS for data mining Steps of the TOPSIS data mining method Dataset TOPSIS model over training data Model comparisons Simulation of model results Conclusions

87 87 87 88 90 91 93 96 97

10 Online Banking Efficiency and Risk Evaluation with Principal Component Analysis Introduction Data and variables Results and analysis PCA-DEA analysis Risk factors Conclusions

99 99 100 101 104 106 106

11 Economic Perspective The traditional economic view The human factor Reality Risk mitigation Risk tolerance Recent events Conclusions

108 108 111 111 114 114 114 116

12 British Petroleum Deepwater Horizon Introduction Deepwater horizon The Oil Pollution Act of 1990

118 118 119 120

viii

Contents

Recovery factors in Macondo Risk management factors Conclusions

120 121 122

13 Bank Efficiency Analysis Introduction Data envelopment analysis (DEA) Neural networks The energy we use Bank branch efficiency analysis Short-term efficiency prediction Conclusions

124 124 126 127 129 129 133 134

14 Catastrophe Bond and Risk Modeling Introduction Catastrophe risk instruments Loss model Demonstration of computation Parameter estimation Error analysis Conclusions

136 136 136 138 138 140 143 144

15 Bilevel Programming Merger Analysis in Banking Introduction A conceptual banking chain with constrained resources Mathematical model Merger evaluation A numerical example for incentive incompatibility Case study: banking chain illustration Post merger Managerial insights Conclusions

145 145 146 148 150 151 152 157 160 161

16 Sustainability and Risk in Globalization Enterprise sustainability Types of risk Contexts of sustainable risk Globalization Supply chain risk management Global business risks Conclusions

163 163 165 166 168 170 171 173

17 Risk from Natural Disasters Introduction Preparing for high-impact, low-probability events Be prepared Risks and emergencies

175 175 176 177 177

Contents ix

Technical tools Emergency management Emergency management support systems Conclusions

178 179 180 181

18 Pricing of Carbon Emission Exchange in the EU ETS Introduction Literature review Price movements Model, data and sample Analysis of EUA logreturns Time series test GARCH effect test Method selection Estimation and forecasting Conclusions

183 183 184 186 188 189 191 192 193 194 197

19 Volatility Forecasting of the Crude Oil Market Introduction Volatility models Historical volatility ARMA(R,M) ARMAX(R,M, b) ARCH(q) GARCH(p,q) EGARCH GJR(p,q) Regime-switching models Data Distribution analysis Results GARCH modeling Markov regime-switching modeling Conclusions

199 199 200 200 201 201 201 202 202 203 203 204 204 206 206 210 214

20 Confucius Three-stage Learning of Risk Management Introduction Self-cultivation Family regulation State harmonization Conclusions

215 215 216 217 218 219

Notes

221

References

238

Index

253

List of Figures 3.1 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 6.1 6.2 6.3 6.4 7.1 7.2 8.1 8.2 8.3 8.4 8.5 8.6 9.1 9.2 10.1 10.2 13.1 14.1 14.2 14.3 14.4

CVaR and VaR Flow chart and functional parts of our approach associating information volume and volatility Flow chart and functional parts of our approach to associate information sentiment and volatility Daily changing rates of the trading volumes of NASDAQ index Sentiment calculation process for the current keyword w Sliding time window learning and forecasting Price volatility forecast result for company MDT over all the time windows Price volatility forecast result for company WAG over all the time windows Price volatility trend forecast accuracies for all the time windows Conceptual modeling of sentiment for volatility forecast The Lexicon approach for sentiment classification Volume of reviews distributed by time of week Volume of reviews distributed by time of day Two-stage DEA model Proposed evaluation process Lorenz curve, January–June 1999 sample Lorenz curve, July–December 1999 sample Lorenz curve, January–June 2000 sample Performance comparison of three samples Cumulative population distribution on all applicants Interval population distribution on all applications PolyAnalyst decision tree See5 decision tree 3D plot of PCA analysis Plots of principal component loadings in different DEA models Backpropagation neural networks Insurance payoffs (million RMB) for the Wenchuan earthquake Frequency histogram of logarithm of earthquake loss with normal fit Historical vs. simulated data distribution of compound Poisson model Historical vs. simulated data of stochastic process x

18 33 35 38 40 41 47 47 47 50 51 52 53 65 65 78 79 79 79 85 86 93 94 103 105 128 137 141 142 143

List of Figures xi

15.1 Supply chain model of the banking process with constrained resources 18.1 EUA price movement in Phase I 18.2 EUA price movement in Phase II 18.3 Logreturns of EUA in Phase I 18.4 Logreturns of EUA in Phase II 18.5 Series correlation of EUA logreturns in Phases I and II 18.6 Residual, standard deviation and logreturns series in Phase I 18.7 Residual, standard deviation and logreturns series in Phase II 19.1 NYMEX crude oil daily price movements 19.2 NYMEX crude oil daily logreturn 19.3 Normal distribution vs. t-distribution 19.4 Innovation, standard deviation, return 19.5 Simulation and forecasting 19.6 Transitional probabilities in Markov regime-switching with GED 19.7 Returns of two regimes in historical time series 19.8 Price of two regimes in historical time series

147 187 188 189 190 192 195 196 204 205 205 207 209 213 213 214

List of Tables 1.1 1.2 2.1 4.1 4.2 4.3 4.4 4.5 5.1 5.2 5.3

COSO ERM cube Risk management responsibilities Sarbanes–Oxley act elements Real estate cycle Northern Rock events Northern Rock retail deposits Northern Rock holdings before and after run Key events for AIG The information volumes calculated from Google Finance A snippet of news entries for the companies ADCT, S and MRO The eight word sets we use in this chapter to calculate the keyword sentiment 5.4 Predicted values of the average forecast error and the volatility trend forecast accuracy ratio 5.5 Forecast results for 177 listed companies during the year 2007 6.1 Selecting 1-grams as features 6.2 Relative accuracies by sentiment 7.1 BCC-efficient scores on performance 7.2 BCC-efficient scores on the level of returns per unit of risk 7.3 Ranking of the BCC-efficient scores of total efficiency and investing risk 7.4 Ranking of the BCC-efficient scores of whole model 8.1 Balanced scorecard perspectives, goals, and measures 8.2 Model risk events in banking 8.3 Scorecard performance validation, January–June 1999 8.4 Scorecard performance validation, July–December 1999 8.5 Scorecard performance validation, January–June 2000 8.6 Summary of performance samples 8.7 Population stability, January–June 1999 8.8 Population stability, July–December 1999 8.9 Population Stability, January–June 2000 8.10 Total population stability index 9.1 Independent variables for Canadian banking data set 9.2 Standardized data regression 9.3 Coincidence matrix – PolyAnalyst decision tree 9.4 Coincidence matrix – See5 decision tree 9.5 Coincidence matrix – TOPSIS L1 model xii

4 8 14 25 28 29 29 30 34 36 39 44 45 51 54 67 68 69 70 74 75 76 76 77 78 81 82 83 84 91 92 94 95 95

List of Tables xiii

9.6 9.7 9.8 9.9 10.1 10.2 10.3 10.4 10.5 10.6 10.7 11.1 11.2 12.1 12.2 13.1 13.2 13.3 13.4 13.5 13.6 13.7 13.8 13.9 13.10 13.11 13.12 14.1 14.2 14.3 14.4 14.5 14.6 15.1 15.2 15.3 15.4 15.5 15.6 15.7

Coincidence matrix – TOPSIS L2 model Coincidence matrix – TOPSIS L∞ model Comparison of model results Simulation results Online banking DEA variables Online banking data DEA combinations and their efficiencies Maximum component loadings matrix in different models Integrated PCA-DEA score Variance explained in integrated PCA-DEA Multivariate linear regression analysis – DV bank revenue Realms of uncertainty Evolution of risk management BP risk factors Factors in recovery Summary statistics of data Estimated neural network parameters Number of branches corresponding to each efficiency interval Statistical results corresponding to each efficiency interval Efficiency score distribution Implication of slight efficiency improvement on branch costs Regression analysis for branch efficiency prediction using October data Number of branches in each efficiency interval DEA-NN3 results Regression analysis for short-term efficiency prediction Comparison of best-practice branches Comparison of DEA and DEA-NN to efficiency measurement Catastrophe loss models Chinese earthquake loss data, 1966–2008 Results of ADF testing Results of parameter estimation Results of K–S test Error analysis Input and output data for the 8 branches in the numerical example Raw data for 30 branches Profit efficiency values A comparison with existing DEA and SFA Correlations Statistics under both CRS and VRS assumptions Top ten promising mergers under UL game structure

95 95 96 96 101 101 102 103 104 105 106 109 110 119 120 129 130 131 131 132 132 133 133 133 134 134 135 138 139 140 141 143 144 151 153 154 156 157 158 158

xiv List of Tables

15.8 18.1 18.2 18.3 18.4 18.5 18.6 18.7 18.8 18.9 18.10 18.11 18.12 18.13 19.1 19.2 19.3 19.4 19.5 19.6 19.7 19.8 20.1

Top ten promising mergers under LL game structure Descriptive statistics of EUA logreturns in Phase I Descriptive statistics of EUA logreturns in Phase II The ADF test of EUA logreturns in Phase I The ADF test of EUA first-order difference in Phase I The ADF test of EUA logreturns in Phase II The ADF test of EUA first-order difference in Phase II The ARCH LM test for EUA logreturns in Phases I and II The Akaike info criterion (AIC) and Schwarz criterion (SC) for the estimated model in Phase I The Akaike info criterion (AIC) and Schwarz criterion (SC) for the estimated model in Phase II Estimation of EUA logreturns in Phase I EUA Logreturns forecasting in Phase I Estimation of EUA logreturns in Phase II EUA Logreturn forecasting in Phase II Statistics on the daily crude oil index changes, February 2006–July 2009 Daily crude oil index logreturn statistics, February 2006–July 2009 GARCH(1,1) estimation using the t-distribution Various GARCH modeling characteristics Markov regime-switching computation example Markov regime-switching using Hamilton’s (1989) model Markov regime-switching using t-distribution Markov regime-switching using GED Risk management links

159 190 190 191 191 191 192 193 193 193 194 195 196 197 204 205 206 208 211 211 212 212 219

Preface The importance of financial risk was revealed by the traumatic events of 2007 and 2008, when the global financial community experienced a real estate bubble collapse from which (at the time of writing) most of the world’s economies are still recovering. Human investment activity seems determined to create bubbles, despite our long history of suffering.1 Financial investment seems to be a never-ending game of greedy players seeking to take advantage of each other, which Adam Smith assured us would lead to an optimal economic system. It is interesting that we pass through periods of trying one system, usually persisting until we encounter failure, and then moving on to another. The United States went through a long stretch where regulation of financial institutions was considered paramount, beginning with the Great Depression of the 1930s. When relative prosperity was experienced, the 1980s saw a resurgence of deregulation, culminating in the Gramm–Rudman Act that dismantled much of the post-depression regulation in favor of a free-wheeling economic system. Some post-2008 theorists have found evidence that this deregulation went too far. It is notable that Canada, with an economy highly integrated with that of the United States (but with more consistent regulation), experienced none of the traumatic real estate issues that plagued the US. We do not pretend to offer solutions to financial economic problems. We do, however, purport to offer a variety of analytic models that can be used to aid financial decision-making. These models are presented in the spirit that they are tools, which can be used for good or bad. But we do contend that investigating these tools is important in helping to better understand our global inter-connected economy, with its financial opportunities and risks. The responsibility for investment decisions remains with human investors. This book presents a number of operations research model applications to financial risk management. It is based on a framework of four perspectives, each with appended current examples, with separate chapters based on published models designed to support financial risk management. The four perspectives used are accounting (explaining the COSO framework in Chapter 1), finance (reviewing some basic conceptual tools in Chapter 3), economic (risk theory in Chapter 11), and sustainability (Chapter 16). Current issues related to each of these perspectives are appended. Chapter 2 supplements the overview introductory chapter by discussing the ethical risk issues highlighted by the Enron case. Chapter 4 supplements the financial perspective chapter with a review of the 2008 real estate crash, both in the US and in Europe. Chapter 12 supplements the economic perspective chapter with a review of the risks associated xv

xvi Preface

with the British Petroleum oil spill in the Gulf of Mexico. Chapter 17 supplements the sustainability chapter with reviews of some natural disaster events. Given this framework with current examples, the focus of the book is on quantitative models presented to support risk management in finance. These models include sentiment analysis, data envelopment analysis, catastrophe bond modeling, chance constrained optimization, bank credit scoring, multiobjective credit scoring, and advanced time series modeling. Chapters 5 and 6 present sentiment analysis models, one of investment analysis, the second of stock price volatility. Chapter 7 presents a data envelopment analysis risk scoring model of internet stocks. Chapter 8 uses statistical credit scorecard modeling, while Chapter 9 applies a TOPSIS credit-scoring model supplemented with simulation. Chapter 10 utilizes principal components analysis to make DEA more efficient in analyzing the efficiency of on-line banking. To supplement the economic perspective, another data envelopment analysis model is used to assess bank branch efficiency in Chapter 13, whereas Chapter 14 describes catastrophe bond loss modeling, and Chapter 15 bilevel mathematical programming in bank merger analysis. The sustainability perspective is supported by use of GARCH-type forecasting models of carbon emissions markets in Chapter 18, while similar tools are used to forecast crude oil prices in Chapter 19. Chapter 20 provides a summary of how risk management can be studies using a Confucius three-stage learning approach. Thus the book provides a variety of tools for assessing different types of financial risk situations.

Note 1.

L. Laeven and F. Valencia (2008) ‘Systemic banking crises: a new database’, International Monetary Fund, Working Paper WP/08/224.

Acknowledgements Chapter 5 follows the work of Li et al. (2009)1 on financial risk analysis. We acknowledge Taylor & Francis for granting us rights to revise and publish this work in a book format. In this chapter we consider a simplified version of Li et al. (2009). Readers may refer to Li et al. (2009) for further theoretical and modeling issues. Chapter 6 follows the work of Wu et al. (2014)2 on financial risk analysis. In this chapter we consider a simplified version of Wu et al. (2014). Readers may refer to Wu et al. (2014) for theoretical and modeling issues. Chapter 7 follows the work of Ho, Wu and Olson (2009).3 We acknowledge World Scientific for granting us rights to revise and publish this work in a book format. Readers may refer to Ho, Wu and Olson (2009) for further theoretical and modeling issues. Chapter 8 follows the work of Wu and Olson (2010).4 We acknowledge Palgrave Macmillan for granting us rights to the revise and publish this work in a book format. Readers may refer to Wu and Olson (2010) for further theoretical and modeling issues. Chapter 9 follows the work of Wu and Olson.5 We acknowledge Idea Group Inc. for granting us rights to revise and publish this work in a book format. Readers may refer to Wu and Olson (2006) for further theoretical and modeling issues. Chapter 10 follows the work of Wu and Wu (2010)6 on banking operations. We acknowledge Emerald for granting us rights to revise and publish this work in a book format. Chapter 13 follows the work of Wu et al. (2006)7 on financial risk analysis. We acknowledge Elsevier for granting us rights to revise and publish this work in a book format. Readers may refer to Wu et al. (2006) for further theoretical and modeling issues. Chapter 14 follows the work of Wu and Zhou (2010)8 on Cat bond and financial risk analysis. We acknowledge Taylor & Francis for granting us rights to revise and publish this work in a book format. Readers may refer to Wu and Zhou (2010) for further theoretical and modeling issues. Chapter 15 follows the work of Wu et al. (2014)9 on financial merger analysis. We acknowledge Wiley and the Production and Operations Management Society for granting us rights to revise and publish this work in a book format. Readers may refer to Wu et al. (2014) for further theoretical and modeling issues.

xvii

xviii

Acknowledgements

Chapter 18 follows the work of Chen et al. (2010)10 on carbon emission pricing. We acknowledge Taylor & Francis for granting us rights to revise and publish this work in a book format. In this chapter we consider a simplified and demonstration version of Chen et al. (2010). Chapter 19 follows the work of Luo et al.11 dealing with carbon emission pricing. We acknowledge Emerald for granting us rights to revise and publish this work in a book format. Chapter 20 follows the work of Wu.12 We acknowledge Inderscience for granting us rights to revise and publish this work in a book format.

Notes 1. N. Li, X. Liang, X. Li, C. Wang, Desheng D. Wu (2009) ‘Network environment and financial risk using machine learning and sentiment analysis,’ Human and Ecological Risk Assessment, 15(2): 227–252. 2. D. Wu, L. Zheng, D.L. Olson (2014) ‘A decision support approach for online stock forum sentiment analysis,’ IEEE Transactions on Systems Man and Cybernetics, 44(8): 1077–1087. 3. Chien-Ta Bruce Ho, Desheng Dash Wu, David L. Olson (2009) ‘A risk scoring model and application to measuring internet stock performance,’ International Journal of Information Technology and Decision Making, 8(1): 133–149. 4. Desheng Dash Wu, David L. Olson (2010) ‘Enterprise risk management: coping with model risk in a large bank,’ Journal of the Operational Research Society, 61(2): 774–787. 5. D. Wu, D.L. Olson (2006) ‘A TOPSIS data mining demonstration and application to credit scoring,’ International Journal of Data Warehousing & Mining, 2(3): 1–10. 6. D. Wu, D.D. Wu (2010) ‘Performance evaluation and risk analysis of online banking service,’ Kybernetics, 39(5): 723–734. 7. D. Wu, Z. Yang, L. Liang(2006) ‘Using DEA-neural network approach to evaluate branch efficiency of a large Canadian bank,’ Expert System with Applications, 108–115. 8. D. Wu, Y. Zhou (2010) ‘Catastrophe bond and risk modeling: a review and calibration using Chinese earthquake loss data,’ Human and Ecological Risk Assessment, 16(3): 510–523. 9. D. Wu, C. Luo, H. Wang, J.R. Birge (2014) ‘Bilevel programming merger evaluation and application to banking operations,’ Production and Operations Management. DOI: 10.1111/poms.12205. Accepted and in press. 10. X. Chen, Z. Wang, D.D. Wu (2013) ‘Modeling the price mechanism of carbon emission exchange in the European Union emission trading system,’ Human and Ecological Risk Assessment, 19(5): 1309–1323. 11. C. Luo, L.A. Seco, H. Wang, D.D. Wu (2010) Risk modeling in crude oil market: a comparison of Markov switching and GARCH models’, Kybernetics, 39(5): 750–769. 12. D.D. Wu (2014) ‘An approach for learning risk management: confucianism system and risk theory,’ International Journal of Financial Services Management. Accepted and in press.

1

Enterprise Risk Management

Introduction Living and working in today’s environment involves many risks. The processes used to make decisions in this environment should consider the need both to keep people gainfully employed (through increased economic activity) and to protect humanity from threats arising from human activity. Terrorism led to the gas attack on the Japanese subway system in 1995, to 9/11 in 2001, and to the bombings of the Spanish and British transportation systems in 2004 and 2005 respectively. But nature has been far more deadly, with hurricanes in Florida, tsunamis in Japan, earthquakes in China, and volcanoes in Iceland. These locations only represent recent, well-publicized events. Nature can strike at us anywhere. We need to consider the many risks that exist, and to come up with strategies, controls, and regulations that accomplish a complex combination of goals. Risks can be viewed as threats, but business exists to cope with risks. No one should expect compensation or profit without taking on some risk. The key to successful risk management is to select those risks that one is competent to deal with, and to find some way to avoid, reduce, or insure against those risks not in this category. Consideration of risk has always been part of business, manifesting itself in the growth of coffee houses such as Lloyd’s of London in the 17th century, spreading risk related to cargoes on the high seas. The field of insurance developed to cover a wide variety of risks, related to external and internal risks covering natural catastrophes, accidents, human error, and even fraud. Enterprise risk management (ERM) is a systematic, integrated approach to managing all risks facing an organization. It focuses on board supervision, aiming to identify, evaluate, and manage all major corporate risks in an integrated framework. The board is responsible for providing strategic input, identifying performance objectives, making key personnel appointments, and providing management oversight. Enterprise risks are inherently 1

2

Enterprise Risk Management in Finance

part of corporate strategy. Thus consideration of risks in strategy selection can be one way to control them. ERM can be viewed as top-down by necessity for this reason.

Definition Risk management can be defined as the process of identification, analysis and either acceptance or mitigation of uncertainty in investment decision-making. Once risk has been processed in this manner, risk management seeks coordinated and economical application of resources to control the probability and/ or impact of adverse events, and to monitor the effectiveness of actions taken.1 Risk management is about managing uncertainty related to a threat. ERM has been recognized as being one of the most important issues in business management in the last decade. There are systematic variations in ERM practices in the financial services industry. There is a need to monitor and address all risks inherent in organizational operations as necessary to avoid economic catastrophe. There is a need to consider all corporate risks within a single ERM framework in order to gain long-run competitive advantage. In the US, recent crises include the 2007 subprime crisis of the banking industry, the Fannie Mae and Freddie Mac crisis in secondary US mortgage markets, the failure of Lehman Brothers, Merrill Lynch’s takeover by Bank of America and insurance industry giant AIG applying for emergency financial support from the Federal Reserve. More recently, the H1N1 virus has sharpened the awareness of the response system worldwide. Risks can arise in many facets of business. Global economic crisis risks are profound and widespread over the last decade. Businesses in fact exist to cope with risk in their area of specialization. But chief executive officers are responsible for dealing with any risk that fate throws at their organization. Risk management began in the financial disciplines. Financial risk management has focused on banking, accounting, and finance. There are many good organizations that have done excellent work to aid organizations dealing with those specific forms of risk, applying many types of models. Risk management can also be applied in other areas, to include accounting. Risk management can be defined as the process of identification, analysis and either acceptance or mitigation of uncertainty in investment decision-making. Risk management is about managing uncertainty related to a threat. Traditional risk management focuses on risks stemming from physical or legal causes such as natural disasters or fires, accidents, death and lawsuits. Financial risk management deals with risks that can be managed using traded financial instruments. The most recent concept, enterprise risk management, provides a tool to enhance the value of systems, both commercial and communal, from a systematic point of view. Operations research (OR) is always useful for optimizing risk management.

Enterprise Risk Management

3

Accounting perspective Accounting responsibilities involve auditing organizational operations to provide stakeholders with accurate, transparent information of finances. This includes assuring that a sound process is in place to detect, deal with, and monitor risk. The accounting approach to risk management is centered to a large degree on the standards promulgated by the Committee on Sponsoring Organizations of the Treadway Commission (COSO), generated by the Treadway Commission beginning in 1992. The Sarbanes–Oxley Act of 2002 outlines regulatory requirements for publicly traded firms to establish, evaluate, and assess the effectiveness of internal accounting controls. SOC has had a synergistic impact with COSO. While many companies have not used it, COSO offers a framework for organizations to manage risk.2 COSO objectives are: 1. Effectiveness and efficiency of operations 2. Reliability of financial reporting 3. Compliance with applicable laws and regulations. To attain these objectives, COSO identifies the components of internal control: • • • • •

Control environment Risk assessment Control activities Information and communication Monitoring.

COSO was found to be used to a large extent by only 11% of the organizations surveyed, and only 15% of the respondents believed that their internal auditors used the COSO 1992 framework in full. Chief executive officers and chief financial officers are required to certify effective internal controls. These controls can be assessed against COSO. This benefits stakeholders. Risk management is now understood to be a strategic activity, and risk standards can ensure uniform risk assessment across the organization. Resources are more likely to be devoted to the most important risk, and better responsiveness to change is obtained. The COSO framework In 2004, COSO published an Enterprise Risk Management – Integrated Framework.3 COSO provides a framework to manage enterprise uncertainty, expressed in their ERM Cube. The cube considers dimension of objective categories, activities, and organizational levels, as shown in Table 1.1.

4

Enterprise Risk Management in Finance

Table 1.1

COSO ERM cube1

Categories

Activities

Levels

Strategic Operations Reporting Compliance

Internal environment Objective setting Event identification Risk assessment Risk response Control activities Information & communication Monitoring

Entity level Division Business unit Subsidiary

Note: 1COSO (2004). Enterprise Risk Management – Integrated Framework: Executive Summary. September.

This framework provides key principles and concepts, a common language, and clear direction and guidance.4 Categories The strategic level involves overarching activities such as organizational governance, strategic objectives, business models, consideration of external forces, and other factors. The operations level is concerned with business processes, value chains, financial flows, and related issues. Reporting includes information systems as well as means to communicate organizational performance on multiple dimensions, to include finance, reputation, and intellectual property. Compliance considers organizational reporting on legal, contractual, and other regulatory requirements (including environmental). Activities The COSO internal control process consists of a series of actions.5 1. Internal Environment: The process starts with identification of the organizational units, with entity level representing the overall organization. The tone is set by the top of the organization. This includes actions to develop a risk management philosophy, create a risk management culture, and design a risk management organizational structure. 2. Objective Setting: Each participating division, business unit, and subsidiary would then identify business objectives and strategic alternatives, reflecting vision for enterprise success. These objectives would be categorized as strategic, operations, reporting, and compliance. These objectives need to be integrated with enterprise objectives at the entity level. Objectives should be clear and strategic, and should reflect the entity-wide risk appetite.

Enterprise Risk Management

5

3. Event Identification: Management needs to identify events that could influence organizational performance, either positively or negatively. Risk events are identified, along with event interdependencies. (Some events are isolated, while others are correlated.) Measurement issues associated with methodologies or risk assessment techniques need to be considered. 4. Risk Assessment: Each of the risks identified in Step 3 is assessed in terms of probability of occurrence, as well as the impact each risk will have on the organization. Thus both impact and likelihood are considered. Their product provides a metric for ranking risks. Assessment techniques can include point estimates, ranges, or best/worst-case scenarios. 5. Risk Response: Strategies available to manage risks are developed. These can include risk acceptance, risk avoidance, risk sharing, or risk reduction. Options have been summarized into the four Ts: a. Treat a risk: Take direct action to reduce impact or likelihood. b. Terminate a risk: Discontinue activity exposing the organization to the risk. c. Transfer a risk: Insurance or contracts. d. Take (or tolerate) a risk: For areas of organizational expertise, they may decide to accept risk with the idea that they are expert at dealing with it. Avoidance is akin to terminating, acceptance to treating, reduction and transfer to transfer above, and seeking risks to toleration. Risks are necessary to lead to situations likely to offer profit, but risks should be taken only after informed business analysis. The effects of risk response on other risks should be considered. 6. Control Activities: Controls needed to mitigate identified risks are selected. Implicit in this step is assessment of the costs of each risk response available, and consideration of activities to reduce risks. 7. Information and Communication: Control and other risk response activities are put in place to ensure appropriate action is taken within the organization. Organizations need to ensure that information systems can measure and report risk accurately. ERM effectiveness and cost should be communicated to stakeholders. 8. Monitoring: As part of an ongoing process, the effectiveness of plan implementation is monitored, feeding back to the control step if problems are encountered. Monitoring includes risk evaluations comparing actual event occurrences with prior estimates of probability, frequency, and cost.

6

Enterprise Risk Management in Finance

Risk appetite Risks are necessary to do business. Every organization can be viewed as a specialist at dealing with at least one type of risk. Insurance companies specialize in assessing the market value of risks, and offer policies that transfer special types of risks to themselves from their clients at a fee. Banks specialize in the risk of loan repayment, and survive when they are effective at managing these risks. Construction companies specialize in the risks of making buildings or other facilities. However, risks come at organizations from every direction. Those risks that are outside of an organization’s specialty are outside that organization’s risk appetite. Management needs to assess risks associated with the opportunities presented to it, and accept those that fit its risk appetite (or organizational expertise), and offload other risks in some way (see step 6 above). Risk appetite is the amount of risk that an organization is willing to accept in pursuit of value. Each organization pursues various objectives to add value, and should broadly understand the risk it is willing to undertake in doing so.6 COSO recommends three steps to determine risk appetite: 1. Developing risk appetite requires consideration of the current level and distribution of risks the organization faces, assessment of the level of risk that the organization can handle, the acceptable level of risk that the organization is willing to accept, and attitude concerning growth, risk, and return. 2. Once the risk appetite appropriate to the organization is agreed upon, it must be communicated throughout the organization. 3. Risks affecting the organization need to be monitored in terms of quantitative and qualitative measures. Examples of specific measures for a business might be: • • • • • • • •

How customer requirements are being met Shareholder expectations Strategic initiatives and growth Financial reporting Operational performance Regulatory compliance Employee health and safety Environmental responsibility.

Matyjewicz and D’Arcangelo gave simple examples of how risk assessment could be applied. First, a matrix of risk level (high or low) and control strength (weak or strong) could be generated for each identified risk. Risk impact could be further categorized, as critical, significant, moderate, low, or insignificant,

Enterprise Risk Management

7

while risk probability could have categories of highly probable, probable, likely, unlikely, or remote. The likely actions of internal auditing were identified. Those risks involving high risk and strong controls would call for checking that inherent risks were in fact mitigated by risk response strategies and controls. Risks involving high risk and weak controls would call for checking for adequacy of management’s action plan to improve controls. Those risks assessed as low call for internal auditing to review accuracy of managerial impact evaluation and risk event likelihood.

ERM process COSO provides a great deal of help in describing a process for risk management implementation.7 This process should be continuous, supporting the organization’s strategy. One possible list of steps could be: 1. Risk Identification: identification of what could go wrong without controls, considering key organizational areas such as mission, customers, people, physical assets, and financial assets. 2. Risk Analysis: consideration of risk likelihood and impact, and ranking them accordingly. 3. Response to Significant Risks: toleration, treatment, transfer, or termination. 4. Identification of Controls: where risks can be effectively dealt with; and the controls needed to effectively respond. 5. Reporting and Monitoring: documentation of risk evaluation, events, impacts, and effectiveness of strategies. It is paramount that a climate of dealing with risks be spread throughout the organization. This can be aided by specific identification of roles, responsibilities, and communications channels, along with clearly stated risk strategy and appetite. Protocols in the form of risk guidelines should include rules and procedures along with specific risk management methodologies, tools, and techniques. It is also important to clearly identify responsibilities. COSO provides typical risk responsibilities, as shown in Table 1.2. Implementation issues Past risk management efforts have been characterized by bottom-up implementation. But effective implementation calls for top-down management, as do most organizational efforts. Without top support, lack of funding will starve most efforts. Related to that, top support is needed to coordinate efforts so that silo mentalities do not take over. COSO requires a holistic approach. If COSO is adopted within daily processes, it can effectively strengthen corporate

8

Enterprise Risk Management in Finance

Table 1.2 Risk management responsibilities1 Entity

Risk management responsibilities

CEO/Board

Set strategic approach and risk appetite Establish risk management structure Identify most significant risk Crisis management

Unit managers

Establish risk awareness culture Set performance targets Ensure implementation Identify and report changes

Individuals

Understand RM processes Report insufficient controls Report loss events and close calls Cooperate with incident investigations

Risk manager

Develop current risk management policy Document policies and structures Coordinate internal controls Compile data and reports

Internal audit manager

Develop risk-based audit program Audit risk processes Provide assurance of risk management Report internal control efficiency/effectiveness

Note: 1The Association of Risk Managers (2010). A Structured Approach to Enterprise Risk Management (ERM) and the Requirements of ISO 31000. COSO.

governance. Another important issue is the application of sufficient resources to effectively implement ERM. One view of ERM, parallel to that of the CMI system used in software engineering, is as follows.8 Level 1: Compliance – review of policy and procedure with a checklist orientation, providing low value to the organization in terms of ERM. Level 2: Control – implementation of control frameworks, still using a checklist orientation, also providing low value to organizations. Level 3: Process – taking a process view across departments, focusing on effectiveness as well as efficiency, to include process mapping. Level 4: Risk Management – use of shared risk language, with the ability to prioritize efforts based on process mapping. Level 5: Enterprise Risk Management – the Nirvana of holistic risk reviews tied to entity strategy based on common risk language, viewing risk management as a process, providing high value to organizational risk management.

Enterprise Risk Management

9

Risk management modeling There have been many models applied to managing risks. There are many published research papers proposing optimization modeling to ERM. We feel that optimization of supply chain systems involving risk is dangerous, in that we think there is a fundamental conflict between optimal planning of complex systems (which seeks to eliminate all excess resources) and the capability of a system to deal with risk. Think of Chicago’s O’Hare Airport. Airlines schedule it to the optimum, seeking maximum revenue through service to the most customers feasible. But anyone who has traveled through O’Hare knows that the system is saturated – the slightest drop of rain results in a cascade of cancelled flights throughout the US and Canada. The airlines have dealt with their risk – they can pay inconvenienced customers the minimum allowed by regulation. The customers on the other hand travel through O’Hare at their own risk (at least with respect to on-time arrival at their destination). Should the airlines back off their optimal schedule, they would have greater slack capacity to absorb the unexpected, which occurs whenever the weather in Chicago turns the least bit nasty, which occurs almost always.

Book outline Chapter 2 provides a review of the Enron event, demonstrating why SOX happened. Chapter 3 will focus on financial risks, not restricted to market risk, credit risk, operational risk, operational risk, liquidity risk. Financial risk has been controlled through hedge funds and other tools over the years, often by investment banks. Chapter 4 discusses the 2007–2008 real estate financing crisis. Chapter 11 discusses the insurance perspective, realizing that many risks can be prevented, or their impact reduced, through loss-prevention and control systems, leading to a broader view of risk management. Chapter 12 demonstrates its importance by reviewing the BP oil spill. Chapter 16 looks at the wider view of sustainability and risk in globalization. Chapter 17 reviews some of the risks to business operations arising from natural disaster, calling for insurance. Chapter 20 presents a Confucian perspective on risk management. Section II of the book presents a variety of quantitative modeling techniques proposed for ERM. Models provide a means to quantify risks, and to aid decision-making concerning issues with complex interactions. Effective management of risks inherently involves tradeoffs. Optimization models may identify solutions with the greatest expected short-term profits, but these solutions also tend to have high levels of risk, especially in the longer term. Simulation models enable consideration of uncertainties, as long as they are expressed in the form of probability distributions. Multiple criteria models

10 Enterprise Risk Management in Finance

focus on the analysis of tradeoffs. Once models are used to examine expected relationships between causes and effects, risk reduction and management can be more effective. The usual forms of management of risks tend to be based upon either financial models, or through frameworks. Risk management modeling tools demonstrated include GARCH, data mining through neural networks, bilevel programming, efficiency analysis through DEA, sentiment analysis, and stochastic simulation. These tools offer advanced techniques available to aid decision-making under risk.

2

Enron

One of the primary reasons for the current emphasis on ethical business practice has been the history of the Enron Corporation. Enron was founded in 1985 to manage a natural gas pipeline. It expanded operations to include trading not only natural gas but also other energy commodities, including gas and electricity. This trading included derivatives on the price of gas, to hedge risk. Enron participated in a joint venture creating an online trading operation offering options and other derivatives to traders in the gas industry. By 2000 it was well known as a trading pioneer in that field, and this led to its stock doing well on Wall Street; in 2001 it was rated as 7th on the Fortune 500.

Risk management Enron had a strong reputation for its use of sophisticated financial risk management tools. It was in a business with long-term fixed commitments that needed hedging to survive the inevitable fluctuations in energy prices. The sophisticated tools included derivatives and transfer of risk to special entities, but the problem was that Enron owned these special entities. Thus it really didn’t transfer risk, but hedged with itself. Enron accounting practices were creative, hiding losses and shuffling debts through complex trades,1 but such practice was actually quite common in the deregulated fervor of the 1990s.2 Enron also became involved in trading electricity in California, and was caught manipulating that market.3 Late in 2001 the Securities and Exchange Commission began investigating Enron, and Enron stock dropped. Enron filed for Chapter 11 bankruptcy on December 2, 2001. During this fall, Enron executives off-loaded stock and gave themselves large compensation packages while urging lower-level employees to retain their stock. Many high-profile court trials ensued, with criminal convictions of CEO Kenneth Lay and President Jeffrey Skilling in 2006. Lay was convicted in May 2006 on five counts of

11

12 Enterprise Risk Management in Finance

securities fraud, wire fraud, and conspiracy to commit the same, but he did not go to prison as he died of a heart attack in July, before sentencing. Skilling was convicted in May 2006 of 28 counts of securities fraud, wire fraud, conspiracy to commit same, false statements, and insider trading. He was sentenced to 24 years and 4 months of prison, and ordered to provide $45 million in restitution to victims. Skilling entered prison in 2006, with appeals pending; in June 2013, his sentence was reduced to 14 years in prison.4 California electricity In 1996, California modified its controls on its energy markets, seeking to increase competition. A spot market for energy began operation in April 1998. In May 2000, significant price increases for energy were experienced, due to shortage of supply. This shortage has been attributed to a cap on retail electricity prices, along with market manipulation and illegal shutdowns of pipelines by Enron. On June 14, 2000, California endured multiple blackouts on a large scale. Drought, as well as delays in approval for new energy plants, also played a role. The fiasco played a part in Governor Gray Davis’s political popularity. In August 2000, San Diego Gas & Electric Company filed a complaint about electricity market manipulation. January 17 and 18, 2001, saw more blackouts, and Governor Davis declared a state of emergency. More blackouts occurred on March 19 and 20. Pacific Gas & Electric filed for bankruptcy in April, and there were more blackouts on May 7 and 8. Energy prices stabilized in September. In December, following Enron’s bankruptcy, allegations were made that Enron had manipulated energy prices, and the Federal Energy Regulatory Commission began investigation in February 2002. The Wall Street Journal published a series of studies of the Enron case. Reasons for the fall of Enron included:5 1. 2. 3. 4. 5.

Bad investments (12/4/2001) Loss of trust in Enron Online traders (12/6/2001) Bad hedging, bad trading, bad assets (12/26/2001) Efforts to expand into market with no experience (12/31/2001) Creation of a complex structure of subsidiaries and financial instruments evading clear explanation to investors (1/11/2002) 6. Accounting statements not providing investors with a complete picture nor a fair assessment of risks (1/16/2002) 7. Adoption of self-dealing partnerships with accountants and lawyers covering up illegal actions 1/21/2002). The worst of Enron’s egregious activities included its California energy market manipulation, which gouged California electricity customers of billions, as

Enron

13

well as betrayal of its own employees, who were encouraged to invest their retirement funds in Enron stock while the executive board sold them out. Accounting impact The first manifestation of change after the fall of Enron was the fall of its audit firm, Arthur Andersen, in June 2002. Andersen was convicted of obstruction of justice, for shredding documents related to Enron. This barred it from auditing US and foreign-based public companies, leaving 1085 firms in need of a new auditor. The Sarbanes–Oxley Act was passed in 2002, requiring additional auditing requirements. The remaining Big Four audit firms were flooded with business by these two events, leading them to drop 500 clients between 2002 and 2005 (and also enabling them to raise audit fees).6

SOX The Sarbanes–Oxley Act of 2002 is a federal law setting standards for US public company boards. It was enacted in reaction to corporate and accounting scandals to include Enron, Tyco International, Adelphia, Peregrine Systems, and WorldCom, each of which involved company collapses that led to heavy losses by investors. SOX provided reforms to include requirements for certification, criminal penalties to chief officers of offending corporations, internal controls, independent audit committees, and regulations of disclosure. The Sarbanes–Oxley Act provides for a number of sections displayed in Table 2.1. Sarbanes–Oxley (SOX) has been heavily studied. On the positive side, evidence of increased transparency of firms under SOX has been reported. The costs of compliance were found to be 0.043% of revenue in 2006, and 0.036% of revenue in 2007, with costs much higher for decentralized companies with multiple divisions.7 Surveys have found that there has been a positive impact on investor confidence in the reliability of financial statements and in fraud prevention. However, most survey respondents have seen the benefits to be less than the cost. Costs incurred are for external auditor fees, insurance for officers and directors, board compensation, lost productivity, and legal costs. Reaction of firms to avoid SOX have included going private, delisting on stock exchanges, and staying small enough to avoid its requirements.8 Smaller international companies have been found to prefer listing in UK stock exchanges rather than US stock exchanges, indicating a cost impact on smaller firms.9 Studies have found that those firms that use avoidance tend to be small, less financially endowed, with weak governance and weak performance. Larger firms in the US have been found to reduce risk-taking (in terms of capital expenditure, R&D, and variance in cash flow and returns).10

14 Enterprise Risk Management in Finance

Table 2.1

Sarbanes–Oxley act elements

Accounting oversight board

Board created to provide independent oversight of public accounting firms offering audit services.

Auditor independence

Standards for external auditor independence to limit conflicts of interest.

Corporate responsibility

Senior executives are individually responsible for financial report accuracy and completeness.

Financial disclosure

Reporting requirements for financial transactions.

Analyst conflicts of interest

Definition of codes of conduct for securities analysts, requiring conflict of interest disclosure.

Commission resources and authority

Practices to restore investor confidence

Studies and reports

Comptroller General and SEC required to perform studies and report findings on effect of consolidation of public accounting firms, role of credit rating agencies, securities violations and enforcement, investment bank participation in earnings manipulation and financial condition obfuscation.

Corporate and criminal fraud accountability

Criminal penalties for manipulation, destruction, or alternation of financial records described, whistle-blower protection provided.

White collar crime penalty enhancement

Increased criminal penalties for white-collar crimes and conspiracies.

Corporate tax returns

CEO required to sign company tax return.

Corporate fraud accountability

Identifies corporate fraud, revises sentencing guidelines.

Conclusions SOX was not solely driven by the Enron case, but that was probably the most salient representative of a general business climate in the US in the 1990s, when the drive to greater profit overrode concern about risk. Thus SOX provides some structure that probably aids systematic management of risk within organizations. Primarily, it makes it harder for business directors to mislead investors, stipulating reporting requirements that more accurately document a firm’s financial performance, and stipulating greater accountability for the board of directors and especially for their audit committee. Systems such as SOX and the International Organization for Standardization (ISO) provide greater security and control, often implemented through enterprise resource planning (ERP) systems that provide standardized processes and reduce human errors. There are some problems, because SOX and ISO are essentially additional bureaucracy that imposes rigidity. Overall, however, SOX and ISO achieve a great deal in the effort to regulate dishonest business behavior.

3

Financial Risk Management

Introduction Traditional risk management focuses on risks stemming from physical or legal causes such as natural disasters or fires, accidents, death and lawsuits. Financial risk management deals with risks that can be managed using traded financial instruments. The events of the 21st century have made it even more critical. Top business management came under suspicion after the scandals at ENRON, WorldCom, and other business entities. In recent times, many investors have experienced difficulties from bubbles. The most spectacular failure in the late 20th century was probably that of Long-Term Capital Management,1 but that was only a precursor to the more comprehensive failure of technology firms during the dotcom bubble around 2001. The global financial community suffered the 2007 subprime crisis of the banking industry, the Fannie Mae and Freddie Mac crisis in secondary US mortgage markets, Lehman Brothers’ failure, Merrill Lynch’s takeover by Bank of America, and industry-giant AIG applying for emergency financial support from the Federal Reserve. The financial world’s failures include the Barings Bank collapse in 1995, as well as the Long-Term Capital Management and subprime mortgage bubble implosion already mentioned. Financial management needs to obtain a return on capital, while simultaneously maintaining positive cash flow. Banks, the petroleum industry, and commodity trading of all types need to consider these fundamental requirements. This chapter provides an overview of risks in human investment activity, and describes a few basic tools to aid in financial enterprise risk management. Concepts: Hedging as a means to trade-off return for assurance Copulas as a means of hedging Value-at-risk as a tool for daily operations of cash management Tranching of investment products

15

16 Enterprise Risk Management in Finance

The Basel Accords have been instrumental in providing guidance for banking risk management. The Basel Committee on Bank Supervision sets standards with the purpose of prudent bank regulation. Basel I was created in 1988 by regulatory representatives of G-10 countries, along with central bank input. Basel I set minimum capital requirements, grouping bank assets into categories of credit risk with differential liquid asset holdings specified for various levels of risk. Basel II was published in 2004, aimed at providing international standards for banking regulations and mitigating against a sequence of related bank failures arising from strong cross-relationships among banks. In addition to minimum capital requirements, Basel II provided for supervisory review and market discipline. After the banking crisis of 2008, Basel III was published in 2010–2011, increasing liquidity requirements and decreasing bank leverage. The traditional financial risk management approach is based on the mean-variance framework of portfolio theory, that is, selection and diversification.2 Over the past 20 years, the field of financial risk management has experienced a fast and advanced growth at an incredible speed. It is widely recognized in finance that risk can be understood as two types: systematic (non-diversifiable) risk, which is positively correlated with the rate of return, and unsystematic (diversifiable) risk, which can be diversified by increasing the number of securities invested. Based on the portfolio theory, the Capital Asset Pricing Model (CAPM) was discovered to price risky assets on perfect capital markets.3 A derivative market grew tremendously, with the recognition of option pricing theory.4 Value-at-Risk models have been popular,5 partly in response to Basel II banking guidelines. Other analytic tools include simulation of internal risk rating systems using past data. Economic risk management is originally based on the Expected Utility Theory by recognizing people’s risk attitude on different sizes of risk – small, medium, large – which is derived from the utility-of-wealth function.6 People are risk averse when the size of the risk is large, and risk neutral when the scale of risk is small.7 Decision-making behaviour when faced with lotteries and other gambles motivates most of the studies on risk issues. The complexity presented by derivatives, to include CDOs and CDSs, is complicated even more by the use of high-speed computer trading. Risks are traditionally defined as the combination of probability and severity, but are actually characterized by additional factors. Markowitz (1952) equated risk with variance, which would be controlled by diversification, considering correlation across investments available, and focused on efficient portfolios non-dominated with respect to risk and return. This leads to the need for some calculus of preferences, such as multi-attribute utility theory. Financial risk management has developed additional tools such as value at risk. Risk characteristics include uncertainties, dynamics, dependence, clusters, and complexities, which motivate the utilization of various operational research tools. Risks can

Financial Risk Management

17

be viewed as having three properties: Probability, dynamics, and dependence. Probability in risk management deals with distribution models. This approach can be dated back to the 1700s, leading to the Bernoulli, Poisson, and Gaussian models of events, the generalized Pareto distributions, and the generalized extreme value distributions to model extreme events. Risk dynamics uses stochastic process theory in risk management. This can be dated back to the 1930s when Markov processes, Brownian motion and Levy processes were developed. Dependence of risks deals with correlation among risk factors. Various copula functions are built, and Fourier transformations are also used. The dependence of returns across investments leads to additional complexities.

Investment collars The idea of a collar model is an option strategy using puts and calls simultaneously to manage investment risk. A put option is purchased by the investor gaining the right to sell the underlying equity shares at a stated exercise price. This provides downside protection. To offset the cost of the put, the investor sells a call option, where underlying shares are sold at another exercise price. The collar refers to the range set by the put and call options. Collars used in this manner are intended to provide an alternative to diversification for risk management, given the difficulties experienced in 2008 with correlation across investments intended to be diversified. The benefit of the collar is to limit downside risk. The cost is that upside benefit is capped (as well as the investor being out the cost of obtaining put and call options). There is additional risk of counterparty default on the put side. Collars have also been applied in the case of adjustable rate mortgages with respect to interest rates. VaR Value at risk (VaR) is one of the most widely used models in risk management (Gordon, 2009). It is based on probability and statistics (Jorion, 1997). VaR can be characterized as a maximum expected loss, given a time horizon and within a given confidence interval. Its utility is in providing a measure of risk that illustrates the risk inherent in a portfolio with multiple risk factors, such as portfolios held by large banks, which are diversified across many risk factors and product types. VaR is used to estimate the boundaries of risk for a portfolio over a given time period, for an assumed probability distribution of market performance. The purpose is to diagnose risk exposure. Definition

Value at risk describes the probability distribution for the value (earnings or losses) of an investment (firm, portfolio, etc.). The mean is a point estimate of a statistic showing a historical central tendency. Value at risk is also a point

18 Enterprise Risk Management in Finance

0.6 Probability density function 0.5 0.4 0.3 0.2

1–α

Maximum loss

0.1 VaR

0.0 –1

Figure 3.1

0

1 Mean

CVaR 2

3

4 Loss

CVaR and VaR

estimate, but offset from the mean. It requires specification of a given probability level, and then provides the point estimate of the return or better expected to occur at the prescribed probability. For instance, Figure 3.1 gives the normal distribution for a statistic with a mean of 10 and a standard deviation of 4 (Crystal Ball was used, with 10,000 replications). However, value at risk has undesirable properties, especially for gain and loss data with non-elliptical distributions. It satisfies the well accepted principle of diversification under normal distribution; however, it violates the fairly well accepted subadditive rule; that is, the portfolio VaR is not smaller than the sum of the component VaR. The reason is that VaR only considers the extreme percentile of a gain/loss distribution without considering the magnitude of the loss. As a consequence, a variant of VaR, usually labeled Conditional-Valueat-Risk (or CVaR), has been used. In computational issues, optimization of CVaR can be very simple, which is another reason for adoption of CVaR. This pioneer work was initiated by Rockafellar and Uryasev (2002), where CVaR constraints in optimization problems can be formulated as linear constraints. CVaR represents a weighted average between the value at risk and losses exceeding the value at risk. CVaR is a risk assessment approach used to reduce the probability that a portfolio will incur large losses assuming a specified confidence level. It is possible to maximize portfolio return subject to constraints including Conditional Value-at-Risk (CVaR) and other downside risk measures, both absolute and relative to a benchmark (market and liability-based). Simulation CVaR-based optimization models can be developed. Value-at-risk (VaR) has become a popular risk management tool. It is often used to measure the risk of loss on a specific portfolio of financial assets, in terms of the probability of losing a specified percentage of the portfolio in mark-to-market value (current market price) over a certain time. If there is a 0.01 probability that a given portfolio has a daily 1% VaR of $100,000, this

Financial Risk Management

19

portfolio will lose $1000 (=0.01*100,000) in a given day with that probability. This infers that a loss of $1000 in a day with this portfolio is expected on 1 day out of 100. Banks commonly report VaR by type of market factors, such as currency rates, commodity prices, interest rates, equity prices, etc. The financial industry has grown to view VaR as a metric indicating relative change in risk for a given investment. In fact, before 2008, investment banks were known to have guided their people to seek investments with higher VaR in the expectation that they would yield higher returns, relying on the portfolio aspects of CDOs to manage the risk.8 There are two broad uses of VaR. The short-range use is for risk management. If the organization’s portfolio measured in mark-to-market terms falls below the VaR at the end of the time horizon, assets are sold off to gain enough cash to cover the deficiency. But there is an alternative use in risk measurement, taking a longer-term view. Here, the aim is to measure the fluctuation in VaR and use it as an indicator of trends in relative risk for the firm or investment department. There are three basic ways to compute value-at-risk. The statistical approach assumes a distribution (normal commonly) which can use the variancecovariance of the investments in a given portfolio. We will demonstrate this calculation in the simple case of one investment (avoiding the more complex formulation involving covariance). Historical simulation looks at past data and simply sorts outcomes, selecting the probability level desired; if you want a 0.99 probability assurance of not exceeding a VaR, use the observation that is the 0.01 lowest. The third method is Monte Carlo simulation, which is the most flexible (you can model any assumption you want, to include select distributions you prefer, and can complicate the model with external probabilities such as the probability of catastrophe). However, Monte Carlo is also the most involved, and provides an estimated solution rather than a precisely calculated outcome. Even the simplest of these methods, using variance calculations, involves some complications in the details. First, the definition’s strict sense is the measure of the probability of the worst case falling below VaR over a given time horizon. This turns out to be fairly difficult to compute. Thus a redefinition of the end-of-period value, to below VaR, is often used. Another complication is the time horizon. One day is the most common, in which case the time horizon isn’t so important; however, there are cases for longer time horizons. Yet another complication is that the distribution of outcomes for positive events (favorable) might differ from that of negative events (losses). This in fact would be expected if outcomes were lognormally distributed. If the normal distribution were used with different distributions for positive and negative outcomes, daily price changes could be separated into positive and negative groups, and each group analyzed separately. Since VaR is worried about losses, the negative group would be of interest.

20

Enterprise Risk Management in Finance

The distribution of financial returns underlying VaR calculations involves fat tails.9 Taleb (2012) argued that we tend to see the world in its stable form, and don’t expect rare events we haven’t seen lately (black swans). We tend to expect a normal distribution, but if in reality the distribution has a higher probability of rare events, the tails also have a higher probability, or become fatter. Another caveat about VaR is that it doesn’t represent the worst that could happen – it represents a point on the distribution of possible outcomes. For a 0.99 probability level, if outcomes do happen to follow the assumed distribution, outcomes worse than the VaR would be expected once out of 100 events. For a 250-day trading year, that amounts to 2.5 the expected times that losses would exceed the VaR.

Copulas Copulas were proposed by David X. Li in 2000.10 The idea is that bundles of investments with risks would be safer than each investment individually, and the copula was a means to identify the relative risk of the pool of risky assets. This was very germane to the financial derivatives based on mortgages that were generated by financial institutions in the 1990s. Using a copula calculation based on the normal distribution (Gaussian copula), banks could bundle high-risk investments together in collateralized debt obligations (CDOs) that had much lower joint risk. CDOs are asset-backed securities that allow their issuer to use cash flows from different debt assets to back bonds which the issuer could then sell to investors. Furthermore, the issuer could market portions of these bonds associated with different levels of risks (tranches), which investors in turn would assume reflected different returns (the assumption being that higher risk yielded higher returns, and the pooled nature of the underlying assets reduced risks). Thus the issuing banks could sell CDO tranches to a variety of buyers interested in different levels of risk. Credit rating institutions saw the logic, and were so comforted by the reduced risk that they gave very favorable credit ratings to CDOs. Furthermore, while the number of mortgages was finite, CDOs could be sold multiple times, vastly increasing the amount of money at risk. As to copulas, the fundamental measure was the probability of joint default. If component investment survival times were independent of each other, the risk for the set would be easy to calculate. However, positive correlation between survival times has been widely observed. Li’s Gaussian copula formulation derived the joint distribution from the marginal distributions of the component investments. Issuers of CDOs came to rely on the Gaussian copula (which assumes a normal distribution of outcomes). That formulation assumed a number of

Financial Risk Management

21

things, including stable correlation across assets over time. But in the housing mortgage industry around 2007–2008, the flaw in this logic was seriously exposed. Furthermore, Salmon (2009)11 pointed out that mortgage pools have greater volatility than most bonds, as there is no guaranteed interest rate, in part because mortgage borrowers do things like being late on payments, quitting making payments, or conversely, paying loans off early. The pooling idea was expected to be safe, first because few living people remembered house prices doing anything but going up, second because housing markets across a country (or across the world) were expected to be independent enough that a down period in one area would be offset by gains in other areas, third because defaults were rare and spread out over time. However, 2007 and 2008 saw widespread decline in highly populated areas of the US, generating high degrees of correlation.

Tranches The financial industry developed some highly innovative investment vehicles after the deregulation of banks. These included collections of mortgages into credit default swaps (CDSs) or collateralized debt obligations (CDOs). Credit default swaps are agreements where the seller of the CDS will pay the buyer should some credit arrangement default. Anybody can purchase a CDS, even those without any interest in the loan upon which the CDS is based. CDSs are marketable. A CDO is an asset-backed security, often used in the mortgage market. The underlying investments involve cash flows, which investors are after. CDOs usually involve collections of loans (such as mortgages) of varying degrees and risk, and possibly different areas and types of housing to diversify risk. The basic idea is that this diversification provides an assurance of collective security while each of the component parts of the CDO can be sold to investors, reflecting different returns due to being sorted into tranches of relative risk. This scheme worked extraordinarily well from deregulation in the late 1990s until 2007, when the unforeseen circumstance of house price declines occurred. Before 2007, loans were pushed on any housing investor in sight, without much regard to their ability to repay, because the initiating banks were not going to hold the mortgages, but were going to sell them as commodities to investment banks. The investment banks in turn were not too concerned about repayment, as they sold their CDOs on to investors. When the underlying house prices started to fall, however, the system collapsed, with ripple effects suffered around the world. What happened was that the probabilities of default jumped up, because borrowers were finding that their house value was dropping to below their mortgage level, and they were “underwater” on their loans. Some borrowers took that as a reason to default. More often, however,

22

Enterprise Risk Management in Finance

the borrowers were overextended by the underlying economic decline occurring at the same time, and were thus unable to pay their mortgages, so the banks foreclosed.

Conclusions The overall lesson seems to be that there is no free lunch. It doesn’t make sense to speculate on things that appear too good to be true. If too many people do that, the system will inevitably catch up with the herd involved in that irrational behavior. An accompanying lesson, unfortunately, is that those agile enough, like some initiating lenders and investment banks, can rip off many investors before the herd sees the light.

4

The Real Estate Crash of 2008

Introduction The current population of the United States has grown up with what seemed to be a steady and reliable increase in value of homes. Owning one’s own home is one of the signs of a prosperous and successful culture. In the 1930s, many people lost title to their homes during one of the greatest failures of human economic systems known. In response, banks and mortgage lending were regulated, bank deposits insured up to a level covering what most people had, and stock-trading practices ostensibly controlled. True, very old people could remember a time when the price of housing dropped, but with the inevitable passage of life with time, this group grew smaller and smaller, and older and older, and less relevant. There also were anomalies in local areas where, for whatever reason, home prices might negatively fluctuate. Reinhart and Rogoff have described five such anomalies, all associated with banking crises (Spain in 1977, Norway in 1987, Finland and Sweden in 1991, and Japan in 1992).1 But house price decline occurred very rarely, and nobody really noticed. There also was a strong feeling that less regulation was always better. Ronald Reagan may have been a democrat in his younger years, but he became the champion of the conservative class. He rode this popularity to the presidency, which was characterized by an emphasis on standing up to aircraft controllers and Sandinistas, along with a conservative preference for less government interference. Even when a democratic president appeared, such as Bill Clinton, his economic regulation (whether due to preference or poll-counting) did not vary that much from conservatives. Many of the regulations imposed during the 1930s were overturned in an effort to let the market run free, which was expected to lead to a golden age of prosperity. Financial research developed many tools that became popular during this period. The Efficient Market Hypothesis held that asset prices are always and everywhere at the correct price, a view that fits well with the concept that that 23

24 Enterprise Risk Management in Finance

economy is regulated best that is regulated least.2 Investors prospered during the 1990s, none more than Long-Term Capital Management (LTCM). LTCM was built on the financial models of Black, Scholes, and Merton, providing tools to price derivatives. Billionaires who could afford to buy into LTCM made billions.3 But human economic activity consists of a complex interaction of many actors, some more exuberant about prospects than others. And yet, while on average for every buyer there is a seller, and the idea of a stable equilibrium seems reasonable, human history is full of periods where a market will go crazy and drive prices beyond reason. These periods are known as bubbles. Consider the following: • • • •

1630 tulip mania 1720 the Mississippi Company 1720 the South Sea Company 1929 stock market crash.

These are only four of many economic crashes, all preceded by excessive rapid growth (bubbles).4 There are many studies of these bubbles. Some bubbles consist of expansion, followed by rising prices, overtrading, and mass participation, followed by an event triggering doubt, a subsequent selling flood, and ultimate collapse. The 1990s saw the crash of LTCM, demonstrating that theoretical models do NOT cover everything, that every model leaves something out, and that life is more complex than anyone understands. Trade in technical stocks made the NASDAQ highly popular, but it also demonstrated a bubble, crashing the confidence of the brave new world of computer techs.

Real estate in 2008 Yet investors still held great confidence in certain sectors of the economy, such as real estate, which was as safe as houses. Risk managers created new tools to make investment safer. Derivatives5 are securities or contracts deriving value from an underlying natural security, such as a stock, a bond, or a mortgage. While derivatives provide some security through diversification, their primary attraction is that they enable high degrees of leverage. The crisis is credited with beginning with the collapse of the US subprime residential mortgage market in 2007, which spread throughout the world due to exposure to US real estate assets through financial derivatives.6 Causes could be the growth in asset securitization, US government initiatives to expand home ownership (thus encouraging fewer loan restrictions), expansionary monetary policy, and weaker regulatory oversight.7 The real estate price boom was furthered by financial institution exploitation of loopholes in capital regulation, allowing them to significantly increase leverage while remaining within required capitalization. Mortgage derivatives allowed investment in riskier and non-liquid assets funded in wholesale markets, without sufficient capital backing. This, along with high dependence on a short-term view and lax regulatory oversight, has

The Real Estate Crash of 2008 25

been credited with inducing the collapse of the bubble in 2008.8 Distress first appeared in 2007 with losses by US subprime loan originators and those holding derivatives based upon such mortgages. In late 2007, losses by Northern Rock, a UK mortgage lender, indicated that the bust was going global. Wiseman (2013)9 outlined four stages in a real estate cycle, with corresponding activities (see Table 4.1).

Table 4.1 Real estate cycle Phase

Characteristics

Peak

Boom

Government raises interest rates to cool economy Consumer confidence high Higher demand for goods & services Aggregate supply near aggregate demand Adequate housing inventory, home prices stable New home starts and real estate activity stable Bank-owned inventory and foreclosures low Low vacancies, high rental prices Speculators overvalue properties

Contraction

Bust

Restrictive and expensive credit financing Decreased consumer confidence, declining consumer demand Aggregate supply > aggregate demand Housing inventory builds as sales decline Home prices start to decline New home starts decline Banked-owned inventory and foreclosures increase Commodity based properties (farmland) decline Competition declines, speculators sell off

Trough

Bust

Government lowers interest rates to stimulate Consumer confidence low, supply > demand Excess housing inventory, low sales level Home prices stabilize at low levels New home starts low Bank-owned inventory & foreclosures high Vacancies high, rental prices low Speculators tend to undervalue properties Buyer’s market

Recovery

Boom

Easy, cheap credit Increasing consumer confidence Increasing consumer demand for goods & services Aggregate supply > aggregate demand Demand for housing exceeds inventory Commodity-based business properties peak in value Home sales increase Home prices increase More new home starts Bank-owned inventory & foreclosures decrease Vacancies decrease, rental prices rise Prices rise, speculators buy Seller’s market

Source: Based on Wiseman (2013).

26 Enterprise Risk Management in Finance

Mortgage system With deregulation, the mortgage industry became more specialized, with a number of organizations playing a role in the overall system. Lenders such as Green Tree Finance led the charge to issue as many mortgages as possible (mobile homes in the case of Green Tree; Ameriquest, Countrywide, Golden West and others for conventional homes).10 These mortgages were sold to banks and other investment agencies, so the mortgage initiators had little concern other than generating lots of mortgages and making a living out of the fees. In fact, many home owners saw the inevitable rise in home value to be an opportunity to make a financial killing by leveraging as many home loans as they could, making purchases for speculation rather than for residence. The purchasers of these mortgages often combined various tranches of different levels of perceived risk, with higher interest rate mortgages associated with higher probability of loan failure. This evolved into the convolution in logic that marginal borrowers, who had no choice but the highest interest rates, were preferred customers. These instruments were sold to investors. Meanwhile, these banks often covered their risk in innovative ways. A credit default swap (CDS) is a credit derivative similar in concept to insurance; should the underlying asset fail, the purchaser receives payment. A collateralized mortgage obligation (CMO) is a certificate built from tranches of mortgage-backed securities. Thus it is a tranched instrument of an instrument that is already tranched. A collateralized debt obligation (CDO) is similar, but can be based on any kind of debt, not just mortgages.11

Northern Rock The Northern Counties Permanent Building Society was established in 1850, serving Newcastle upon Tyne. The Rock Building Society was established in 1865. Both were building societies, and they merged to become Northern Rock, a mutually owned savings and mortgage bank. The Building Societies Act of 1986 allowed building societies to convert to public banks, allowing them access to wholesale money markets. Northern Rock went public in 1997, enabling it to sell shares on the stock market. It borrowed on capital markets, lent this money to customers, turned the loans into bonds, and sold the bonds. Sampath wrote about the organizational risk in reputation, using Northern Rock as a case in point.12 Basel II addressed operational risk, credit risk, and market risk. In a Pillar 2, it mentioned strategic risk, reputation risk, and nonstandard risk, but these last three categories of risk have no specific capitalization provisions. This is because there is less data available and this makes it difficult to quantify exposure. Sampath argued that Northern Rock demonstrated failure in management of these less quantifiable risks.

The Real Estate Crash of 2008 27

Within months of capitalization in 1997, Northern Rock was accused of arbitrarily changing the terms of its accounts and reducing the interest it paid to its depositors. In 1996 Northern Rock was forced to remove the earlyrepayment penalty clauses it had inserted into some of its fixed-rate mortgages. In 1997 it was criticized for increasing its mortgage rates in anticipation of rising base rates that did not materialize. In 1998 it was cited as among the worst mortgage institutions in terms of press coverage. Thus Northern Rock’s actions jeopardized its reputation risk. In 2002 it instituted a social and customer service commitment in efforts to use corporate social responsibility as a means to resurrect its image. It also set up programs to support gay and lesbian foundations, tackled domestic violence, and set aside 5% of pre-tax profits for charity. In 2004 Northern Rock floated bonds, shifting away from home loans. It depended heavily on short-term loans from other banks rather than on depositors to get the cash it needed to lend in mortgages. In mid-2007 it was caught by an increase in funding costs, which it was unable to pass on to its clients. It also overlooked hedging loans for the two months between mortgage initiation and securitization. The United Kingdom had not seen a bank run since 1866, when London bank Overend Gurney overextended its resources during the boom in railroad and dock construction. The US saw many bank runs in the 1930s, which led to protection by the Federal Deposit Insurance Corporation. In the United Kingdom, this protection was provided by the banking industry. On September 13, 2007, the BBC announced that Northern Rock Bank had sought Bank of England support, which was provided the next morning. While undoubtedly not the intended outcome, the effect was that Northern Rock depositors lined up to withdraw their deposits on September 14.13 Northern Rock was a mortgage bank, focusing on prime mortgages. It had a much heavier proportion of nonretail funding, in the form of short-term borrowing in capital markets and securitized notes – and the global credit crisis in the summer of 2007 led to a massive reduction in short-term funding and interbank lending. The French BNP Paribas closed three investment funds invested in US subprime mortgages on August 9, 2007, joining the difficulties many were experiencing in renewal of short-term borrowing. While Northern Rock had little, if any, subprime lending, it drew from the same pools for shortterm funding. Thus Northern Rock approached the Financial Services Authority (FSA) on August 13 for help. The FSA and the Bank of England sought to quietly deal with the crisis by finding a private buyer for Northern Rock, but failed, and had to announce its assistance to Northern Rock on August 14. Thus the public learned of Northern Rock’s problems, and depositors joined the lines to get their money.

28

Enterprise Risk Management in Finance

Table 4.2 Northern Rock events 2007

Event

July 25

Northern Rock issues optimistic outlook

August 9

BNP Paribas suspends three investment funds with subprime mortgages

August 13

Northern Rock informs regulators of funding difficulties

August 14

Bank of England alerted of Northern Rock difficulties

September 4

Money market problems increase, LIBOR reaches 9 year peak

September 12

Bank of England announces it would support banks through short-term loans, but not massive injection of funds

September 13

BBC reveals Northern Rock asked for, will receive BOE aid

September 14

BOE and others reveal Northern Rock will receive help, lines queue at Northern Rock

September 17

After stock market close, British Government announces guarantee of all Northern Rock deposits in turbulent period

September 19

BOE announces injection of liquidity into money markets, extension to mortgage debt

September 20

Government guarantee extended to unsecured wholesale funding

October 9

Government guarantee extended to new retail deposits

Source: Extracted from Goldsmith-Pinkham and Yorulmazer (2010).14

Northern Rock’s difficulties in obtaining short-term funds were not due to its lending practices, but rather to the system’s inability to provide funds. This in turn was due to the subprime mortgage issues of 2007. The events are outlined in Table 4.2. Table 4.3 shows types of retail deposits at Northern Rock before and after the run. Table 4.4 shows the holdings of Northern Rock before (June 2007) and after (December 2007) the liquidity crisis. As a mutual, Northern Rock dealt with retail mortgages. Securitized notes were medium- to long-term, which outstripped retail deposits after going public. Covered bonds were illiquid longterm liabilities against segregated mortgage assets. Wholesale liabilities (money markets) were nonretail funding not covered as covered bonds or securitized notes. While Northern Rock did not make subprime loans, they were vulnerable in a market where housing values were in decline. Because they were overleveraged, they suffered the first bank run in Great Britain in over a century. Northern Rock failed strategically. It shifted away from its traditional market of mutual mortgage lending, seeking perceived higher profits in broader lending. Sampath attributes its troubles to underestimation of reputational risk,

The Real Estate Crash of 2008 29

Table 4.3 Northern Rock retail deposits (millions of pounds) December 2006

December 2007

Change

10,201 5,573 4,105 2,752 22,631

4,351 3,035 1,371 1,712 10,469

−5,850 −2,538 −2,734 −1,040 −12,162

Postal accounts Branch accounts Offshore & other Internet & telephone Totals Source: Extracted from Shin (2009).

Table 4.4 Northern Rock holdings before and after run (millions of pounds)

Securitized notes Wholesale Retail Covered bonds BOE Loan Totals

June 2007

December 2007

Change

45,698 26,710 24,350 8,105 – 104,863

43,070 11,472 10,469 8,938 28,473 102,422

−2,628 −15,238 −13,881 +833 +28,473 −2,441

Source: Extracted from Shin (2009).

which jeopardized public confidence. Restrictive and sharp practices destroyed its credibility and reputation, and it was hit with a run that was withstood only through Bank of England intervention.

AIG AIG is the world’s largest insurance company. It began in China. In 1926 it opened operations in the US to write insurance on American risks outside the US, and in the 1930s it started to buy American insurance companies.15 It became very large, and by the time of the real estate crisis in 2007, this had a bearing on the risk AIG was exposed to; on paper, it could say that it had offloaded some risk to a reinsurer, but since it owned the reinsurer, the risk was retained.16 Starting in 1999, AIG and its subsidiaries issued a large number of CDSs. These provided a very strong revenue stream for the firm when market conditions were stable, and there were low default rates. A feature of CDSs purchased by AIG from investment banks was credit support annexes, standard contracts attached to swap agreements mandating that the instrument be marked to

30 Enterprise Risk Management in Finance

Table 4.5 Key events for AIG Date

Event

February 11, 2008

AIG announced write down of $4.88 billion in CDSs

September 15, 2008

AIG reported to be seeking $40 billion in capital to avoid downgrading by credit rating firms

September 17, 2008

Fed authorized loan of $85 billion to AIG, giving Government 79.9% equity in AIG and veto power over dividends, to be repaid in 24 months

October 9, 2008

Fed authorized bailout package of another $37.8 billion in securities in exchange for cash collateral

Source: Extracted from Egginton et al. (2010).

market price nightly. The investment banks were buying insurance that the CDS would not fall below a certain value, and a check every night made it more likely that AIG would have to pay off the swap.17 Events related to the 2008 real estate crisis are shown in Table 4.5. While AIG made a lot of money issuing CDSs before 2008, investment banks took the opportunity to purchase many CDSs that paid off in 2008. In fact, they purchased more CDSs than the value of the underlying mortgage assets. By September 16, 2008, AIG was in severe difficulty, its stock down to $3.75 (it had been $63.44 a year earlier).18 The failure of AIG has been attributed to high-leverage trading, just as with large banks such as Lehman Brothers and Bear Stearns.19 Other problems cited were lack of transparency with respect to the risk of CDSs and CDOs; adverse selection, in that investment banks knew more about the risks associated with the coverage they purchased from AIG than AIG did; and the high magnitude of unhedged CDOs held by AIG ($562 billion). The issue with unhedged CDOs was complicated in that the conventional expectation is that they are naturally diversified, but the mortgage markets upon which they were based turned out to have a highly correlated downward trend.

Risk management The Northern Rock case demonstrates that there is more to risk management than simply dealing with the hedging aspects of financial instruments. Linsley and Slack focused on the ethical criticisms of Northern Rock.20 After becoming a publicly traded bank, with the ability to access money market (wholesale) financing, it found itself financing long-term mortgages with short-term funding, which created a technical financial issue, as its cash flow was subject to the variations in short-term financing. In the summer of 2007, the sub-prime

The Real Estate Crash of 2008 31

mortgage issues led to less liquidity on the wholesale money markets, placing Northern Rock under great stress. Linsley and Slack looked at Northern Rock’s press releases over the period 2005–2008, finding that it emphasized robustness and strength in the early period, based on its claimed performance and growth, the strength of the housing market and economy, a sound strategy and business model, and its ability to manage lending risk. There was no reference to stakeholder relationships, and Linsley and Slack detected no evidence of Northern Rock caring about customer welfare. Northern Rock did benefit favorably from its previous mutual status, its local nature, and The Northern Rock Foundation, which provided a charitable presence. The crisis led stakeholders to adjust their view of Northern Rock. Northern Rock’s communications implied that the crisis was an external problem that would disappear with time. The press release announcing that Her Majesty’s Treasury had guaranteed a portion of deposits appears to have had little acknowledgement of depositor concerns, and to have underestimated the negative impact. The passive nature of Northern Rock’s response was attributed to be the reason for the protection release announcement, leading to the run on the bank. Clearly we infer that bank risk management involves more than monitoring VaR, and even more than hedging. Banks by their very nature depend upon customer confidence. The Northern Rock case demonstrates the negative impact of not paying attention to developing and maintaining good relationships with its depositors.

5

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

Introduction There is a widespread need for effective forecasting of financial risk using readily available financial measures, but the complicated environment facing financial practitioners and business institutions makes this very challenging. The concept of financial volatility, a required parameter for pricing many kinds of financial assets and derivatives, is critical, because it is widely expected that financial volatility implies financial risk. Therefore, accurate prediction of financial volatility is extremely important. Efficient prediction of financial volatility has been an extremely difficult task, but we can now offer a scalable and customizable mathematical model to achieve this goal, employing two approaches to forecast the volatility using financial information available online. First, we carry out a comparative study between two different machine-learning techniques – artificial neural networks (ANN) and support vector machines (SVM) – to forecast trading volume volatility. Second, we utilize semantic techniques to probe correlations between information sentiment and asset price volatility.

Information volume and volatility Fluctuation in trading volume, just like that of stock price, vividly reflects market behavior. We investigate associations between trading volume volatility and online information volume, with the intent of forecasting the former. Online financial information volume has been assumed to be an important element affecting the financial trading volume volatility. We forecast volatility, relying partly on online information, using both an ANN-based and an SVM-based approach. We compare these forecasting models, to observe their performances. The basic architecture of this approach is displayed in Figure 5.1. 32

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

Training

Google finance

33

Forecasting

Financial news volumes

Yahoo finance

Financial trading volumes

GARCH-ANN approach

Generated rules

Forecast result

GARCH-SVM approach

Generated rules

Forecast result

Trading volume changing rates and corresponding volatilities

Comparative study Trading volume changing rates and corresponding volatilities Financial trading volumes

Yahoo finance

Data from immediate past

Financial news volumes

Google finance

Most recent data

Figure 5.1 Flow chart and functional parts of our approach associating information volume and volatility

We downloaded online financial information from Google Finance (http:// www.finance.google.com), as shown in Figure 5.1. The data was post-processed to obtain information volumes for various stocks and indices. Table 5.1 shows a snippet of the post-processing result, with each sub-row on top indicating the date, and the one below the volume, of the news. In Table 5.1, N, D, M, I, A, and G stand for NASDAQ, DOW, MSFT, INTC, AAPL and GOOG, respectively. 0629 stands for June 29, 2006. Information sentiment and volatility Representing financial information purely by its volume might be misleading, thus undermining the efficiency of the forecast. In order to investigate the impacts of online information upon financial time series more comprehensively, we explore its content as well, specifically emotional or sentimental polarity. An essential step is to evaluate the sentiment of each news entry, primarily accomplished by the methodology of bags-of-words. We aim to obtain a real value for each news entry whose sign denotes the authors’ judgmental state, while the absolute value indicates emotion intensity. HowNet (http://www.keenage.com/html/e_index.html) is a commonsense online knowledge base indicating inter-conceptual relations and inter-attribute relations of concepts utilizing Chinese lexicons and their English equivalents.1

34

Enterprise Risk Management in Finance

Table 5.1

The information volumes calculated from Google Finance 0629

0630

0703

0704

0705

0706

0707

...

7

4

5

2

4

4

3

...

D

0630

0703

0704

0705

0706

0707

0708

...

2

3

3

4

5

12

1

...

M

0628

0629

0630

0701

0703

0704

0705

...

5

8

3

1

2

2

4

...

I

0703

0705

0706

0707

0710

0711

0712

...

2

2

2

3

2

1

2

...

0630

0701

0703

0705

0706

0707

0710

...

7

1

4

3

6

2

2

...

0717 3

0718 5

0719 2

0720 8

0721 7

0724 4

0725 5

... ...

N

A

G

HowNet is used to calculate the sentiment of news pieces through its set of keywords. Each news piece is decomposed and converted into a keyword array in the same sequence as the words appear in the chapter, each word being assigned a specific sentiment value based on the HowNet word corpus. The overall sentiment for the whole chapter is acquired by combining the sentiment values of all of its keywords. After the sentiment time series is obtained, it is fed into the machine-learning system (SVM in particular) as one of the exogenous inputs. By assigning sentiment as one element of the feature vector for a listed company, non-linear correlation between online news sentiment and financial volatility can be quantitatively analyzed, which might eventually lead to more effective and efficient volatility forecasting. The basic architecture of this approach can be seen in Figure 5.2. Financial news used in this phase of the study was acquired from a variety of online sources, and experiments were carried out on a huge body of companies listed on US stock markets. Aggregated statistical results are the key tool to substantiate nonlinear correlation between the two entities. Table 5.2 is a snippet of the financial news entries with their calculated sentiment values; in the time window, 2007–1 indicates that this is a news entry in the first week of 2007. Volatility model and modified non-linear GARCH model Volatility refers to the standard deviation or variance of the change in value of a financial instrument within a specific time span. The GARCH system is widely employed in modeling financial time series that exhibit time-varying volatility clustering. In this section, we develop a GARCH system by incorporating financial information into the usual framework.

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

35

Webpage crawling

Internet

Cleaning Parsing

Data preprocessing

Segmenting Internet

Stemming

Sentiment calculation

How Net corpse

Historical quote Sentiment time series

Asset price volatility/ Trading volume volatility time series

Dynamic training and forecasting GARCH-SVM approach

Predicted time series

Prediction error

Figure 5.2 Flow chart and functional parts of our approach to associate information sentiment and volatility

The GARCH model, proposed by Bollerslev in 1986,2 can be formulated as yt = µ t + ε t ,

(5.1)

ε t |ψ t −1 ~ N (0, σ t2 ), p

σ t2 = α 0 +

∑α σ i

i =1

(5.2) q

2 t −i

+ ∑ β j ε t2− j , (a0 > 0, ai, bi ≥ 0) j =1

(5.3)

News title

An Insecure Future for McAfee

Stocks end 2006 with best gains in three years

Marathon Oils lower earnings top forecasts

Still looking good

2007010102057465

2007010202063354

2007050203261590

2007060203595695

2007–22

2007–18

2007–1

2007–1

Time window

ADCT

MRO

S

ADCT

Company symbol

A snippet of news entries for the companies ADCT, S and MRO

News ID

Table 5.2

Perhaps its fatigue with the options scandal that has now spread to more than 100 technology companies. Perhaps its ... NEW YORK (MarketWatch) – U.S. stocks finished the year with strong gains Friday, with all three major stock averages booking their best performance since 2003. The Dow Jones Industrial Average ($INDU : $INDUNews) SAN FRANCISCO (MarketWatch) – Marathon Oil Corp. reported Tuesday a drop in first-quarter earnings, clipped by lower oil and gas prices and a decline in production. For the three months ended March 31, Marathon (MRO : MRONews) ANNANDALE, Va. (MarketWatch) – It’s been a little bit over two months since the triggering of a rare, and historically very bullish, technical signal. (Read my March 22 column.) Can we count on the bullish winds of that signal blowing into the ...

News body

11.8

−0.8

17.8

45.4

News sentiment

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

37

where the daily return yt is sum of the deterministic mean return μt and a stochastic term εt, also known as the shock, forecast error, residual, innovation, etc.,3 ψt−1 represents the information set available at time t, and σt2 is the timevarying variance of both yt and εt. In our approach, we substitute yt with the daily changing rate of either the trading volume or the asset price. The GARCH model uses εt as a function of those exogenous inputs, which have some affect on financial volatility. The GARCH model bases its conditional distribution on the information set available at time t. Freisleben and Ripper 4 point out that the parameter βi in Equation (5.3) describes the stock return’s immediate reaction to new events in the market, mostly in the form of financial news. Meanwhile, the fast development of the internet enables us to acquire the online financial information in a real-time, exhaustive fashion. Considering these factors, designating financial information volume as one variate of εt is justifiable. Therefore we formulize εt using the following two equations, ε t = yt − ζ,

(5.4)

ε t = f t (Wt , ε 9t ) = g t (Wt ) + θ t ε 9t ,

(5.5)

where ζ is a constant and Wt is the on-line financial information volume on day t. Consequently a modified GARCH model can be expressed as yt = µ t + ε t ,

(5.6)

ε t |ψ t −1 ~ N (0,σ t2 ) , p

α t2 = α 0 +

∑α

i

χ t − i (σ t2− i ) +

i =1

r

(5.7) q

∑β ϕ j

j =1

t−j

( yt2− j ) (5.8)

+ ∑ γ k φt − k (Wt2− k ) k =1

where p, q, r represent the three time lags, the three unknown functions, xt–i, φt − j, and ϕ t − k represent the undetermined nonlinear correlations. Financial time series exhibit specific features that make a GARCH model a preferable alternative. We assume that the volatility of financial trading volume shares similar characteristics to that of stock price in exhibiting GARCH effects. Figure 5.3 demonstrates that the daily changing rates of the trading volumes of NASDAQ index within the period October 11, 1984–October 16, 2006 exhibit volatility clustering. A kurtosis value of 11.53 also implies that there is an underlying fat tail effect. We have discovered obvious GARCH effects exhibited by online financial information volume time series in tests conducted on more than 100 stocks in the US stock markets.

38

Enterprise Risk Management in Finance

1

0.5

0

–0.5

–1

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

5500

Figure 5.3 Daily changing rates of the trading volumes of NASDAQ index, October 11, 1984–October 16, 2006

Daily volatility model In associating information volume with trading volume volatility, we define the variance of the trading volume’s daily changing rates within a period starting on the previous day and ending on the current day. For a specific stock or index, let vt denote trading volume on day t, so that the daily changing rate yt of the trading volume is denoted as yt = ln

vt . vt −1

(5.9)

If D is defined as the width of the calculating window, the volatility st2 can be calculated by computing the variance of the yt within the (t − D + 1)th and the tth day as D −1

2 t

σ =

∑ (y

t −i

− yt )2

i=0

D −1

,

(5.10)

where D −1

yt =

∑y i=0

D

t −i

.

(5.11)

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

39

Daily volatilities are calculated based on a sliding volatility window of a certain length. Each day’s volatility represents the variation in the value of trading volume over the past few days until the current day. Using GARCH-based SVM to associate information sentiment with asset price volatility This section investigates the relationship between volatility and information sentiment. We incorporate the average sentiment value into the current forecasting model, and predict how volatility will move in the immediate future using a more comprehensive perspective. Sentiment analysis of financial news The sentiment of a news chapter is based upon the sentiment values of all its keywords, that is, those containing emotional polarity. We used the English lexicon released by HowNet to form the essential word sets based upon which the sentiment calculation for a keyword is implemented. We generated eight word sets: POSITIVE, NEGATIVE, PRIVATIVE, MODIFIER i (i = 1,2,3,4,5). Each MODIFIERi is bound with a weight value WEIGHTi denoting the intensity of the words listed in this set. The intensity of the words increases with weight value. Table 5.3 defines these eight sets. We calculated keyword sentiment by counting matches, using the algorithm flowcharted in Figure 5.4. The sentiment for an entire chapter is calculated by summing the sentiments for all the keywords contained in that chapter.

Table 5.3 The eight word sets we use in this chapter to calculate the keyword sentiment Word sets

Description

POSITIVE

A list of English words that have positive emotional polarity, which includes a set of 4363 words.

NEGATIVE

A list of English words that have negative emotional polarity, which includes a set of 4574 words.

PRIVATIVE

A list of privative English words, which includes 14 words. {no, not, none, neither, never, hardly, seldom, barely, scarcely, ain’t, aren’t, isn’t, hasn’t, haven’t}

The following five sets are modifiers, whose intensities decrease while i increases. MODIFIER1 MODIFIER 2 MODIFIER3 MODIFIER4 MODIFIER5

64 modifier words, with WEIGHT1 = 2. 25 modifier words, with WEIGHT−2 = 1.8. 22 modifier words, with WEIGHT−3 = 1.6. 15 modifier words, with WEIGHT−4 = 1.4. 11 modifier words, with WEIGHT−5 = 0.8.

40 Enterprise Risk Management in Finance

Get the current word w to be analyzed, and set its sentiment value v to 0

Is w a positive word?

N

Is w a negative word?

Set v to 1

Set v to –1

Get the list of words which are within k words before w

N

Y Get the list of words which are either within m words before w or n words after w

Multiply v with –1

Does any of the word in this list appear in MODIFIERi Y N

Multiply v with WEIGHTi

Output v

End

Figure 5.4

Output v as 0

Y

Y

Does any of the word in this list appear in PRIVATIVE?

N

Sentiment calculation process for the current keyword w

End

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

W1

…

Wi–1

Wi

Wi+1

…

WT

Wi+1

…

WT

41

Training

W1

…

Wi–1

Wi Forecasting

Figure 5.5

Sliding time window learning and forecasting

Using GARCH-based SVM to associate information sentiment and volatility As mentioned, volatility will be calculated on a time window basis, meaning that the time series will first be segmented into a series of time windows. The GARCH-based SVM approach is carried out on a sliding window basis; each SVM is trained on data collected from the current and the previous time window. The well-trained SVM is then used to predict volatility in the next time window. Figure 5.5 illustrates the process of sliding time window machine learning, where Wi−1 and Wi constitute the training input and output respectively, and Wi and Wi+1 constitute the forecasting input and output respectively. Each time window in Figure 5.5 corresponds to an input and output matrix formatted for machine learning and prediction. These matrices contain expanded companies in order to generate aggregated statistics for the entire stock market. If there are M listed companies of interest, let denote the training input and output matrix tuple to forecast the volatility in time window Wi, giving: −  p2  2 p2 σ i −1(1) yi −1(1) Si −1(1)   p2  − 2 p2 I = σ i −1(2) yi −1(2) Si −1(2)    ...   p2  2 p−2 σ i −1( M ) yi −1( M ) Si −1( M )

and σ ip 2 (1)   p2  σ (2)  O =  i , ...   p2 σ i ( M )

42 Enterprise Risk Management in Finance

where σ ip−21( k ) (k = 1,2,3, ... M ) is the asset price volatility within time window − Wi−1 for company k, yip−21( k ) represents the average daily changing rate of the asset price of company k in Wi−1, and Si2−1( k ) is the sum of the sentiment values for all the news entries relating to company k within the time window Wi−1. Accordingly, denote by the forecasting input and output matrix tuple for Wi, and we have

−p 2  p2  2 σ i (1) yi (1) Si (1)    −p 2 2 p2 I 9 = σ i (2) yi (2) Si (2)    ...   p2  −p 2 2 σ i ( M ) yi ( M ) Si ( M )

and σ ip+21(1)   p2  σ (2)  O9 =  i +1 . ...   σ ip+21( M )

Empirical results and analysis The empirical studies are roughly composed of two parts. In our first experiment, we utilize both GARCH-based ANN and SVM to study the correlations between financial information volume and trading volume volatility, and conduct a comparative study of different machine-learning techniques in financial volatility forecasting. The daily volatility model serves as the primary approach to calculating volatility. The data sets used in this step are limited to the trading quote and news data for two indices and two listed companies in the US stock markets. The time horizon spans three months. In the second experiment, we apply the GARCH-based SVM model to data from all of 2007 for 177 listed companies in the US markets, but in this experiment a timewindow-based volatility model is used instead. For both experiments, historical financial quotation data is downloaded and formatted from Yahoo Finance (http://www.finance.yahoo.com).

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

43

Trading volume volatility forecasting We utilize both ANN and SVM to forecast volatilities. The SVM toolbox adopted in this paper is the LS-SVMlab (http://www.esat.kuleuven.ac.be/sista/ lssvmlab/). The least squares SVM (LS-SVM)5 is a reformulation of the regular SVM. The kernel function we used is the radial basis function (RBF). Financial information was acquired from Google Finance, yielding a comprehensive set of online financial information for the previous three months from more than 500 financial portals. The width of the volatility calculating windows (D) in both phases is set to 20 days. We use σ^t2 to denote forecast volatility on day t; if σ^t2 − σ 2t−1 and σt2 − Δσ 2t−1 share the same sign, we say that an accurate forecast for the volatility trend Δσt2 has been achieved. The aforementioned ratio is defined as the percentage of the days on which the forecast volatility trend has been accurately forecast. We conducted experiments on two indices (NASDAQ and DOW) and two stocks (MSFT and INTC) with time spans from June 30, 2006 to September 28, 2006; from June 29, 2006 to September 26, 2006; from June 28, 2006 to September 26, 2006; and from July 3, 2006 to September 28, 2006, respectively. The optimal parameters for RBF in SVM (the regularization parameter gam and the bandwidth sig2) are set to sig2 = 50 and gam = 10. The results of these experiments are shown in Table 5.4, where s/i stands for stock/index, and N, D, M, and I stand for NASDAQ, DOW, MSFT, and INTC, respectively. C represents the size of the samples, H the number of the hidden nodes, and p, q, r the three time lags. In order to compare different results, we altered the parameter values for both GARCH-based ANN and SVM three times. Table 5.4 gives predicted values of the average forecast error and the volatility trend forecast accuracy ratio for these three scenarios. Results indicate that GARCH-based SVM outperforms GARCH-based ANN for volatility forecasts, whereas GARCHbased ANN achieves a better forecast result for the volatility trend. This is primarily because SVM is characteristic of the use of kernels, the absence of local minima, the sparseness of the solution and the capacity control obtained by optimizing the margin, which enables its better generality in overcoming overfitting phenomenon. In addition, SVM is a preferable solution, especially for small-scaled sample set, with our experiment as a case in point. Nonetheless, considering the forecast for volatility trend, the ANN-based approach considerably outplays the other. Our studies show that if we take the online information volume as an exogenous input, the forecast performance for the trading volume volatility considerably outperforms the price return volatility. Therefore trading volume is more likely to be affected by online financial information. In addition, we have found out that the larger the value of D, the smaller the average forecast error, which shows the volatility clustering feature of the financial time series.

44 Enterprise Risk Management in Finance

Table 5.4 Predicted values of the average forecast error and the volatility trend forecast accuracy ratio s/i

model

C

H

p

q

r

ē (%)

ratio (%)

N

ANN SVM ANN SVM ANN SVM

20 20 10 10 10 10

5 – 5 – 4 –

4 4 5 5 3 3

9 9 8 8 8 8

1 1 2 2 3 3

11.08 9.60 9.33 7.30 7.67 7.62

83.33 58.33 73.91 52.17 69.57 47.83

ANN SVM ANN SVM

10 10 15 15

8 – 5 –

6 6 9 9

6 6 4 4

1 1 2 2

10.17 9.42 12.96 11.34

68.00 64.00 64.71 64.71

M

ANN SVM ANN SVM ANN SVM

20 20 10 10 10 10

5 – 8 – 6 –

4 4 5 5 5 5

9 9 8 8 5 5

1 1 2 2 3 3

9.72 8.39 9.39 6.69 11.04 7.81

83.33 66.67 82.61 60.87 69.23 57.69

I

ANN SVM ANN SVM

10 10 15 15

6 – 6 –

5 5 9 9

5 5 9 9

3 3 3 3

14.31 13.55 16.72 16.09

61.54 38.46 64.71 41.18

D

Additionally, a better forecast performance can be achieved if we square the moving average component of the input vector, which substantiates one of the GARCH theory’s contentions that there is a significant correlation between the squared residuals of financial time series. Volatility forecasting with sentiment analysis Correlations between financial news sentiment value and stock price volatility of listed companies are investigated using GARCH-based SVM regression. A time window with a length of seven days (five trading days) is used. The forecast is implemented progressively, with sliding time windows, yielding aggregated statistics for all companies. Before being fed into the SVM model, both quote and news data are segmented into time windows. The financial news in this experiment is downloaded from over 200 English portals on the internet. The SVM toolbox utilized in this experiment is the open source library LIBSVM (http://www.csie.ntu.edu.tw/~cjlin/libsvm/). The news entries have a time horizon ranging from January 1, 2007 to December 3, 2007, spanning 49 time windows. In total, a set of 177 listed companies in the US stock markets are studied. There are 153,468 pieces of news for all the companies, all of which are tagged with their sentiment values computed. The kernel function used is

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

45

the RBF. Two major performance metrics are introduced in this experiment to evaluate the aggregated forecast performance for all the 177 companies: squared correlation coefficient (SCC) and volatility trend forecast accuracy (VTFA). SCC and VTFA are computed based on the forecasting values for each time window. The squared correlation coefficient evaluates the correlation of all the explanatory variables to the response variable. The closer this value is to 1, the better regression result is achieved. The volatility trend forecast accuracy is the proportion of companies with an accurately predicted volatility trend among all the companies. The definition of volatility trend is consistent with the first experiment. Table 5.5 presents a demonstration of the asset price volatility for 177 companies using information sentiment during the year 2007. Note that the penalty parameter c and the RBF kernel parameter g are set as c = 64 and g = 1/3.

Table 5.5 Forecast results for 177 listed companies during the year 2007 Time window for forecast output

Time window for forecast input

Time windows for training samples

Week3 Week4 Week5 Week6 Week7 Week8 Week9 Week10 Week11 Week12 Week13 Week14 Week15 Week16 Week17 Week18 Week19 Week20 Week21 Week22 Week23 Week24 Week25 Week26

Week2 Week3 Week4 Week5 Week6 Week7 Week8 Week9 Week10 Week11 Week12 Week13 Week14 Week15 Week16 Week17 Week18 Week19 Week20 Week21 Week22 Week23 Week24 Week25

Week1–2 Week2–3 Week3–4 Week4–5 Week5–6 Week6–7 Week7–8 Week8–9 Week9–10 Week10–11 Week11–12 Week12–13 Week13–14 Week14–15 Week15–16 Week16–17 Week17–18 Week18–19 Week19–20 Week20–21 Week21–22 Week22–23 Week23–24 Week24–25

SCC (R2) 0.662534 0.669676 0.873588 0.248732 0.584079 0.89559 0.987398 0.98691 0.882611 0.794255 0.491824 0.743331 0.615906 0.176804 0.981393 0.742617 0.565846 0.967438 0.932697 0.906246 0.940825 0.8779 0.06059 0.084561

VTFA 0.655367 0.525424 0.59887 0.694915 0.59887 0.632768 0.305085 0.649718 0.706215 0.548023 0.734463 0.661017 0.638418 0.59887 0.525424 0.672316 0.683616 0.559322 0.711864 0.621469 0.468927 0.824859 0.632768 0.615819 Continued

46

Enterprise Risk Management in Finance

Table 5.5 Continued Time window for forecast output

Time window for forecast input

Time windows for training samples

Week27 Week28 Week29 Week30 Week31 Week32 Week33 Week34 Week35 Week36 Week37 Week38 Week39 Week40 Week41 Week42 Week43 Week44 Week45 Week46 Week47 Week48 Week49 Week50

Week26 Week27 Week28 Week29 Week30 Week31 Week32 Week33 Week34 Week35 Week36 Week37 Week38 Week39 Week40 Week41 Week42 Week43 Week44 Week45 Week46 Week47 Week48 Week49

Week25–26 Week26–27 Week27–28 Week28–29 Week29–30 Week30–31 Week31–32 Week32–33 Week33–34 Week34–35 Week35–36 Week36–37 Week37–38 Week38–39 Week39–40 Week40–41 Week41–42 Week42–43 Week43–44 Week44–45 Week45–46 Week46–47 Week47–48 Week48–49

SCC (R2)

VTFA

0.394098 0.404327 0.781131 0.855221 0.302073 0.496343 0.700048 0.991294 0.998906 0.985648 0.921044 0.661739 0.656383 0.932568 0.99876 0.970877 0.464741 0.97924 0.214432 0.325735 0.843826 0.928572 0.724147 0.999941

0.615819 0.570621 0.638418 0.525424 0.610169 0.548023 0.644068 0.536723 0.587571 0.689266 0.632768 0.59322 0.638418 0.451977 0.711864 0.60452 0.508475 0.666667 0.587571 0.564972 0.429379 0.508475 0.672316 0.553672

Observe that an average of 60.3225% is achieved for the volatility trend forecast and an average of 71.2593% is achieved for the squared correlation coefficient upon all the 177 companies with the 48 forecast time windows. Figures 5.6 and 5.7 illustrate the price volatility forecasts for two specific companies out of the 177, MDT and WAG. Figure 5.8 shows the VTFA for all the time windows, corresponding to Table 5.5. As shown in Figures 5.5 and 5.6, the predicted values, under most circumstances, correspond well to the actual values, although for occasional huge oscillations the forecast result is not very good. For both asset price volatility and trading volume volatility forecasts, an average of over 60% of VTFA can be achieved, substantiating the existence of firm correlations between information sentiment and volatility trend. The experiment also indicates that during the days when there are a large number of news items, incorporating information sentiment into the machine-learning model can noticeably improve volatility trend forecasting; an average of over

3 2.5 2 1.5 1 0.5 0

47

Week3 Week5 Week7 Week9 Week11 Week13 Week15 Week17 Week19 Week21 Week23 Week25 Week27 Week29 Week31 Week33 Week35 Week37 Week39 Week41 Week43 Week45 Week47 Week49 Week51

Price volatility

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

Time window MDT’s real price volatility MDT’s predicted price volatility

1.4 1.2 1 0.8 0.6 0.4 0.2 0

Price volatility forecast result for company MDT over all the time windows

Week3 Week5 Week7 Week9 Week11 Week13 Week15 Week17 Week19 Week21 Week23 Week25 Week27 Week29 Week31 Week33 Week35 Week37 Week39 Week41 Week43 Week45 Week47 Week49 Week51

Price volatility

Figure 5.6

Time window WAG’s real price volatility WAG’s predicted price volatility

Time window

Figure 5.8

Price volatility trend forecast accuracies for all the time windows

Week51

Week49

Week47

Week45

Week43

Week41

Week39

Week37

Week35

Week33

Week31

Week29

Week27

Week25

Week23

Week21

Week19

Week17

Week15

Week13

Week9

Week11

Week7

90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 20.00% 10.00% 0.00%

Week5

Price volatility forecast result for company WAG over all the time windows

Week3

Proportion of companies with accurately predicted price volatility trend

Figure 5.7

48

Enterprise Risk Management in Finance

70% of SCC was achieved for both price volatility and trading volume volatility forecasting, giving convincing evidence to the correlations between these factors. This experiment extends existing studies to a large set of companies. Empirical results are reflected by aggregated statistics, indicating the effects of information on entire stock markets. The results of this phase, although only focused upon US markets, provides a vivid description of the macro influences of financial news on financial volatility. It is of critical value to strategic decision-making by financial practitioners who seek to gain a panoramic picture of the general market.

Conclusions We have introduced GARCH-based ANN and SVM to investigate the correlations between financial trading volume volatility and online information volume. This enables effective prediction of financial risk in a network environment. Both methods are capable of achieving favorable prediction results; GARCHbased ANN performs better in predicting the volatility trend than GARCHbased SVM, while GARCH-based SVM outperforms GARCH-based ANN in forecasting the volatility itself. Moreover, online information is converted to sentiment values, which constitutes another key input element for the machinelearning models. Empirical studies indicate solid correlations between asset price volatility and information sentiment, which is well captured and stored by the SVM. Aggregated statistics show that a good forecast performance can be achieved by use of GARCH-based SVM method under sliding time windows. These empirical studies can be useful to financial investors, portfolio holders, academicians, etc. in the sense that they provide an alternative tool to forecast volatility and trend.

6

Online Stock Forum Sentiment Analysis

Introduction Over the past few decades, behavior finance and risk management have attracted a great deal of attention from both researchers and practitioners seeking to explain investor sentiment behavior, risk and loss perceptions, factors affecting investment strategy and investor behavior related to ongoing market trends.1 There is a great deal of unstructured data and information of public sentiments and opinions on market fluctuations posted on the internet today. This includes, for example, discussion forums, blogs, or message boards such as Facebook and Twitter, major sources of big data. Investors’ opinions and sentiments greatly impact on market volatility.2 This chapter develops a sentiment ontology for conducting context-sensitive sentiment analysis of online opinion posts in stock markets by integrating popular sentiment analysis into machine-learning approaches based on support vector machine (SVM) and generalized autoregressive conditional heteroskedasticity (GARCH)3 modeling.

Architectural design of GARCH-SVM based on sentiment index Figure 6.1 presents the conceptual flow chart for the methodology we use to conduct context-sensitive sentiment analysis of online opinion posts based on multiple sources of data. We first manually label sentiment polarity of a subset of postings. Then using sentiment analysis and these labeled postings, we identify features from written stock forum text to automatically predict sentiment polarity of other postings. We then aggregate postings for each stock on a daily basis. Thereafter we use an aggregated sentiment index and SVM classifiers to build GARCH-SVM4 models to predict future stock price volatility. As

49

50 Enterprise Risk Management in Finance

Data preprocessing

Cleaning

Internet Multisource data

Sentiment analysis

Sentiment aggregation

Sentiment time series

Segmenting Simulation, prediction

Trading, statement Dynamic training and forecasting Historical Time series

Figure 6.1

Time series of price, return, volatility

Econometrics, data mining (e.g. GARCH-SVM)

Conceptual modeling of sentiment for volatility forecast

can be seen from Figure 6.1, two sources of data are collected and incorporated into the model: sentiment-related data and historical financial time series. Sentiment analysis In market prediction, sentiment analysis technology is employed to automatically classify unstructured reviews as positive or negative, and then identify investor sentiment as either ‘bullish’ or ‘bearish.’ We consider two approaches for sentiment analysis: the machine-learning-based approach and the lexiconbased approach.5 In the machine-learning-based approach, an n-gram model is necessary for sentiment classification. The n-gram model takes characters (letters, spaces, or symbols) as the basic units. The advantages of the n-gram method are: (1) It is language-free, and can be applied to texts in English, traditional Chinese and simplified Chinese; (2) Linguistics processing, word segment and part-ofspeech tagging of the text are unnecessary; (3) It has good fault tolerance to identify spelling mistakes, and the requirement for extant knowledge of text is low; (4) Dictionaries and regulations are unnecessary. If the n-gram model is selected, classification accuracy would decline with the increase of order, i.e. 1-grams>2-grams>3-grams. We therefore first selected 1-grams as potential features from the training set, then manually adjusted the characters according to the following principle: When selecting 1-grams, there will be many punctuation marks (commas and periods) that contribute little to classification. Table 6.1 demonstrates a simple example. In order to improve classification accuracy, those punctuation marks should be eliminated manually. Figure 6.2 demonstrates the basic process of the lexicon approach, which includes pretreatment, word segmentation, POS (part of speech) tagging, polarity (‘bullish’ or ‘bearish’) tagging, combining and results output.

Online Stock Forum Sentiment Analysis

Table 6.1

Text, document

Figure 6.2

Pretreatment

51

Selecting 1-grams as features

File

1-grams

The large-cap stock is not bad.

The large-cap stock Is not bad

Word segmentation

POS tagging

Polarity Tagging

Combining

Categorized Text, document

The Lexicon approach for sentiment classification

We chose the ICTCLAS System, developed by the Institute of Computing Technology of the Chinese Academy of Sciences (http://ictclas.org/) for word segmentation, and POS (part of speech) tagging. Because adjectives play an important role in the sentiment classification of online reviews,6 we selected adjectives as the keywords. Meanwhile, in order to take account of the importance of the negative words, such as ‘no’ and ‘not,’ if negative words appeared they were given equal importance to adjectives. After word segmentation and POS tagging, each piece was decomposed and converted into a keyword array, each of which was assigned a specific sentiment value based on the HowNet word score. The overall sentiment for complete reviews was acquired by combining sentiment values of all their keywords. Keywords in the same sentences are grouped together. Then all the sentiment values of the keywords are added together; the sum represents the overall sentiment of the specific review. Data Both stock forum data and financial time series data are used in this study. The stock forum data utilized a corpus of stock reviews taken from a wellknown finance website, Sina Finance. • Sina Finance (http://finance.sina.com.cn/) was founded in 1999. It was the first choice of global Chinese financial portal because it has more than one third of the market share of Chinese financial web sites. • Customers registering on this platform can freely seek and share financial information on a daily basis.

52 Enterprise Risk Management in Finance

The financial dataset consisted of 50 military-sector stocks. • Stocks were chosen to focus on the military sector, because their stock forum showed a wide range of activity. • For a period of four months, from July to November 2012, we downloaded every message posted to these forums. We then employed our voting algorithm with the larger training set to assess each review and determine its sentiment. Figure 6.3 shows that 96.82% of the total reviews occurred during the working day, indicating high volume during trading days, and low activity during weekends and holidays. Figure 6.4 shows that most reviews occurred at 9 am, 10 am, and 4 pm, indicating that investors prefer to communicate with each other at the opening and closing stages of the stock market. At the same time, many reviews occurred at 7 pm, during non-working hours. This differs from the American stock market. We analyzed the content of the reviews and found that forecast reviews would always be updated after 3 pm. One possible reason is that the stock market in China closes at 3 pm, and investors will then make predictions about the future stock market. We chose the reviews updated after 3 pm on the trading day to create a corpus of online stock reviews on the whole stock market, from July 15, 2012 to November 15, 2012. This corpus was divided into two sets: a training corpus and a forecast corpus. We first manually labeled sentiment polarity (‘bullish’ or ‘bearish’) in the training corpus. Then, using sentiment analysis and the labeled training reviews, we identified features from the written text of the stock forum to automatically predict the sentiment polarity of the other reviews. When creating the forecast corpus, the number and date of the reviews were recorded.

600 500 400 300 200 100 0 1 Figure 6.3

2

3

4

5

Volume of reviews distributed by time of week

6

7

Online Stock Forum Sentiment Analysis

53

80 60 40 20 0 1 Figure 6.4

3

5

7

9

11

13

15

17

19

21

23

Volume of reviews distributed by time of day

Methodology comparison Classification performance for the machine-learning approach and the lexicon approach is compared. The resulting computation shows that the statistical machine learning approach has a classification accuracy of 81.82%, higher than that of the semantic approach, which has a classification accuracy of 75.58%. Classification accuracy is the degree of closeness of computed results to actual (true) value. Note that this comparison is statistically significant at the 0.95 level ( p value = 0.018). Table 6.2 shows that classification accuracy of the statistical machine-learning approach is reasonably robust with respect to the size of the training set when the size is more than 600. The p value = 0.018 suggests that this is true at the significance level of 95%. Reviews were classified by our algorithms into one of three types: bullish (optimistic), bearish (pessimistic), and neutral. Here the chi-square test shows that there are significant differences in classification accuracy between the semantic method and the statistical method. Because the statistical approach possessed higher accuracy when compared with the semantic approach, we opted for the statistical approach to assign labels for the reviews: bullish, bearish and neutral. When the size of training set was relatively small, classification accuracy could be improved by expanding the training set. But when training reviews were increased to a certain number, accuracy declined, and expanding the training set caused an increase in training time. Therefore, when choosing the size of training set, it is necessary to balance efficiency and accuracy. To achieve this balance, we first manually labeled about 30,000 reviews to three distinct sentiments, 1 for bullish, −1 for bearish, and 0 for neutral sentiment – referred to as ‘manual labels.’

54

Enterprise Risk Management in Finance

Table 6.2

Relative accuracies by sentiment Accuracy Sentiment

Age High-Low Positive Negative Firm size Positive Negative Price-to-book Positive Negative

Low (small)

Medium

High (large)

X2

p value

66.7 50 69.1 73.1 70.5 76.6 73.8 66.7 75

62.5 58.4 71.4 65.7 66.7 75 65.6 62.2 76.6

68.1 70.5 78.4 58.4 41.65 67.3 60.9 50 69.1

1.434

0.488

9.329

0.009

8.385

0.015

Sentiment and stock price volatility We analyzed the relationship of sentiment and stock price volatility trend at the individual stock level. Given that there appears to be a link from sentiment to the index at the aggregated level, we drilled down to the individual stock level to explore possible relations. Our analysis used the normalized stock price and normalized sentiment index for each stock. This normalization allowed us to stack up all the stocks together and then conduct the analysis using pooled data. We selected three kinds of stock features: age, firm size, and price-to-book value. The price-to-book ratio is a financial ratio used to compare a firm’s current market price to its book value. A higher price-to-book ratio usually implies that investors expect management to create more value from a given set of assets, all else equal. The price-to-book ratio can have greater discriminating power to identify value stocks and growth stocks. For each day, we formed ten equal-weighted portfolios according to the stock features of firm size (ME), age, and price-to-book value. We also calculated portfolio stock price volatility accuracy for all the individual stocks. Table 6.2 shows that we then reported average portfolio accuracy over days in which sentiment values from the previous day-end was positive, and days in which it was negative. Table 6.2 shows the following interesting patterns. In comparing the accuracy of aggregated bullish and bearish daily ticker data, we found no significant difference in prediction accuracy between young stocks and old stocks, given that the p value equals 0.488, which is greater than 0.05 at the significance level of 95%. One possible reason is that the Chinese stock market is relatively new, so market investors are not fully considering stock age as an evaluation factor for stock returns.

Online Stock Forum Sentiment Analysis

55

There were significant differences in prediction accuracy between small stocks and large ones. When the firm size was relatively small, sentiment sensitivity was high and accuracy was high, but when the scale of stock was relatively large, sentiment sensitivity was low and accuracy was low. This suggests that our sentiment-based GARCH-SVM approach works better for small companies, which are more sensitive than large ones to various online reviews. We also examined whether investor sentiment had a significant influence on value and growth stock returns. Based on data for the period July 1963 to December 2000, Siegel7 divided annual returns for stock classified into size quintiles (small cap to large cap) and price-to-book quintiles (value to growth). Growth stocks returned 6.41% per year compared with value stocks, which returned 23.28% for small cap stocks. Value stocks returned 13.59% and growth stocks returned 10.28% for large stocks. We obtained results consistent with those of previous studies.8 Our computational results in Table 6.2 suggest that there are significant differences in sentiment-oriented prediction accuracy between high price-to-book stocks (value stocks) and low price-to-book stocks (growth stocks). The model predicting the accuracy and sentiment sensitivity of the growth stocks was lower than that of value stocks: prediction accuracy and positive sentiment sensitivity were 73.8% and 66.7% respectively for value stocks, higher than the 60.9% and 50% for growth stocks. In comparing the accuracy of aggregated bullish and bearish daily ticker data, the results support that bearish labels have higher predictive accuracy than bullish labels. Furthermore, bullish sentiment was heavily influenced by the phenomenon of wishful thinking, reducing its predictive accuracy.

Conclusions There is a growing trend to use investor sentiment as an investment guide in the stock market. In the equity market when the market sentiment is bullish, most investors expect upward movement of stock prices; otherwise when the majority of investors expect downward movement for stock price, the market sentiment is called bearish. In the stock market, effective forecasting of financial risk measured by volatility is a critical activity, but is also very challenging. Therefore, accurate prediction of financial volatility is important. Although stock price volatility has been extensively discussed in the prior stock market literature, little published research considers the difference in the predictive power between bullish and bearish stock messages. We employed sentiment analysis technology based using both a machinelearning approach and a lexicon approach to automatically classify unstructured reviews as positive or negative, and then identify investor sentiment

56 Enterprise Risk Management in Finance

as either ‘bullish’ or ‘bearish.’ Empirical studies indicated solid correlations between stock price volatility trend and stock forum sentiment, which was well captured and stored by the SVM. Computational results demonstrated that the statistical machine-learning approach has a classification accuracy of 81.82%, which is higher than that of the semantic approach, with a classification accuracy of 75.58%, significant at the 95% level. Further analysis show that: • The proposed sentiment-based GARCH-SVM approach worked better for small companies, which are more sensitive than large companies to various online reviews. • Investor sentiment had a particularly strong effect for value stocks relative to growth stocks: the model predicting accuracy and positive sentiment sensitivity is 73.8% and 66.7% for value stocks respectively, higher than 60.9% and 50% for growth stocks.

7

DEA Risk Scoring Model of Internet Stocks

Introduction In financial markets, there are many kinds of investments, with stock the most popular. When investors choose which stock to invest in, they may expect high returns from investing in high performance companies. However, the greatest concern for investors is whether their investment has the potential for high returns, and whether the high performance companies will always yield high returns. Even after the dotcom collapse, US internet stock has remained a popular investment. However, investors are still concerned about future Internet Bubbles. Thus, the US internet stock market is a useful research focus with respect to financial performance. From an accounting perspective, the return on equity (ROE) ratio is an important indicator to measure the performance of a company, because the goal of a company is to maximum the stockholders’ equity. The DuPont model breaks ROE into three parts: profit margin, total asset turnover and financial leverage.1 It enables us to identify the existence of many indicators that influence the performance of a company. Hence, multiple indicators are considered. Data Envelopment Analysis (DEA) is a performance evaluation method capable of considering multiple inputs and multiple outputs. In this research, we aim to formulate an evaluation process combining the DEA method with the concept of ROE. Investors can use this as a stock selection method, and managers can use it for performance evaluation.

Different methods of performance evaluation Performance evaluation considers a number of attributes (or criteria) and covers multiple levels. Items chosen for evaluating performance include both quantifiable and non-quantifiable indicators. These may be mutually exclusive, 57

58 Enterprise Risk Management in Finance

related or independent of each other. In addition, the problems being faced are extremely complex and unpredictable. A number of techniques have been proposed. Objectivity, fairness and feasibility are crucial for performance evaluation. This study reviews seven methods applicable to the evaluation of performance. They are (1) Multivariate Statistical Analysis,2 (2) Data Envelopment Analysis,3 (3) Analytic Hierarchy Process,4 (4) Fuzzy Set Theory,5 (5) Grey Relation Analysis,6 (6) Balanced Scorecard,7 and (7) Financial Statement Analysis.8 The fundamental theories of the seven methods, and their advantages and disadvantages when applied to performance evaluation, are described in detail below: Multivariate statistical analysis A statistical method used to quantify complex issues or events and to arrange them systematically for the purpose of classification, inference, evaluation and forecast. Strengths:

i. It is based on traditional methods of statistics, with solid theoretical foundation. ii. The system is complete and could be applied in almost all areas of research. Weaknesses:

i. It requires a large sample size and normal distribution. ii. Methods not including statistical testing cannot be used systematically, which hampers further interpretation of the results. Data envelopment analysis Based on the concept of Pareto optimality. When measuring the efficiency value of DMU, only the production margin is required. The production margin would then be compared with actual production to calculate the efficiency values. Strengths:

i. DEA could be used to handle problems with multiple inputs and outputs. ii. It would not be influenced by different scales. iii. The results of DEA evaluation on efficiency is a composite indicator, and could be used to apply the concept of total production factors in economics. iv. The weighted value in the DEA model is the product of mathematical calculation, and hence free from human subjectivity. v. DEA can deal with interval data as well as ordinal data.

DEA Risk Scoring Model of Internet Stocks

59

vi. The results of the evaluation by DEA could provide more information on the data used, which could be used as a reference in the decision-making process. Weaknesses:

i. ii. iii. iv.

It yields the efficient frontier, which may be quite large. If the sample size is too small, the outcome is less reliable. There should not be too many variables. The degree of relation between the input and output variables (indicators) is not considered.

Analytic hierarchy process An approach to quantify subjective estimates. Complex and non-systematic issues are treated systematically in a stepwise process, yielding weighted value of options (indicators). Strengths:

i. ii. iii. iv.

Easy to apply. The results are subject to consistency checking. Solid theoretical foundation and is objective. Easier to handle qualitative problems.

Weaknesses:

i. When there are great differences across experts, diverse results yield little value. ii. Fails to discuss the relation between factors (indicators). Fuzzy set theory Provides an overall evaluation on events or phenomenon influenced by a number of factors, by way of building up of subordinate functions. Accordingly, the qualitative and quantitative values of the indicators would be interchangeable, and a value in real numbers would be assigned to each factor under evaluation. Priority would then be assessed. Strengths:

i. It can deal with a large number of uncertain problems. ii. Since it is a simulation of human thought and decision processing, it is compatible with human behavior.

60

Enterprise Risk Management in Finance

Weaknesses:

i. The degree of subordination is indicated by a value between 0 and 1, so the results of evaluation would be subject to influence by the choice over subordination function. ii. The relation between variables (indicators) is not discussed. Grey relation analysis Based on the homogeneity or heterogeneity of the trend development of factors to find out if there is grey relation between two indicators and, if so, the extent of this relationship. Strengths:

i. No rigid requirement for sample size. ii. Can still be applied when the distribution of data is uncertain. iii. Is based on data analysis, and is free from traditional subjectivity in decision making. iv. The method of calculation is simple and easy to apply. Weaknesses:

i. Cannot directly handle qualitative issues (is non-quantifiable). ii. The criteria for choosing grey relation coefficient value would directly affect the final evaluation result. Balanced scorecard A performance evaluation system containing four components for evaluation. This is also called a strategic management system, which could help firms translate strategy into actions. The four components are finance, customer, internal process, and learning and growth. Strengths:

i. Can integrate information, and put various key factors for the success of the organization into a single report. ii. Avoids information overload, since the indicators used for performance measurement are the key indicators. Weaknesses:

i. The procedure for the application of BSC is complex and time consuming. Financial statement analysis People use this approach with the belief that the result of business activities of the firm would be reflected in its financial statement

DEA Risk Scoring Model of Internet Stocks

61

Strengths:

i. Objective: It is the reflection of actual events. ii. Concrete: All data in the financial statement can be quantified. iii. Measurable: Since the data in the financial statement can be quantified, they are measurable. Weaknesses:

i. There is no criterion for selecting a ratio that is agreeable by all users. ii. The figures in the financial statement have been added or simplified, and cannot satisfy the needs of all users. iii. The financial statement cannot express qualitative information, such as ability, morale, potential and trust. Each of the above seven methods can be independently applied to evaluating performance. However, none of them is perfect. There is a saying, ‘Whenever there is an advantage, it entails a drawback.’ Researchers can only choose a method to evaluate performance that has the least number of drawbacks for that study’s particular situation. In contrast to other approaches such as AHP, Multivariate Statistical Analysis and Grey Relation Analysis, DEA requires little assumption about a functional form among variables; no prior information on weight assigned to input/output variables is required. Thus, DEA provides a very good tool to objectively gauge the DMU performance. DEA has been widely used to yield new insights into activities (and entities) previously evaluated by other methods such as TOPSIS and fuzzy methods.9

Basics of data envelopment analysis DEA is a non-parametric approach to build an efficiency frontier to measure relative efficiency for a set of homogeneous decision-making units (DMUs) between multiple inputs and outputs. The theory of DEA can be traced back to Farrell, who proposed using a production frontier to evaluate the technical efficiency. He divides efficiency into overall efficiency (OE) (or economic efficiency), technical efficiency (TE), and allocative efficiency (AE). Overall efficiency is composed of technical efficiency and allocative efficiency. Technical efficiency shows the maximum products that factories can produce, given their specific inputs. Allocative efficiency shows the inputs that enterprises should put into, given their specific price and product technology. If we multiply technical efficiency by allocative efficiency, we get economic efficiency. That is: OE = AE × TE

62 Enterprise Risk Management in Finance

This is the efficiency measuring model that Farrell proposed in 1957.10 There are two major DEA models: One is the CCR model proposed by Charnes, Cooper and Rhodes,11 and the other is the BCC model proposed by Banker, Charnes and Cooper, allowing variable returns to scale.12 The CCR model is inputoriented, and the BCC is output-oriented. In this research, the output-oriented BCC model is adopted because the variable returns to scale assumption is more realistic, and the goal of companies is to maximize their outputs. Min hs =

m

∑V X j

js

+ Ds

j =1

p

Subject to

∑U Y

k ks

=1

k =1

p

m

∑ U Y − ∑V X k ki

k =1

j

ji

− Di ≤ 0 , i = 1,2,. . . n

j =1

Vi ≥ 0, j = 1,2, . . . m; Uk ≥ 0, k = 1,2,. . . p Di is a constant and we can use Di as the index of the return scale of the DMU. The standard is as follows: Di > 0 → DMU is under decreasing returns to scale Di = 0 → DMU is under constant returns to scale Di < 0 → DMU is under increasing returns to scale. Using duality theory and the slack variable to transform the equation, we get: p  m  − + Max H s + ε ∑ SVjs + ∑ SVks  k =1  j =1  n

Subject to H sYks − ∑ Yki λ i + SVks+ = 0 i =1

n

X js − ∑ X ji λ i − SVjs− = 0 i =1

n

∑λ

i

=1

i =1

+

λ i ≥ 0 , i = 1,2, . . . n; SVjs− ≥ 0 , j = 1,2, . . . m; SV kS ≥ 0, k = 1,2,. . . p

DEA Risk Scoring Model of Internet Stocks

63

We can see that compared with the CCR model, the BCC model adds the limitation of

n

∑l

i

= 1, to make sure that the production frontier will be raised to

i =1

the origin.

The proposed approach DEA can deal with multiple inputs and outputs simultaneously, and DEA models are broadly used in many fields. DEA is believed to be one of the most commonly used approaches to measure company performance in the financial industry. In this section, we propose a model to combine DEA models with a financial analysis tool to evaluate efficiency of online companies. Financial ratio analysis has been the standard technique used in economics to examine business and managerial performances.13 Due to its simplicity and ease of understanding, the analytical ratio measure has been widely applied in many areas such as in financial investment and insurance industries. Two of the most preferred analytical ratios are return on equity (ROE) and return on assets (ROA), both providing insight into a financial institution that allows management to make strategic decisions that can dramatically affect its structure and profitability. ROA is defined as the ratio of net income divided by total assets, and estimates how efficient we are at earning returns per dollar of assets. ROA has been merged into DEA to evaluate efficiency and effectiveness of an organization.14 ROE is calculated by dividing net income by average equity, and identifies how efficiently we use our invested capital. Companies that boast a high ROE with little or no debt are able to grow without large capital expenditures, allowing the owners of the business to withdraw cash and reinvest it elsewhere. ROE is just as comprehensive as ROA, and could be a better indicator than ROA in terms of identifying a firm’s profitability and potential growth, that is, the potential risk that a firm can take. Moreover, from the accounting perspective, when we use the ROA ratio to measure company performance, the ROE ratio has to be used simultaneously to see whether a high ROE ratio arises from financial leverage, or whether the company ROA is high. So in this research, we use the ROE ratio to make the evaluation process more complete. ROA =

Net Income Assets

ROE =

Net Income Equity

Many investors fail to realize, however, that two companies can have the same return on equity, yet one can be a much better business. The DuPont model

64

Enterprise Risk Management in Finance

provides a tool to decompose ROE into three elements in the calculation of ROE: The net profit margin, the asset turnover, and the equity multiplier. By examining each input individually, we can discover the sources of a company’s return on equity and compare it to its competitors. Using the DuPont model, the ROE ratio can be decomposed into:

ROE =

Net Income Sales Assets × × Sales Assets Equity (Effectiveness) (Efficiency) (Equity multiple)

= ROA × Equity multiple Asset turnover (efficiency) is used to measure the ability of the firm to use its assets, and can be deemed as the operational efficiency of a company. Profit margin (effectiveness) is used to diagnose the effectiveness of a company; it measures not only the competitiveness of the product but the expense control ability of a company. The equity multiple can be used to understand the capital structure of a company, and companies can use financial leverage to control their capital structure. So investors should take a higher risk to gain high returns. Hence, we use the concept of ROE to test whether investing in companies with high financial performance can get high returns or not, because it is important to think of return and risk simultaneously when choosing an investment target. Based on the concept of measuring firm performance by efficiency and effectiveness, this research adopts the two-stage DEA model15 to evaluate the performance of online companies. The two-stage DEA approach is shown in Figure 7.1 The risk-scoring model is depicted in Figure 7.2, to include two sub-processes. In contrast to Figure 7.1, Figure 7.2 introduces another process to understand whether investing in companies with high financial performance can get high returns or not. These two processes can be used to evaluate the company from both enterprise (company performance) and investor (the returns per unit of risk available) perspectives; hence this evaluation process can give investors and managers more accurate criteria to make decisions. Variable selection In order to measure DMU efficiency, the selection of the input variables and output variables is very important. In the internet industry, existing literature defines a good set of variables to measure online company performance, including both financial data and non-financial data.

DEA Risk Scoring Model of Internet Stocks

Intermediate variable

Output variable

Stage 2

Stage 1 Figure 7.1

Effectiveness

Efficiency

Input variable

65

Two-stage DEA model Process I Intermediate variable

Stage 1

Input variable

Effectiveness

Efficiency

Input variable

Output variable

Stage 2

Return level per unit of risk

Output variable

Process II Figure 7.2

Proposed evaluation process

Some criteria are set up in order to facilitate the input and output selection as follows: 1. The variables adopted by a paper measuring the internet industry using the DEA approach can be considered. 2. There are limited papers measuring internet industry performance using the DEA approach, so a two-stage DEA approach will be used for measuring operating efficiency and effectiveness. 3. For measuring the investing risk, there is only one paper using the DEA approach to measure the relationship between return and risk. Hence, the variables adopted by that paper have been considered. 4. All the variables must conform to the ROE concept. We based our variable selection on various financial measures employed in the literature for evaluation of the efficiency of a financial institution.16 In Evaluation Process 1, we chose operating expense, employees, total assets,

66

Enterprise Risk Management in Finance

revenue, gross profit, EPS and net income as the main variables to measure the performance of online companies based on the literature review and suggestions by experts. In order to measure efficiency, we used total assets, total equity and operating expense as input variables to measure how much money they can earn (revenue) and how many profits (gross profit) they can generate. Total assets was chosen as an input because it is the sum of intangible asset, current asset, and fixed asset – and the intangible asset which is shown in the balance sheet as a result of a merger or takeover is hard to measure. The variable of operating expense was chosen as an input because many internet companies do not report the number of marketing expense or R&D expense. In order to measure effectiveness, we used revenue and gross profits as input variables, to see how well companies controlled their expenses to generate money (net income) and how much income was shared with stockholders (EPS). In Evaluation Process 2, we choose Beta, book value to market value (BV/ MV), and rate of return as the main variables to measure how much return a company can generate given the same risk. Beta was chosen as an input because investing risk can be divided into systematic risk and non-systematic risk. For investors, the non-systematic risk can be dispersed by diversification effect. So, if systematic risk is the only concern, then the Beta coefficient is a better way to measure the risk. BV/MV was chosen as an input because it is also a commonly used ratio to measure investing risk.

Empirical study The sample for this study includes listed online companies in the United States. There are 127 such companies in NASDAQ categories, and they are grouped into three categories: internet service providers, internet information providers, and internet software and services providers. Because the data for DMUs cannot be negative, the data of DMUs could not be collected, and because some accounting periods were different across companies, we chose only 27 listed online companies. Data was collected from Yahoo! Finance (http://finance.yahoo.com/) and EDGAR Online (http://edgar.brand.edgaronline.com/default.aspx) for 2006. The DEA result The DEA efficiency scores are a percentage value which varies between 0% and 100%. If the efficiency score is equal to 100%, then the score is the best efficiency and hence the unit is the most efficient unit. The DEA result for Evaluation Process 1 is shown in Table 7.1. We can see that 10 out of 27 listed online companies, namely GOOG, TZOO, JCOM, AMZN, UNTD, DTAS, ADAM, EGOV, EBAY, and RATE, are BCC-efficient on operating efficiency. There are 6 out of 27 listed online companies, namely

DEA Risk Scoring Model of Internet Stocks

Table 7.1

67

BCC-efficient scores on performance

Company name

Stock code

Google, Inc. A.D.A.M., Inc. j2 Global Communications, Inc. eBay, Inc. Bankrate, Inc. Travelzoo, Inc. Digitas, Inc. NIC Inc. United Online, Inc. Amazon.com, Inc. Varsity Group, Inc. Citrix Systems, Inc. Sabre Corp Websense, Inc. Aquantive, Inc. Digital River, Inc. IAC/InterActive Corp Sabre Corp Sohu.com Inc. Priceline.Com, Inc eCollege.com Open Solutions, Inc. Sina Corp iPass, Inc. Online Resources Corp Corillian Corp SupportSoft, Inc.

GOOG ADAM JCOM EBAY RATE TZOO DTAS EGOV UNTD AMZN VSTY CTXS YHOO WBSN AQNT DRIV IACI TSG SOHU PCLN ECLG OPEN SINA IPAS ORCC CORI SPRT

Operating efficiency score 100 100 100 100 100 100 100 100 100 100 99.03 93.81 91.64 90.67 89.54 86.64 84.27 82.46 82.12 81.63 79.94 75.56 72.87 71.68 68.58 67.17 54.22

Effectiveness score 100 100 100 65.88 64.24 50.54 40.03 35.16 34.96 30.44 93.39 56.05 100 76.26 36.69 96.26 67.88 32.18 79.56 100 23.92 59.98 67.65 20.92 100 13.37 17.19

ADAM, ORCC, JCOM, PCLN, GOOG, and YHOO, which are BCC-efficient on effectiveness. Moreover, of the ten companies that are BCC-efficient on operating efficiency, seven companies are not BCC-efficient on effectiveness. These are TZOO, AMZN, UNTD, DTAS, EGOV, EBAY, and RATE. These seven companies can use their resources to generate profits very well; however, they cannot use their profits to generate income very well. This may be because when comparing to the companies that are BCC-efficient on both dimensions, these seven companies do not control their expenses well, or they issue more stock so that EPS becomes diluted. On the other hand, of the six companies that are BCC-efficient on effectiveness, three are not BCC-efficient on operating efficiency. They are ORCC, PCLN, and YHOO. These companies can use their profits to generate income

68

Enterprise Risk Management in Finance

very well; however, they cannot use their resources to generate profits very well. This may be because when compared with the companies that are BCC-efficient on both dimensions, these companies spend more costs on their products, or the sales volume or price is lower than market, or their capital has mainly come from inside (stockholder’s equity), not borrowed from outside (liabilities). The DEA efficiency scores are a percentage value which varies between 0% and 100%. In order to see the total efficiency of each company, we time the efficiency scores of operating efficiency and effectiveness. The BCC efficiency scores are also a percentage value between 0% and 100%. It can be observed that only three companies perform best in both dimensions, showing that they are BCC-efficient in both operating efficiency and effectiveness, which are ADAM, GOOG and JCOM. These three companies can use their resources to generate profits, as well as using their profits to generate income. The DEA result for Evaluation Process 2 is shown in Table 7.2.

Table 7.2 BCC-efficient scores on the level of returns per unit of risk Company name

Stock code

j2 Global Communications, Inc. Priceline.com, Inc. Varsity Group, Inc. A.D.A.M., Inc. Online Resources Corp Yahoo!, Inc. Google, Inc. Digital River, Inc. Sohu.com, Inc. Websense, Inc. Bankrate, Inc. Travelzoo, Inc. Amazon.com, Inc. Aquantive, Inc. Citrix Systems, Inc. United Online, Inc. Sina Corp eBay, Inc. NIC, Inc. Digitas, Inc. Sabre Corp Open Solutions, Inc. IAC/InterActiveCorp eCollege.com Corillian Corp iPass, Inc. SupportSoft, Inc.

JCOM PCLN VSTY ADAM ORCC YHOO GOOG DRIV SOHU WBSN RATE TZOO AMZN AQNT CTXS UNTD SINA EBAY EGOV DTAS TSG OPEN IACI ECLG CORI IPAS SPRT

Risk score 100 100 100 100 60.89 45.93 42.39 39.62 37.20 30.88 30.25 26.70 24.11 23.62 23.27 22.48 22.40 19.20 18.92 15.67 15.67 14.76 10.94 10.24 9.04 8.86 5.02

DEA Risk Scoring Model of Internet Stocks

69

We can see there are 4 out of 27 listed online companies, namely JCOM, PCLN, VSTY, and ADAM, which are BCC-efficient on investing risk. These four companies have the highest level of returns per unit of risk, which means that if investors choose them, they can expect the highest level of returns. There is a trend between total efficiency and investing risk. Table 7.3 shows the total efficiency and risk for the top ten and bottom four companies. If investors choose companies with high scores in total efficiency, they can enjoy a higher level of returns. Otherwise, if the investors choose the companies with low scores in total efficiency, they will get a lower level of returns. Moreover, through the comparison in Table 7.4, we can see the companies that are BCC-efficient on operating efficiency but not BCC-efficient on effectiveness will perform worse in total efficiency. For example, TZOO, AMZN, UNTD, DTAS, EGOV, EBAY, and RATE are BCC-efficient on operating efficiency, but their performance on effectiveness is bad, and so the total efficiency of these companies will be lower. On the other hand, the companies that are BCC-efficient on effectiveness but not on operating efficiency will perform better in total efficiency than the companies that are BCC-efficient on operating efficiency but not on effectiveness. For example, ORCC,PCLN, and YHOO are BCC-efficient on effectiveness, but their performance on operating efficiency is not good. However, these companies have good scores in total efficiency. So, we can see that the main dimension that influences the total efficiency of a company is its effectiveness. In the internet industry, the effectiveness of a company is more important than its operating efficiency.

Table 7.3 Ranking of the BCC-efficient scores of total efficiency and investing risk Rank

Total efficiency

Investing risk

1 2 3 4 5 6 7 8 9 10 24 25 26 27

ADAM JCOM GOOG VSTY YHOO DRIV PCLN WBSN ORCC EBAY ECLG IPAS SPRT CORI

ADAM JCOM PCLN VSTY ORCC YHOO GOOG DRIV SOHU WBSN ECLG CORI IPAS SPRT

70 Enterprise Risk Management in Finance

Table 7.4

Ranking of the BCC-efficient scores of whole model

Rank

Operating efficiency

Marketability

Total efficiency

Investing risk

1 2 3 4 5 6 7 8 9 10 24 25 26 27

ADAM JCOM GOOG AMZN UNTD DTAS TZOO EGOV EBAY RATE IPAS ORCC CORI SPRT

ADAM JCOM GOOG PCLN ORCC YHOO DRIV VSTY SOHU WBSN ECLG IPAS SPRT CORI

ADAM JCOM GOOG VSTY YHOO DRIV PCLN WBSN ORCC EBAY ECLG IPAS SPRT CORI

ADAM JCOM PCLN VSTY ORCC YHOO GOOG DRIV SOHU WBSN ECLG CORI IPAS SPRT

There are only 2 out of the 27 listed online companies, namely ADAM and JCOM, that are BCC-efficient on operating efficiency, effectiveness and investing risk. These companies operate well, and investors can get the highest returns by choosing them. GOOG is BCC-efficient on operating efficiency and marketability; however, its efficiency score on investing risk is low. This means that GOOG can use its resources to generate profits, as well as using its profits to generate income, but investors cannot get the high returns they hope for by investing in GOOG. On the other hand, there are 3 of the 27 listed online companies, namely IPAS, CORI, and SPRT, which are BCC-inefficient on operating efficiency, effectiveness and investing risk. These companies operate their company inefficiently, and investors get the lowest returns by choosing them.

Conclusions This paper proposed a new performance measurement model, combining the ROE and DEA approach, for both investors and managers. From empirical study, we draw the following conclusions: 1. The main contribution of this study is to propose an accurate evaluation process combining the ROE concept with the DEA model. The evaluation process not only considers the performance of a company, but also considers the return to investors as the whole performance to measure a company. For investors, this model can be used to be the stock selecting strategy; based on the relative score of each DMU, investors can easily rank the priority of

DEA Risk Scoring Model of Internet Stocks

71

the stocks. And for managers, this model can be used to be the performance measurement model; based on the relative score of each DMU, the managers can know the position where their company and also their competitors stand, and they can know in each dimension whether their company is performing well or needs to be improved. 2. Based on the research results, the main dimensional influence for a company’s total efficiency is effectiveness. For the internet industry, the company’s effectiveness is more important than its operating efficiency in influencing performance. Hence, investors may focus on the net income and EPS when they want to measure the US internet stock market. For managers, most of the internet companies perform well on operating efficiency, which means the companies can use their resources to generate profit well. However, most of the internet companies have lower efficiency scores on effectiveness. So companies should control their expenses, hence their net income will be higher. 3. If investors choose companies with high scores in operating performance, they should gain a higher level of returns. If the investors choose companies with low scores in operating performance, they are likely to gain a lower level of returns. As with any study, this research is not without limitations. Three limitations are noted: 1. The DEA model cannot use negative numbers.17 However, the listed online companies can have negative net income or equity. In order to compare these listed online companies on the same basis, companies with different financial periods were excluded, leaving only 27 companies. The number of decision making units (DMUs) is greater than the twice of the summation of the number of input and output variables, so the sample size used in this study still complies with the requirement of DEA approach. 2. Non-financial data are not included, which is also an important dimension to measure online companies. This is because some data cannot be measured or may be confidential. 3. There are few previous research reports using DEA to evaluate the performance of the internet industry, resulting in the lack of theoretical backup in the input and output variable selection. 4. DEA can be combined with classical risk management such as value-at-risk18 to develop new methodologies for optimizing risk management.19

8

Bank Credit Scoring

Introduction The concept of enterprise risk management (ERM) developed in the mid-1990s in industry, expressing a managerial focus. ERM is a systematic, integrated approach to managing all risks facing an organization.1 It has been encouraged by traumatic recent events such as 9/11/2001 and business scandals, including Enron and WorldCom. Consideration of risk has always been with business, manifesting itself in 17th-century coffee houses such as Lloyd’s of London spreading risk related to cargoes on the high seas. Businesses exist to cope with specific risks efficiently. Uncertainty creates opportunities for businesses to make profits. Outsourcing can offer many benefits, but also has a high level of inherent risk, and ERM seeks to provide means to recognize and mitigate risks. The field of insurance developed to cover a wide variety of risks, both external and internal, covering natural catastrophes, accidents, human error, and even fraud. Financial risk has been controlled through hedge funds and other tools over the years, often by investment banks. With time, it was realized that many risks could be prevented, or their impact reduced, through loss-prevention and control systems, leading to a broader view of risk management. The subprime crisis made companies become increasingly strict about the effective functions of ERM. The failure of the credit rating mechanism troubles companies that need timely signals about the underlying risks of their financial assets. Recent development in major credit ratings agencies such as Standard & Poor’s (S&P) and Moody’s have integrated ERM as an element of their overall analysis of corporate creditworthiness.

Risk modeling It is essential to use models to handle risk in enterprises. Risk-tackling models can be (1) an analytical method for valuing instruments, measuring risk 72

Bank Credit Scoring

73

and/or attributing regulatory or economic capital; (2) an advanced or complex statistical or econometric method for parameter estimation or calibration used in the above; or (3) a statistical or analytical method for credit risk rating or scoring. Risk management tools can include creative risk financing solutions, blending financial, insurance and capital market strategies (AIG).2 Capital market instruments include catastrophe bonds, risk exchange swaps, derivatives/options, catastrophe equity puts (cat-e-puts), contingent surplus notes, collateralized debt obligations, and weather derivatives. Many risk studies in banking involving analytic (quantitative) models have been presented. Crouhy et al. (1998) provided comparative analysis of such models.3 Value-at-risk models have been popular,4 partially in response to the Basel II banking guidelines. Other analytic approaches include the simulation of internal risk rating systems using past data. Jacobson et al. found that Swedish banks used credit rating categories, and that each bank reflected its own risk policy.5 One bank was found to have a higher level of defaults but without adversely affecting profitability, due to its constraining high-risk loans to low amounts. Elsinger et al. (2006) examined the systemic risk from overall economic systems as well as the risk from networks of banks with linked loan portfolios.6 The overall economic system risk was found to be much more likely, while linked loan portfolios involved high impact but very low probability of default. The use of scorecards has been popularized by Kaplan and Norton (1992)7 in their balanced scorecard, as well as other similar efforts to measure performance on multiple attributes.8 In the Kaplan and Norton framework, four perspectives are used, each with possible goals and measures specific to each organization. Table 8.1 demonstrates this concept in the context of bank risk management: This framework of measures was proposed as a means to link intangible assets to value creation for shareholders. Scorecards provide a focus on strategic objectives (goals) and measures, and have been applied in many businesses and governmental organizations with reported success. Papalexandris et al. (2005)9 and Calandro and Lane (2006)10 both have proposed use of balanced scorecards in the context of risk management. Specific applications to finance,11 homeland security,12 and auditing13 have been proposed. Model risk pertains to the risk that models are either incorrectly implemented (with errors) or that make use of questionable assumptions or assumptions that no longer hold in a particular context. It is the responsibility of the executive management in charge of areas that develop and/or use models to determine what models this policy applies to. Lhabitant (2000)14 summarized a series of cases where model risk led to large banking losses. These models vary from a trading model in pricing stripped

74 Enterprise Risk Management in Finance

Table 8.1

Balanced scorecard perspectives, goals, and measures

Perspectives

Goals

Measures

FINANCIAL

Survive Succeed

Cash flow Quarterly sales, growth, operating income by division Increase in market share, Increase in Return on Equity

Prosper CUSTOMER

New products Responsive supply Preferred suppliers Customer partnerships

INTERNAL BUSINESS

Technology capability Manufacturing excellence Design productivity New product innovation

INNOVATION & LEARNING

Technology leadership Manufacturing learning Product focus Time to market

% sales from new products, % sales from proprietary products On-time delivery (customer definition) Share of key accounts’ purchases, ranking by key accounts # of cooperative engineering efforts Benchmark vs. competition Cycle time, unit cost, yield Silicon efficiency, engineering efficiency Schedule: actual vs. planned Time to develop next generation Process time to maturity % products equaling 80% of sales New product introduction vs. competition

mortgage-backed securities, to risk and capital models in deciding on the structured securities, to decision models in issuing a gold card. Table 8.2 summarizes some model risk events in banking. Sources of model risk arise from the incorrect implementation and/or use of a performing model (one with good predictive power) or the correct implementation/use of a non-performing model (one with poor predictive power). To address these risks, vetting of a statistical model is comprised of two main components: vetting and validation. Vetting focuses on analytic model components, includes a methodology review, and verifies any implementation, while validation follows vetting and is an ongoing systematic process to evaluate model performance and to demonstrate that the final outputs of the model are suitable for the intended business purpose.

Performance validation in credit rating Performance validation/back testing focuses on credit rating in two key aspects: discriminatory power (risk discrimination) and predictive accuracy (model calibration). Discriminatory power generally focuses on the model’s ability to

Bank Credit Scoring

Table 8.2

Model Model risk

75

Model risk events in banking Trading and position management models

Decision models in retail banking

Risk and capital models

Booking with a model that does not incorporate all features of the deal; booking with an unvetted or incorrect model; incorrect estimation of model inputs (parameters); incorrect calibration of the model; etc.

Incorrect statistical projections of loss, making an incorrect decision (e.g. lending decision) or incorrectly calculating and reporting the Bank’s risk (e.g. default and loss estimation) as a result of an inadequate model, etc.

Use of an unvetted or incorrect model; poor or incorrect estimation of model parameters; testing limitations due to a lack of historic data; weak or missing change control processes, etc.

rank-order risk, while predictive accuracy focuses on the model’s ability to predict outcomes accurately (e.g. probability of defaults, loss given defaults, etc.). Various statistic measures can be used to test the discriminatory power and predictive accuracy of a model. Commonly used measures in credit rating include the divergence, Lorenz curve/CAP curve and the Kolmogorov-Smirnov (KS) statistic.15

Case study: credit scorecard validation The predictive scorecard that is currently being used in a large Ontario bank is validated. This bank has a network with a total of more than 8000 branches and 14,000 ATM machines operating across Canada. This bank successfully conducted a merger of two brilliant financial institutions in 2000, and became Canada’s leading retail banking organization. It has also become one of the top three online financial service provider by providing online services to more than 2.5 million online customers. The scorecard system used in retail banking strategy will then need to be validated immediately due to this merger event. The scorecard system under evaluation predicts the likelihood that a 60–120-day delinquent account (mainly on personal secured and unsecured loans and lines of credit) will be cured within the subsequent three months. By breaking up funded accounts into three samples based on their limit issue date, the model’s ability to rank order accounts based on creditworthiness was validated for individual samples and compared to the credit bureau score. Tables 8.3, 8.4, and 8.5 give the sample size, mean, and standard deviation of these three samples: Sample 1 involves accounts from January 1999 to June 1999, Sample 2 from July 1999 to December 1999, and Sample 3 from January 2000

76 Enterprise Risk Management in Finance

Table 8.3

Scorecard performance validation, January–June 1999

Scorecard Beacon Good N Mean Std. Dev

Scorecard Beacon/ (No Bureau Application Bureau score) alone alone Empirical

26783 250 24

25945 734 55

26110 734 55

673 222 22

26783 42 9

26783 208 21

Bad

N Mean Std. Dev

317 228 23

292 685 55

296 685 55

21 204 13

317 40 9

317 188 22

Total

N Mean Std. Dev

27100 249 24

26.237 733 55

26406 733 55

694 221 22

27100 42 9

27100 207 21

KS%

KS Value Score

39 240

37 733

37 735

44 215

14 45

36 204

0.869 1.17

0.792 1.11

0.790 1.12

0.877 3.03

0.070 1.17

0.814 1.17

Divergence Bad%

Table 8.4

Scorecard performance validation, July–December 1999

Scorecard Beacon Good N Mean Std. Dev

Scorecard Beacon/ (No Bureau Application Bureau score) alone alone Empirical

20,849 248 24

20,214 728 54

20,302 728 54

547 222 23

20,849 42 9

20,849 206 21

Bad

N Mean Std. Dev

307 231 23

296 691 55

297 692 55

10 208 12

307 40 9

307 191 22

Total

N Mean Std. Dev

21,256 248 24

20,510 727 54

20,599 727 54

557 222 22

21,156 42 9

21,156 206 21

KS%

KS value Score

33 246

26 731

26 731

42 216

10 43

29 200

0.528 1.45

0.450 1.44

0.435 1.44

0.624 1.80

0.040 1.45

0.498 1.45

Divergence Bad%

to June 2000. Cases of 90 days’ delinquency or worse, and accounts that were closed with a ‘NA (non-accrual)’ status or that were written off were included as bad performance. Good cases were defined as those that did not meet the definition of ‘bad.’ The ‘bad’ definition is evaluated at 18 months. Three samples of cohorts were created and compared. Specified time periods refer to month-end dates. For the performance analyses, the limit issue dates were considered,

Bank Credit Scoring

Table 8.5

77

Scorecard performance validation, January–June 2000

Scorecard Beacon Good N Mean Std. Dev

Scorecard (No Bureau Application Bureau Beacon/ score) alone alone Empirical

23,941 246 24

23,254 723 54

23,361 723 54

580 223 21

23,941 41 9

23,941 205 21

Bad

N Mean Std. Dev

533 225 21

490 683 51

495 683 51

38 216 16

533 38 9

533 187 20

Total

N Mean Std. Dev

24,474 245 24

23,744 723 54

23,856 723 54

618 222 20

24,474 41 9

24,474 204 21

KS%

KS value Score

38 239

33 704

33 704

26 215

14 43

34 202

0.843 2.18

0.606 2.05

0.598 2.07

0.147 6.15

0.078 2.18

0.789 2.18

Divergence Bad%

while the population analyses used the application dates. In order to validate the relative effectiveness of the scorecard, we conducted statistical analysis, and reported results for the following statistical measures: divergence test, Lorenz curve, Kolmogorov–Smirnov (K–S) test, and population stability index. Statistical results and discussion Tables 8.3, 8.4 and 8.5 document scorecard performance validation for January–June 1999, July–December 1999 and January–June 2000 respectively. Table 8.6 presents the summary analysis for performance samples. The numbers in these tables are rounded off to demonstrate the nature of score values assigned to different customer accounts. This also helps prevent revealing the bank’s business details, for security. Using the rounded number values, we can easily compute the divergence values close to those in the last row of each table. For example, relating to the scorecard in Table 8.3: Mean Good = 250, Std. Dev. Good = 24, Mean Bad = 228, Std. Dev. Bad = 23. The difference (Mean Good – Mean Bad) is equal to 22, and the average variances sum is equal to 552.5. The divergence is the fraction 484/552.5 = 0.876, which is very close to the 0.869 based on non-rounded values given in Table 8.3. The corresponding Lorenz curves are depicted in Figures 8.1, 8.2, 8.3 and 8.4. Note that all figures are based on best applicant. In a dataset that has been sorted by the scores in ascending order, with a low score corresponding to a risky account, the perfect model would capture all the ‘bads’ as quickly as possible. The Lorenz curve assesses a model’s ability to effectively rank order these accounts. For example,

78

Enterprise Risk Management in Finance

Table 8.6

Good

Bad

Total

KS%

Summary of performance samples January–June 99

July–December 99

January–June 00

N Mean Std. Dev N

26,783 250 24 317

20,849 248 24 307

23,941 246 24

Mean Std. Dev

228 23

231 23

533 225 21

N Mean Std. Dev KS value Score

27,100 249 24 39 240

21,156 248 24 33 246

24,474 245 24 38 239

0.869 1.17 574 459

0.528 1.45 504 569

0.843 2.18 533 1,353

Divergence Bad% * Low score override ** Instore accounts

100%

Percent of bads captured

90% 80% 70% 60% Scorecard Beacon Beacon/Empirica Scorecard (No Bureau score) Applicant Alone Bureau Alone Random Exact

50% 40% 30% 20% 10%

% 30 % 35 % 40 % 45 % 50 % 55 % 60 % 65 % 70 % 75 % 80 % 85 % 90 % 95 % 10 0%

25

%

%

% 20

15

10

5%

0%

0%

Percent into distribution

Figure 8.1 Lorenz curve, January–June 1999 sample

if 15% of the accounts were bad, the ideal or exact model would capture all these bads within the 15th percentile of the score distribution (the x-axis). Similarly, the K–S and divergence statistics determine how well the models distinguished between ‘good’ and ‘bad’ accounts by assessing the properties of their respective distributions. The results indicate that the scorecard is a good predictor of risk. Amongst the three selected sampling periods, the two periods January–June 99 and January–June 00 highlight a slightly better predictive ability than the period July–December 99. Scorecard also performs better than, though not by a significant margin, the credit bureau score.

Percent of bads captured

100% 90% 80% 70% 60%

Scorecard Beacon Beacon/Empirica Scorecard (No Bureau Score) Applicant Alone Bureau Alone Random Exact

50% 40% 30% 20% 10%

%

%

%

%

%

% 50 % 55 % 60 % 65 % 70 % 75 % 80 % 85 % 90 % 95 % 10 0%

45

40

35

30

25

%

20

%

15

5%

10

0%

0%

Percent into distribution

Figure 8.2 Lorenz curve, July–December 1999 sample 100%

Percent of bads captured

90% 80% 70% 60% Scorecard Beacon Beacon/Empirica Scorecard (No Bureau Score) Applicant Alone Bureau Alone Random Exact

50% 40% 30% 20% 10%

%

% 60 % 65 % 70 % 75 % 80 % 85 % 90 % 95 % 10 0%

55

50

% 45

%

% 40

35

% 30

% 25

% 20

% 15

% 10

5%

0%

0%

Percent into distribution

Figure 8.3 Lorenz curve, January–June 2000 sample 100%

Percent of bads captured

90% 80% 70% 60% 50% Jan–Jun’99 Jul–Dec’99 Jan–Jun’00 Random Exact

40% 30% 20% 10%

Percent into distribution

Figure 8.4

Performance comparison of three samples

75 % 80 % 85 % 90 % 95 % 10 0%

70 %

65 %

60 %

55 %

50 %

45 %

40 %

35 %

30 %

25 %

20 %

15 %

5% 10 %

0%

0%

80

Enterprise Risk Management in Finance

The performance statistics for the three selected samples shown in Tables 8.3, 8.4 and 8.5 indicate the superiority of the scorecard as a predictive tool. The scorecard was found to be a more effective assessor of risk for the earlier sample, January–June 99, and the latest sample, January–July 00, but was slightly less effective for the July–December 99 sample. There was a more distinct separation between ‘goods’ and ‘bads’ for the above-mentioned first two samples than the last: the maximum difference between the ‘good’ and ‘bad’ cumulative distributions was 39% and 38% respectively, versus 33% for the remaining sample. Similarly, the divergence values were 0.869 and 0.843, versus 0.528 for the less effective sample. It is possible that the scorecard was better able to separate ‘good’ accounts from ‘bad’ ones for the earlier sample. On the other hand, the process to clean up delinquent unsecured line of credit accounts starting from mid-2001 may result to more bad observations for the latest sample (those accounts booked between January 00 and June 00 with an 18-month observation window will catch up with this clean-up process). This can be evidenced by the bad rate of 2.18% for the January–June 00 sample, compared to 1.45% for the July– December 99 sample, and 1.17% for the January–June 99 sample. If most of these bad accounts in the clean-up have a low initial score, the predictive ability of the scorecard on this cohort will be increased. Population distributions and stability We conduct a comparison analysis between the initial sample used to develop the model and subsequent sampling periods, which provides insight into whether or not the scorecard is being used to score a different population. The analyses considering all applicants are included, but outliers have been excluded, that is, invalid scorecard points. We consider four sampling periods for the cumulative and interval population distribution charts: the FICO development sample, January–June 99, July–December 99, and January–July 00, which had a very notable population shift across the samples where the recent applicants clearly were scoring lower points than before. On the other hand, the development sample was markedly distinct from the three selected samples. We now use the population stability index to estimate the change between the samples. As mentioned in Section 4, a stability index of < 0.10 indicates an insignificant shift; 0.10–0.25 requires some investigation; and > 0.25 means that a major change has taken place between the populations being compared. Tables 8.7, 8.8 and 8.9 present a detailed score distribution report together with the six-month population stability index for each of the above three selected samples, included funded accounts only. Computation shows that the indexes for the three samples on funded accounts are greater than 0.1, and the more recent samples score a lower index than the older samples: 0.2027 for the January–June 99 sample, 0.1461 for the July–December 99 sample, and

527207

Total

100

1.96 1.66 2.59 4.50 6.70 7.86 9.21 10.34 12.02 12.51 12.10 11.40 7.15

(5) −0.0517 −0.0310 −0.0324 −0.0254 −0.0128 −0.0093 −0.0010 0.0113 0.0290 0.0378 0.0441 0.0420 −0.0006

(6) 0.2749 0.3483 0.4440 0.6394 0.8398 0.8944 0.9895 1.1226 1.3187 1.4328 1.5739 1.5835 0.9910

(7)

Ratio (5)/(4)

−1.2912 −1.0546 −0.8119 −0.4471 −0.1746 −0.1116 −0.0106 0.1156 0.2767 0.3596 0.4535 0.4596 −0.0090

(8)

.2027

0.0668 0.0327 0.0263 0.0114 0.0022 0.0010 0.0000 0.0013 0.0080 0.0136 0.0200 0.0193 0.0000

(9)

7.13 11.89 17.72 24.77 32.74 41.53 50.84 60.06 69.17 77.90 85.59 92.78 100.00

(10)

1.96 3.62 6.21 10.71 17.41 25.27 34.48 44.83 56.84 69.35 81.45 92.85 100.00

(11)

Ascending cumulative of Ascending January– cumuContribution Weight June 99 lative of to index of evidence (%) FICO % (8)*(6) in [(7)]

Note: Population Stability Index (sum of contribution): 0.2027 The contribution index can be interpreted as follows: < = 0.10 indicates little to no difference between the FICO development score distribution and the current score distribution. 0.10 to 0.25 indicates some change has taken place. > = 0.25 indicates a shift in the score distribution has occurred.

100

7.13 4.76 5.83 7.04 7.98 8.79 9.31 9.21 9.11 8.73 7.69 7.20 7.22

1430 1209 1888 3284 4885 5735 6716 7543 8762 9121 8826 8310 5216

37601 25093 30742 37128 42055 46355 49068 48577 48034 46023 40541 37940 38050

280

72925

(4)

(3)

(2)

Proportion FICO FICO Change development January– development January– (5)–(4) (%) June 99 (%) (#) June 99 (#)

Score range

(1)

Population stability, January–June 1999

Table 8.7

527207

Total

Note: Population Stability Index (sum of contribution): 0.1461.

100

7.13 4.76 5.83 7.04 7.98 8.79 9.31 9.21 9.11 8.73 7.69 7.20 7.22

1447 1352 2106 3609 5452 6169 7009 7454 7908 7774 7362 6716 3993

37601 25093 30742 37128 42055 46355 49068 48577 48034 46023 40541 37940 38050

280

68351

(4)

100

2.12 1.98 3.08 5.28 7.98 9.03 10.25 10.91 11.57 11.37 10.77 9.83 5.84

(5) −0.0502 −0.0278 −0.0275 −0.0176 0.0000 0.0023 0.0095 0.0169 0.0246 0.0264 0.0308 0.0263 −0.0138

(6)

Proportion July– FICO change development December (5)–(4) 99 (#) (%)

(3)

July– FICO development December 99 (#) (#)

Population stability, July–December 1999

(2)

(1)

Score range

Table 8.8

(8)

0.2968 −1.2146 0.4156 −0.8781 0.5284 −0.6379 0.7498 −0.2880 0.9999 −0.0001 1.0265 0.0261 1.1018 0.0969 0.1685 1.1836 0.2389 1.2699 0.2646 1.3029 0.3370 1.4007 0.3114 1.3654 0.8094 −0.2114

(7)

Ratio (5)/(4)

Weight of evidence in [(7)]

.1461

0.0609 0.0244 0.0175 0.0051 0.0000 0.0001 0.0009 0.0029 0.0059 0.0070 0.0104 0.0082 0.0029

(9)

Contribution to index (8)*(6)

7.13 11.89 17.72 24.77 32.74 41.53 50.84 60.06 69.17 77.90 85.59 92.78 100.00

(10)

2.12 4.10 7.18 12.46 20.43 29.46 39.71 50.62 62.19 73.56 84.33 94.16 100.00

(11)

Ascending cumulative of July– Ascending Cumulative December 99 (#) of FICO %

527207

Total

Note: Population stability index (sum of contribution): 0.1036.

100

7.13 4.76 5.83 7.04 7.98 8.79 9.31 9.21 9.11 8.73 7.69 7.20 7.22

1928 1838 3136 4784 6505 7212 8250 8762 8769 8451 7850 6736 4182

37601 25093 30742 37128 42055 46355 49068 48577 48034 46023 40541 37940 38050

280

78403

(4)

(3)

100

2.46 2.34 4.00 6.10 8.30 9.20 10.52 11.18 11.18 10.78 10.01 8.59 5.33

(5) −0.0467 −0.0242 −0.0183 −0.0094 0.0032 0.0041 0.0122 0.0196 0.0207 0.0205 0.0232 0.0140 −0.0188

(6)

Score range

(2)

Proportion FICO FICO change January– development January– development (5)–(4) (%) June 00 (%) (#) June 00 (#)

(1)

Population stability, January–June 2000

Table 8.9

(8) −1.0648 −0.7082 −0.3770 −0.1434 0.0393 0.0451 0.1227 0.1930 0.2050 0.2109 0.2639 0.1772 −0.3024

(7) 0.3448 0.4925 0.6859 0.8664 1.0401 1.0462 1.1306 1.2129 1.2276 1.2348 1.3020 1.1939 0.7391

Ratio (5)/(4)

.1036

0.0498 0.0171 0.0069 0.0013 0.0001 0.0002 0.0015 0.0038 0.0043 0.0043 0.0061 0.0025 0.0057

(9)

7.13 11.89 17.72 24.77 32.74 41.53 50.84 60.06 69.17 77.90 85.59 92.78 100.00

(10)

2.46 4.80 8.80 14.91 23.20 32.40 42.92 54.10 65.28 76.06 86.07 94.67 100.00

(11)

Ascending Weight of Contribution Ascending cumulative cumulative of January– to index evidence of FICO (%) June 00 (%) (8)*(6) in [(7)]

< = 0.10 0.10 to 0.25 Contribution index < = 0.10

.0787

.0979

.0829

.0898

May00

.0940

April00

January-01

March00

.0962 .1097 .1313 .1236 February-01 March-01 April-01 May-01

.0959

Contribution Februaryindex January-00 00

Table 8.10 Total population stability index

.0615

June-01

.0826

.0696

July-01

.0999

.0940

September00

.0926

October00

.0693

November00

.0656

December00

.0701

.0907

.0816

.0817

.0915

August-01 September-01 October-01 November-01 December-01

.0919

AugustJune-00 July-00 00

.0771

January-02

Bank Credit Scoring

85

0.1036 for the January–June 00 sample. We also computed the monthly population stability, showing the monthly index for total applications (funded or not funded) in the two years starting from January 2000. This result further confirms the declining trend with the monthly indexes for the past 20 months all resting within 0.1 (see Table 8.10). As indicated in Figures 8.5 and 8.6, more of the latest sample accounts had a lower score compared to the older samples; a tendency of scores to drop over time has been revealed. All of the three samples had a score distribution higher than the development sample. The stability indices revealed that the greatest population shift occurred when the scorecard was originally put in place, then the extent of shift reduced gradually across time. The indexes stayed within 0.1 for the past 20 months.

Conclusions

100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%

Score Figure 8.5 Cumulative population distribution on all applicants

280&UP

270–279

260–269

250–259

240–249

230–239

220–229

210–219

200–209

190–199

180–189

170–179

FICO Development Jan–Jun99 Jul–Dec99 Jan–Jun00 0–169

Cumulative distribution

Maintaining a certain level of risk has become a key strategy to make profits in today’s economy. Risk in enterprise can be quantified and managed using various models. Models also provide support to organizations seeking to control enterprise risk. We have discussed risk modeling, and reviewed some common risk measures. Using the variation of these measures, we demonstrate support to risk management through validation of predictive scorecards for a large bank. The scorecard model is validated and compared to credit bureau scores. A comparison of the K–S value and the divergence value between scorecard and bureau score in the three samples indicated scorecard to be a better tool than

86

Enterprise Risk Management in Finance

Interval population distribution

14% 12% 10% 8% 6% FICO Development Jan–Jun99 Jul–Dec99 Jan–Jun00

4% 2%

280&UP

270–279

260–269

250–259

240–249

230–239

220–229

210–219

200–209

190–199

180–189

170–179

0–169

0%

Score Figure 8.6 Interval population distribution on all applications

bureau score to distinguish the ‘bads’ from the ‘goods.’ In practice, however, the vetting and validation of models may encounter many challenges. For example, when retail models under vetting are relatively new to the enterprise, when there are large amounts of variables and data to manipulate and there is limited access to these datasets due to privacy restrictions, when validation tests are not standardized and there are demands for ability to change the measure if results do not look favorable, these challenges become apparent.

Appendix A8.1: Informal Definitions a. Bad accounts refer to cases 90 days delinquent or worse, accounts closed with a ‘NA (non-accrual)’ status or that were written off. b. Good accounts were defined as those that did not meet the ‘bad’ definition. c. Credit score is a number that is based on a statistical analysis of a person’s credit report, and is used to represent the creditworthiness of that person – that is, the likelihood that the person will pay his or her debts. d. A credit bureau is a company that collects information from various sources and provides consumer credit information on individual consumers for a variety of uses. e. Custom score refers the score assigned to existing customers or new applicants. f. Beacon score is the credit score produced at the most recognized agency Equifax in Canada. g. The FICO score is the credit score from Fair Isaac Corporation, a publicly traded corporation that created the best-known and most widely used credit score model in the United States.

9

Credit Scoring using Multiobjective Data Mining

Introduction The technique for order preference by similarity to ideal solution (TOPSIS) is a classical method first developed by Hwang and Yoon,1 subsequently discussed by many.2 TOPSIS is based on the concept that alternatives should be selected that have the shortest distance from the positive ideal solution (PIS) and the longest distance from the negative ideal solution (NIS), or nadir. The PIS has the best measures over all attributes, while the NIS has the worst measures over all attributes. TOPSIS provides a mechanism that is attractive in data mining,3 because it can consider a number of attributes in a systematic way without very much subjective human input. TOPSIS does include weights over the attributes that are considered. However, such weights can be obtained through regression of standardized data (where measurement scale differences are eliminated).4 This allows machine learning, in the sense that data can be analyzed without subjective human input. This chapter demonstrates the method to automatically classify credit score data into groups of high expected repayment and low expected repayment, based upon the concept of TOPSIS.

TOPSIS for data mining The overall approach is to begin with a set of data, which in traditional data mining practice is divided into training and test sets. The data can consist of continuous or binary numeric data, with the outcome variable being binary. The training set data is used to identify maximum and minimum measures for each attribute. This training set is then standardized over the range of 0 to 1, with 0 reflecting the worst measure and 1 the best measure over each attribute. Then relative weight importance is obtained by regression over the standardized data to explain outcome performance in the training data set. 87

88 Enterprise Risk Management in Finance

(An intermediate third data set could be created for generation of weights if desired.) Steps of the TOPSIS data mining method The algorithm we propose thus consists of the following steps: Step 1: Standardize data

In accordance with the presentation above, the training data set is standardized so that each observation j over each attribute i is between 0 and 1. Let the decision matrix X consist of m indicators over n observations. The normalized matrix transforms the X matrix. For indicator i = 1 to m, identify the minimum xi− and the maximum xi+. Then each observation xij for j = 1 to n can be normalized by the following formulas: For measures to be maximized:

yij =

xij − xi− xi+ − xi−

(9.1)

For measures to be minimized: yij = 1 −

(9.2)

xij − xi− xi+ − xi−

which yields values between 0 (the worst) and 1 (the best). Step 2: Determine ideal and nadir solutions

The ideal solution consists of standardized values of 1 over all attributes, while the nadir solution consists of values of 0 over all attributes. Step 3: Calculate weights

In decision analysis, these weights would reflect relative criterion importance (as long as scale differences are eliminated through standardization). Here we are interested in the relative value of each attribute in explaining the outcome of each case. These m weights wi will all be between 0 and 1, and will have a sum of 1. Because weights are continuous, we use ordinary least squares (OLS) regression over the standardized data to obtain the i = 1 to m different weights from regression βi coefficients. 0 ≤ wi ≤ 1,

m

∑w i =1

i

=1

Credit Scoring using Multiobjective Data Mining

89

Step 4: Calculate distances

TOPSIS operates by identifying Di+, the weighted distance from the ideal, and Di−, the weighted distance from the nadir. Different metrics, such as L1, L 2, or L∞, could be used.5 Least absolute value regression (L1) has been found to be useful when it is desired to minimize the impact of outlying observations, and has been shown to be effective in a variety of applications, such as real estate valuation6 and sports ratings.7 The ordinary least squares metric is L 2, widely used. The Tchebychev metric (L∞) focuses on the extreme performance among the set of explanatory variables. Each metric focuses on the different features described. Olson8 found L1 and L 2 to provide similar results, both better than L∞. The weights from Step 3 are used. Lee and Olson9 compared different metrics for predicting outcomes of binary games, and found L 2 and L∞ to provide similar results, both better than L1. Thus, none of these metrics is clearly superior to the others for any specific set of data. For the L1 metric, the formula for the weighted distance from the ideal is: D1j + =

m

∑ w × (1 − y ) i

ij

for j = 1 to n

(9.3)

i =1

The weighted distance from the nadir solution is: D1j − =

m

∑ w × (y ) i

ij

for j = 1 to n

(9.4)

i =1

The formulas for the L2 metric are very similar: Dj2 + =

m

∑ w × (1 − y ) 2 i

2 ij

for j = 1 to n

(9.5)

i =1

The weighted distance from the nadir solution is: Dj2 − =

m

∑ w × (y ) 2 i

2 ij

for j = 1 to n

(9.6)

i =1

The L∞ metric (the Tchebychev metric) by formula involves the infinite root of an infinite power, but this converges to emphasizing the maximum distances. The weights become irrelevant. Thus the L∞ distance measures are: Dj∞ − = MAX{ yij }; Dj∞ + = MAX{1 − yij }

(9.7)

Step 5: Calculate closeness coefficient

Relative closeness considers the distances from the ideal (to be minimized) and from the nadir (to be maximized) simultaneously through the TOPSIS formula: Cj =

DjL − D

L− j

+ DnL +

(9.8)

90 Enterprise Risk Management in Finance

Step 6: Determine cutoff limit for classification

The training data set contained a subset of observations in each category of interest. In a binary application (such as segregating training observations into loans that were defaulted, Neg, and loans that were repaid, Pos), the proportion of Neg observations, PNeg, is identified. The closeness coefficient, Cj, has high values for cases that are close to the ideal and far from the nadir, and thus can be sorted with low values representing the worst cases. Thus the rank of the largest sorted observation in the Neg subset JNeg would be PNeg × (Neg + Pos). The cutoff limit, CLim, can be identified as a value greater than that of ranked observation, JNeg, but less than that of the next largest ranked observation. Step 7: Apply formula

For new cases with unknown outcomes, the relative closeness coefficient Cj can be calculated by Formula (9.7) and compared with the cutoff limit obtained in Step 6. The only data feature that needs to be considered is that it is possible for the test data to contain observations outside the range of data used to determine training parameters. IF yij < 0 THEN yij = 0 IF yij > 1 THEN yij = 1 This retains the standardized features of the test data. The model application is then obtained by applying the rules to the test data: IF Cj < CLim THEN classification is Negative IF Cj > CLim THEN classification is Positive Model fit is tested by traditional data mining coincidence matrices.

Dataset A company’s financial performance can be represented by various ratios taken from financial statements.10 Such ratios provide useful information to describe credit conditions from various perspectives, such as financial conditions and credit status. The diagnostic process involves multiple criteria. We present a real set of loan cases from Canadian banking. The data reflects operations in 1995 and 1996; there are 177 observations for 1995 (17 defaulting, 160 good), and 126 (11 defaulting, 115 good) for 1996. While the dataset is unbalanced (banks would hope that it was), it is typical. Models for decision trees can be susceptible to degeneration, as they often classify all observations in the ‘good’ category.11 This did not prove to be a problem with this dataset, but Laurikkala

Credit Scoring using Multiobjective Data Mining

Table 9.1

91

Independent variables for Canadian banking data set

Variable Total Assets Capital Assets Interest Expense Stability of Earnings Working Capital Total Current Liabilities Total Liabilities Retained Earnings Shareholders Equity Net Income Earnings before Tax & Depr. Cash Flow from Operations

TA CA IE INSTAB WC CL TL RE SE NI EBITDA CF

Training set minimum value

Training set maximum value

332 107 0 34.781 −403,664 33 33 −486,027 −430,935 −238,326 −132,388 −41,387

421,029 269,188 70,938 74,672.86 169,523 578,857 584,698 225,719 298,903 97,736 158,401 95,427

Goal Maximize Maximize Minimize Maximize Maximize Minimize Minimize Maximize Maximize Maximize Maximize Maximize

(2002)12 and Bull (2005)13 provide procedures to deal with such problems of unbalanced data if they detrimentally affect data mining models. The dataset consisted of the outcome variable (categorical: default, good) and 12 continuous numeric independent variables, as given in Table 9.1. This dataset demonstrates many features encountered with real data. Most variables are to be maximized, but here, three of the twelve variables would have the minimum as preferable. There are negative values associated with the dataset as well.

TOPSIS model over training data Step 1: Standardize data

The data set was standardized using Formulas (9.1) and (9.2). Step 2: Determine ideal and nadir solutions

The ideal solution here is a vector of standardized scores of 1: {1 1 1 1 1 1 1 1 1 1 1 1} reflecting the best performance identified in the training set for each variable. The nadir solution is conversely: {0 0 0 0 0 0 0 0 0 0 0 0}. All n observations would have a standardized score vector consisting of m (here m = 12) values between 0 and 1.

92

Enterprise Risk Management in Finance

Table 9.2 Standardized data regression

Variable

Regression coefficient βi

P-Value

Absolute value of βi

Proportional weight

TA CA IE INSTAB WC CL TL RE SE NI EBITDA CF

−1.4205 0.5263 −1.7013 −0.3245 −0.1028 0.3010 −0.6058 0.3551 2.1597 3.3446 −2.2372 0.7510

0.926 1.000 0.043 0.453 1.000 1.000 0.977 0.477 0.935 0.051 0.084 0.056

1.4205 0.5263 1.7013 0.3245 0.1028 0.3010 0.6058 0.3551 2.1597 3.3446 2.2372 0.7510

0.103 0.038 0.123 0.023 0.007 0.022 0.044 0.026 0.156 0.242 0.162 0.054

13.8298

1.000

Totals

Step 3: Calculate weights

Weights were obtained by regressing over the standardized data with the outcome of 0 for default and 1 for no default. Table 9.2 shows the results of that regression (using ordinary least squares). This model had an R-Square value of 0.287 (adjusted R-Square of 0.235) – relatively weak. Correlation analysis indicated a great deal of multicollinearity (demonstrated by the many insignificant beta coefficients in Table 9.2), so a trimmed model using the three uncorrelated variables of NI, EBITDA, and CF was run. This trimmed model had an R-Square of 0.245 (adjusted R-Square of 0.232), but the predictive capability of this model was much weaker than that of the full model. Multicollinearity would be a problem with respect to variable β coefficient significance, but since our purpose is prediction of the overall model rather than interpretation of the contribution of each independent variable, this is not a problem in this application. Therefore, the full regression model was used. Weights obtained in Step 3 are therefore given in the last column of Table 9.2. However, these weights should not be interpreted as accurate reflections of variable prediction importance due to the model’s multicollinearity, which makes these weights unstable given overlapping information content. Step 4: Calculate distances

Three metrics were used for the TOPSIS models in this study. For the L1 model, the values for Di1+ were obtained by generating by Formula (9.3) for each observation over each variable in the training set, and Di1− obtained by Formula (9.4). Formulas (9.5) and (9.6) were used for the L 2 model, and the formulas given in (9.7) for the L∞ model.

Credit Scoring using Multiobjective Data Mining

93

Step 5: Calculate closeness coefficient

Formula (9.8) was applied to the distances obtained in step 4 for the training set. Step 6: Determine cutoff limit for classification

The 177 closeness coefficient values were then sorted, obtaining a 17th ranked closeness coefficient and an 18th ranked closeness coefficient. For the L1 model, these were 0.56197 and 0.561651. Thus an L1 cutoff limit of 0.5615 was obtained for application on the test set and for classification of future values. For the L 2 model, the corresponding numbers were 0.410995 and 0.412294, yielding a cutoff limit of 0.411. For the L∞ model, these numbers were 0.624159 and 0.624179, and a cutoff limit of 0.62416 was used. Step 7: Application of model

The last step is to apply models to test data. Results are given.

Model comparisons The original raw data was used with two commercial data mining software tools (PolyAnalyst and See5) for decision tree models. The PolyAnalyst decision tree model used only two variables, NI and WC. The decision tree is given in Figure 9.1. This model had a 0.865 correct classification rate over the test set of 126 observations, as shown in Table 9.3. IF NI < 1250 AND IF WC < 607 THEN 0 (Neg) ELSE IF WC >= 607 THEN 1 (Pos) ELSE IF NI >= 1250 THEN 1 (Pos)

Yes IF NI 1256

No

IF NI ≤ –26958

Yes

IF CF > 24812 Yes

Yes

No No

THEN Positive outcome predicted

IF CA ≤ 828 Yes No

Yes THEN Negative outcome predicted

Figure 9.2

IF IE > 2326

THEN Negative outcome predicted

THEN Positive outcome predicted

THEN Negative outcome predicted

No THEN Positive outcome predicted

See5 decision tree

The errors in this model were a bit more proportional in the bad case of assigning actual default cases to the predicted on-time payment category. However, cost vectors were not used, so there was no reason to expect the model to reflect this.

Credit Scoring using Multiobjective Data Mining

95

See5 software yielded the following decision tree, using four independent variables. Table 9.4 shows the results for the decision tree model obtained from See5 software, which had a correct classification rate of 0.817, just a little worse than the PolyAnalyst model (although this is for the specific data, and in no way is generalizable): Table 9.4

Coincidence matrix – See5 decision tree

Actual 0 (Neg) Actual 1 (Pos)

Model 0 (Neg)

Model 1 (Pos)

8 13 21

3 102 105

11 115 126

Finally, the TOPSIS models were run. Results for the L1 model are given in Table 9.5, with a correct classification rate of 0.944. Table 9.5 Coincidence matrix – TOPSIS L1 model

Actual 0 (Neg) Actual 1 (Pos)

Model 0 (Neg)

Model 1 (Pos)

6 2 8

5 113 118

11 115 126

The results for the L 2 model are given in Table 9.6, with a correct classification rate of 0.921. Table 9.6

Coincidence matrix – TOPSIS L2 model

Actual 0 (Neg) Actual 1 (Pos)

Model 0 (Neg)

Model 1 (Pos)

6 5 11

5 110 115

11 115 126

The results for the L∞ model are given in Table 9.7, with a correct classification rate of 0.844. Here all three metrics yield similar results (for this data, superior to the decision tree models; but that is not a generalizable conclusion). Table 9.7 Coincidence matrix – TOPSIS L∞ model

Actual 0 (Neg) Actual 1 (Pos)

Model 0 (Neg)

Model 1 (Pos)

6 2 8

5 113 118

11 115 126

96

Enterprise Risk Management in Finance

The results for the different models are given in Table 9.8. Table 9.8 Comparison of model results Model

Actual 0 Model 1

Actual 1 Model 0

Proportion correct

2 3 5 5 5

6 13 2 5 2

0.937 0.873 0.944 0.921 0.944

Polyanalyst decision tree See5 decision tree TOPSIS L1 TOPSIS L 2 TOPSIS L∞

These models were applied to one data set, demonstrating how TOPSIS principles can be applied to data mining classification. In this one small (but real) data set for a common data mining application, the TOPSIS models gave better fit to test data than did two well-respected decision tree software models. This does not imply that the TOPSIS models are better, but it provides another tool for classification. The TOPSIS models are easy to apply in spreadsheets, however much data has been fit into that spreadsheet. Any number of independent variables could be used, limited only by database constraints.

Simulation of model results The Monte Carlo simulation provides a good tool to test the effect of input uncertainty over output results.14 Simulation was applied to examine the sensitivity of the five models to perturbations in test data. Each test data variable value was adjusted by adding an adjustment equal to: Perturbation × Uniform random number × Standard normal variate The perturbations used were 0.25, 0.5, 1, and 2. These values reflect increasing noise in the data. The adjustments are standard normal variates with mean 0 and standard deviation found in the training dataset for that variable. Simulation results are shown in Table 9.9. Table 9.9 Simulation results

Perturbation

PADT Min

PADT Max

C5 Min

0 0.25 0.50 1.0 2.0

0.9365 0.7381 0.7063 0.6905 0.6587

0.9365 0.9365 0.9444 0.8968 0.8968

0.8730 0.6746 0.5952 0.5317 0.5238

C5 Max

L1 Min

0.8730 0.9444 0.8651 0.7619 0.8413 0.7143 0.7937 0.6349 0.7857 0.5714

L1 Max 0.9444 0.9286 0.8968 0.8492 0.8175

L2 Min

L2 Max

L∞ Min

L∞ Max

0.9206 0.9206 0.9444 0.9444 0.7619 0.9127 0.6825 0.8889 0.6190 0.8730 0.6349 0.8492 0.5873 0.8333 0.4683 0.7460 0.5476 0.7937 0.3810 0.6508

Credit Scoring using Multiobjective Data Mining

97

The first decision tree model was quite robust, and in fact retained its predictive power in most of all five models as perturbations were increased. The second decision tree model included more variables and a more complex decision tree. However, it not only was less accurate without perturbation, it also degenerated much faster than the simple two-variable decision tree. While this is not claimed as generalizable, it is possible that simpler trees could be more robust. (As a counterargument, models using more variables may have less reliance on specific variables subjected to noise, so this issue merits further exploration.) The L1 and L 2 TOPSIS metrics had less degeneration than the four-variable decision tree, but a little more than the two-variable decision tree. The L1 TOPSIS model was less affected by perturbations than was the L 2 model, which in turn was quite a bit less affected than the L∞ model. This is to be expected, as the L1 model is less affected by outliers, which can be generated by noise. The L∞ model focuses on the worst case, which is a reason for it to be adversely impacted by noise in the data.

Conclusions TOPSIS is attractive in that it follows automatic machine learning principles. TOPSIS was originally presented in the context of multiple-criteria decisionmaking, where the relative importance of the decision-maker’s preference was a factor, and subjective weights were input. In the data mining application presented here, the weights are obtained from the data itself, removing the subjective element. Weights here reflect how much each independent variable contributes to the best ordinary least squares fit to the data. Standardizing the data removes differences in scale across independent variables. Thus the TOPSIS models provide a straightforward way to classify data with any number of independent variables and observations. The classical methods for classification, decision trees, are valuable tools. Decision trees have a useful feature in that they provide easy-to-interpret sets of rules, as shown in Step 5. In the spirit of data mining, the TOPSIS models presented can provide an additional tool for comparative analysis of data. Three metrics were presented here. The L2 metric is traditionally used, although the L1 and L∞ metrics are just as valid. The L1 metric is usually considered less susceptible to the influence of outlier data, as squaring the distance from the measure of central tendency in the L2 metric has a greater impact. In the Tchebychev L∞ metric, the greatest difference determines the outcome, which is attractive in some contexts. If outliers are not intended to have greater influence, the L1 metric might therefore be preferred. If all variables are to be considered to the greatest degree, the L∞ metric is attractive. Here, however, we confirm prior results cited, and find that the L2 metric seems to perform quite well.

98 Enterprise Risk Management in Finance

Simulation was used to demonstrate relative model performance under different levels of noise. While simulation of data mining models involves quite a bit of extra computation, it can provide insight as to how robust models are expected to be. Nevertheless, future research with TOPSIS data mining is suggested. The possible direction includes developing new techniques to derive weights, instead of using the linear regression approach in this chapter.

10

Online Banking Efficiency and Risk Evaluation with Principal Component Analysis

Introduction Online banking has always been important from different stakeholder perspectives.1 Improving the efficiency of Internet banking is now considered to be important. Banks hope that internet banking will help them maintain profitable growth by enabling them to automate work, reduce costs, and retain customers simultaneously. 2 Internet banking may help reduce expenditure on ‘bricks and mortar,’ and reduce capital expenditures. 3 Internet banking can give customers 24-hour access, and provide convenience for customers. Cost-effective use of the internet can attract many users to online banking services, but there has been little research examining the superiority of banks providing online banking services over those that do not. The role of online banking in leading to better decisions and creating more profit needs study. The Economist has reported that online banking costs were approximately $0.01 per transaction in 2000, accounting for only 1% of all banking transactions.4 Online banking costs are increasing rapidly. Jupiter5 reported that there were an estimated 30 million families in the US using internet banking in 2004, and forecast that the numbers would reach 56 million in 2008. The growth of internet banking is due to economic globalization and the maturation of computer technology. Economic globalization encourages banks to serve and exploit international markets. Opening up a new virtual bank branch requires high expenditure and faces limitations, but online banking websites can provide services to customers throughout the world as long as customers are able to surf the internet. The reduction of computer prices makes their purchase accessible to most people. Meanwhile, perceived improvement in computer security makes customers and banks sufficiently comfortable to use them for their business. The consequent rapid increase of online banking transactions creates an urgent need to examine the efficiency of online banking. 99

100 Enterprise Risk Management in Finance

Furthermore, efficiency evaluation is a source of ideas for bank managers, and motivates them to improve the quality of online banking services. Serrano-Cinca et al. (2005)6 suggested that financial information alone might not always be sufficient to judge an online business. Therefore, constructing performance evaluation methods that can combine a number of possible inputs and outputs is attractive. Banks can evaluate their online banking performance using online performance measurement considering an efficient frontier of tradeoffs across selected criteria. Previous literature on online banking performance evaluation includes the following approaches: linear regression (e.g. the logit model in Furst et al.), DEA (e.g. Sherman and Gold, 19857; Soteriou and Zenios, 19998), free disposal hull (e.g., Tulkens, 19939), the stochastic frontier approach (also called econometric frontier approach, e.g. Berger and Humphrey, 199210), the thick frontier approach (e.g. Berger and Humphrey, 199711), the distribution free approach (e.g., Berger et al., 199312), and others. The main differences between these approaches lie in how much restriction is imposed on the specification of the best practice frontier and the assumptions of random errors and inefficiency.13 DEA is recognized as useful for performance analysis in banking industry because of its advantages in allowing for dynamic efficiency without requirements for prior assumptions on the specification of the best practice frontier. This translates into practical advantages for DEA over other methods, in that an explicit specification of a mathematical form for the production function is not needed, and DEA can provide an integrated efficiency score to decide the level of efficiency of a specific bank. Conversely, there can be computational difficulty if there are too many input and output variables. Principal component analysis (PCA) is a flexible approach that can be used to reduce the number of variables and to classify independent variables. PCA can thus provide powerful support to DEA by providing a means to reduce the number of input variables and integrating output variables. PCA has been applied in many scientific disciplines, including statistics, economics, finance, biology, physics, chemistry, and so on. In internet banking, Eriksson et al. (2008)14 used PCA to identify determinants of customer satisfaction. PCA has also been used to study the acceptance of internet banking by customers, with a primary conclusion suggesting that banks place effort on making internet banking user-friendly.

Data and variables To analyze efficiency, we looked into a few banks chosen from the UK and the US. Data was gathered from 2007 annual reports (see the reference list for such reports) of these banks divided into the input variables and the output variables. Table 10.1 gives input and output variables: Table 10.2 presents online banking data for ten large banks, six from the US and four from the UK: Bank of America,15 Citigroup,16 HSBC,17 Barclays,18

Performance Evaluation and Risk Analysis of Online Banking 101

Table 10.1 Online banking DEA variables INPUT variables

Designator

OUTPUT variables

Total deposits Operating cost Employees Equipment

A B C D

Total revenue Daily visits

Designator 1 2

Table 10.2 Online banking data

Bank Bank of America Citigroup HSBC Barclays Chase Wells Fargo Lloyds Royal Bank of Scotland SunTrust Wachovia

Total Operating deposits cost Employees Equipment Revenue

Daily reach

805177 826230 278693 657058 740728 344460 393092 595908

89881 61488 39042 26398 110560 22824 11134 28106

210000 387000 330000 135000 180667 159800 70000 226400

9404 8191 6054 5992 3779 1294 4014 3524

68068 81698 87601 40410 71400 39390 58180 108934

4524000 623000 136000 1083000 2030000 1051000 623000 134000

119877 449120

5234 9465

32323 29940

1739 6605

8251 17653

129000 551000

Chase,19 Wells Fargo,20 Lloyds,21 Royal Bank of Scotland,22 SunTrust,23 and Wachovia.24

Results and analysis We first conducted a plain DEA analysis and then used PCA scenario analysis and PCA-DEA modeling. Table 10.3 presents the scores produced from normal DEA models based on all 45 combinations of variables in the data. A score of 1.000 (in bold) indicates efficiency. It can be seen in the ABDCD12 column that seven of the ten banks are found efficient. The problem is that using more variables obviously finds more efficient solutions. The ABCD12 model generates too many efficient DMUs/ties due to many input and output variables. But many variables and factors can affect each other in real world; for example, operating cost is usually related to the number of employees in the corporation. Therefore PCA is applied to reduce these measures when applying DEA. To conduct PCA analysis, we use various combinations of variables to see if we can reduce the number of variables and/or detect structure in the relationships between variables. Table 10.4 presents maximum component loadings matrix in different models with different data. All models are weighted with a positive sign

C12

1.000 0.254 0.319 0.476 0.644 0.391 1.000 0.579 0.337 1.000

B12

0.865 0.254 0.429 0.707 0.316 0.794 1.000 0.742 0.432 1.000

A12

1.000 0.395 1.000 0.403 0.652 0.747 0.651 0.582 0.348 0.280

0.592 0.324 0.468 0.223 0.661 1.000 0.472 1.000 0.155 0.103

D12

0.592 0.094 0.028 0.223 0.661 1.000 0.191 0.047 0.091 0.103

D2

0.234 0.323 0.468 0.218 0.611 0.985 0.469 1.000 0.154 0.087

1.000 0.492 1.000 0.764 0.652 0.924 1.000 1.000 0.488 1.000

AB12

1.000 0.198 0.087 0.764 0.488 0.892 1.000 0.090 0.462 1.000

AB2

0.309 0.447 1.000 0.362 0.307 0.545 1.000 1.000 0.390 0.357

AB1

1.000 0.514 1.000 0.522 0.789 0.783 1.000 1.000 0.466 1.000

AC12

1.000 0.134 0.087 0.372 0.522 0.543 0.413 0.040 0.192 0.854

AC2

0.521 0.501 1.000 0.403 0.605 0.582 1.000 1.000 0.420 0.709

AC1

1.000 0.507 1.000 0.462 0.834 1.000 0.786 1.000 0.365 0.280

AD12

1.000 0.149 0.087 0.344 0.786 1.000 0.308 0.060 0.192 0.218

AD2

0.366 0.458 1.000 0.294 0.611 0.985 0.677 1.000 0.271 0.154

AD1

1.000 0.254 0.429 0.707 0.644 0.794 1.000 0.742 0.432 1.000

BC12

1.000 0.174 0.060 0.705 0.522 0.791 0.961 0.082 0.423 1.000

BC2

0.390 0.254 0.429 0.360 0.476 0.330 1.000 0.742 0.307 0.709

BC1

1.000 0.399 0.555 0.774 0.661 1.000 1.000 1.000 0.445 1.000

BD12

1.000 0.199 0.068 0.774 0.661 1.000 1.000 0.094 0.445 1.000

BD2

0.234 0.339 0.555 0.343 0.611 0.985 1.000 1.000 0.311 0.357

BD1

1.000 0.423 0.526 0.558 1.000 1.000 1.000 1.000 0.353 1.000

CD12

1.000 0.135 0.038 0.375 0.952 1.000 0.413 0.062 0.185 0.854

CD2

0.447 0.400 0.526 0.414 0.751 0.985 1.000 1.000 0.319 0.709

CD1

1.000 0.199 0.087 0.774 0.786 1.000 1.000 0.094 0.462 1.000

ABD2

0.366 0.458 1.000 0.362 0.611 0.985 1.000 1.000 0.390 0.357

ABD1

1.000 0.149 0.087 0.375 0.952 1.000 0.413 0.062 0.192 0.854

ACD2

0.521 0.501 1.000 0.414 0.751 0.985 1.000 1.000 0.420 0.709

ACD1

1.000 0.199 0.068 0.774 0.952 1.000 1.000 0.094 0.445 1.000

BCD2

0.447 0.400 0.555 0.414 0.751 0.985 1.000 1.000 0.319 0.709

BCD1

1.000 0.199 0.087 0.774 0.952 1.000 1.000 0.094 0.462 1.000

ABCD2

0.521 0.501 1.000 0.414 0.751 0.985 1.000 1.000 0.420 0.709

ABCD1

1.000 0.514 1.000 0.764 0.789 0.924 1.000 1.000 0.488 1.000

1.000 0.525 1.000 0.774 0.834 1.000 1.000 1.000 0.488 1.000

1.000 0.542 1.000 0.558 1.000 1.000 1.000 1.000 0.466 1.000

1.000 0.434 0.555 0.774 1.000 1.000 1.000 1.000 0.445 1.000

1.000 0.542 1.000 0.774 1.000 1.000 1.000 1.000 0.488 1.000

ABC12 ABD12 ACD12 BCD12 ABCD12

1.000 0.198 0.087 0.764 0.522 0.892 1.000 0.090 0.462 1.000

ABC2

0.521 0.501 1.000 0.403 0.605 0.582 1.000 1.000 0.420 0.709

ABC1

Notes: A: Deposit; B: Operation cost; C: The number of employees; D: Equipment; 1: Revenue; 2: Web reaches.

1.000 0.075 0.019 0.372 0.522 0.305 0.413 0.028 0.185 0.854

0.865 0.174 0.060 0.705 0.315 0.791 0.961 0.082 0.423 1.000

1.000 0.134 0.087 0.293 0.488 0.543 0.282 0.040 0.192 0.218

0.390 0.254 0.319 0.360 0.476 0.297 1.000 0.579 0.307 0.709

C2

0.145 0.254 0.429 0.293 0.124 0.330 1.000 0.742 0.302 0.357

B2

A2

0.269 0.315 1.000 0.196 0.307 0.364 0.471 0.582 0.219 0.125

D1

B1

A1

C1

DEA combinations and their efficiencies

Table 10.3

Performance Evaluation and Risk Analysis of Online Banking 103

on the first component and other variables, thus, the first component is named ‘overall measure of efficiency,’ and is the higher-weighted value in general. In all models, variables can be reduced so that three principal components can be used to explain more than 85% of the variation in the raw data. In this research we chose the principal components with the accumulative contribution ratio equal to or above 90%, and applied them to reevaluate the efficiency of the online banking. For instance, in model ABCD12 we can conclude that only three principal components are used, because the accumulative contribution of three principal components is 90.2%. The calculation is easy due to the reduction of data complexity, and the ranks are reasonable. The data extraction process is shown in Figure 10.1, where every vector represents each of the six variables. How each variable contributes to the three principal components is indicated by the direction and length of the vector. Table 10.4 Maximum component loadings matrix in different models Model ABCD1 ABCD2 ABC12 ABD12 ACD12 BCD12 ABCD12

PC1

PC2

PC3

0.851 0.895 0.892 0.906 0.868 0.849 0.881

−0.561 0.750 0.719 0.804 0.705 0.739 −0.675

0.504 −0.598 −0.467 0.640 −0.526 0.662 0.674

Component 3

Notes: A: Deposit; B: Operation cost; C: Number of employees; D: Equipment; 1: Revenue; 2: Web reaches.

1 0 Employees Total revenue Equipment

–1 –1

Total deposits

–0.5

Operating cost Daily visits

0 0.5 Component 2

1 1

–1

–0.5

0 Component 1

Figure 10.1

3D plot of PCA analysis (ABCD12 model)

0.5

104

Enterprise Risk Management in Finance

In the three-dimensional solid area it is shown that the first principal component, represented by the Component 1 axis, has positive coefficients for all six variables. The second principal component, represented by the Component 2 axis, has negative coefficients for variables Total deposits (A), Operating cost (B), Equipment (D) and Daily visits (2), and positive coefficients for the remaining two variables. The Component 3 axis has all positive and negative coefficients for the variables. This figure indicates that these components distinguish between online banking that has high values for the three sets of variables, and low for the rest. This approach can effectively achieve dimensionality reduction without losing too much information. PCA-DEA analysis This section analyzes online banking using an integrated PCA-DEA model. As internet usage grew, it considerably changed the channel between banks and clients. We used both financial and non-financial variables. The main objective is to construct a framework with DEA and PCA approaches, and use it for the measurement of online banking based on data collected from annual reports and web metrics. The basic DEA model in Table 10.2 shows us that we have seven banks that are efficient. To reduce the number of variables and obtain more accurate results, we ran a PCA-DEA analysis at 75%. This left us with three efficient banks: Bank of America, Lloyds and the Royal Bank of Scotland. Table 10.5 presents integrated PCA-DEA scores, and Table 10.6 gives variance explained by integrated PCA-DEA. Only the Bank of America gets 100% efficiency in all models. Note that in Model ABCD1, where all inputs and outputs except for Daily Reach are taken Table 10.5

Integrated PCA-DEA score

PCA-DEA – 75%

ABD12

ABC12

ABCD1

ACD12

ABCD12

Bank of America Citigroup HSBC Barclays Chase Wells Fargo Lloyds Royal Bank of Scotland SunTrust Wachovia Min Max Mean Standard Deviation

100% 48% 86% 47% 65% 100% 87% 100% 36% 29% 29% 100% 70% 0.2808

100% 43% 80% 56% 60% 75% 100% 94% 45% 50% 43% 100% 70% 0.2239

100% 14% 5% 39% 54% 67% 38% 5% 20% 28% 5% 100% 37% 0.3005

100% 50% 76% 53% 99% 100% 97% 100% 39% 32% 32% 100% 75% 0.2805

100% 48% 75% 57% 77% 98% 100% 100% 41% 38% 38% 100% 73% 0.2579

Performance Evaluation and Risk Analysis of Online Banking 105

into account, only one bank gets a score of 100% and the average efficiency is only 37%. In the other models where the daily reach is taken into account the average ranges from 70 to 75%. To analyze all efficiency scores in Table 10.2, we ran PCA on all 45 sets of scores for all banks. Figure 10.2 gives a plot of principal component loadings in different DEA models, where different scores can be understood from a different perspective. The plots of principal component loadings are converted from the matrix of component loadings and show a set of directional vectors. The computed component loadings result in meaningful naming for both horizontal (the first component) and vertical axis (the second component). The horizontal axis in Figure 10.2 is from west to east, representing the ‘overall Table 10.6 Variance explained in integrated PCA-DEA Variance A B C D 1 2

ABD12

ABC12

ABCD1

ACD12

ABCD12

65.53 15.05 14.39 5.04 53.85 46.15

73.66 19.39 0.00 6.95 53.85 46.15

71.45 19.65 8.89 0.00 53.85 46.15

65.53 15.05 14.39 5.04 100.00 0.00

68.21 20.02 11.77 53.85 46.15

1.0 AB1

0.9 0.8

ABD1

0.7

ACD1&ABCD1 D1

0.6

Component 2

0.5

ACD1&ABC1

AD1

A1

BD12 B1

Cost oriented (Vertical)

BCD1

AD12 D12

0.4 BC1

0.3

A12

0.2 C1

0.1 0.0

CD1 AC12 ACD12 ABD12 AB12 ABC12 ABCD12 iency ll Effic Overa BD12

(Horizontal) B12

–0.1 –0.2 –0.3

BC12

C12

D2

CD12 BCD12

–0.4 –0.5

AD2

Online oriented (Vertical) A2

–0.6 –0.7 –0.1

0.0

0.1

0.2

0.3

0.4

0.5

CD2&ACD2 (AB2&BC2&ABC2&ABD2 &BCD2&ABCD2) BD2

B2 C2 AC2

0.6

0.7

0.8

0.9

1.0

Component 1

Figure 10.2

Plots of principal component loadings in different DEA models

106

Enterprise Risk Management in Finance

Table 10.7 Multivariate linear regression analysis – DV bank revenue

Constant A TotalDeposits B Cost C Employees D Equipment 2 Daily Reach

Standardized coefficients

T

Sig.

Correlation coefficients

NA 0.319 −0.011 0.749 −0.246 −0.030

0.710 0.632 −0.017 1.698 −0.553 −0.056

0.517 0.561 0.987 0.165 0.610 0.958

0.487 0.473 0.783 0.275 0.077

measure of efficiency’; the more efficient ones overall will be located in the right direction. From the origin to north and south are the ‘cost oriented’ and ‘online oriented’ models respectively. Interestingly enough, such a finding is consistent with existing work based on data from other nations. Risk factors This section seeks to detect the key variables that contribute the most to bank revenue. We ran both multivariate linear regression and correlation analysis and present computed values in Table 10.7. It can be seen that the Employees variable has the largest effect on revenue with a regression coefficient value of 0.749 and correlation coefficient value of 0.783. This means that allocation of Employees will affect profit or loss the most, holding the other variables constant. That is to say, misutilization of Employees will bring huge potential risks, which should be considered by banks’ managers when they decide to enter the online banking market. This is because the risks have become important factors affecting the survival of enterprises. Meanwhile, the Basel Committee on Banking Supervision requires that every bank must have an efficient regulation and risk management system where enterprise risk managers should take an important role on the board of directors.25 Three other variables, that is, total deposits, cost, and equipment, have less effect on revenue, and daily reach has no effect on revenue at all. It may be easy to understand that the people who click the online banks’ websites do not necessarily conduct transactions over the websites. Therefore, future research should examine whether the number of transactions will significantly affect revenue.

Conclusions Various financial and non-financial variables have been used in this study to analyze the online banking service of some giant banks. PCA is employed to identify the variables contributing the most information content for DEA

Performance Evaluation and Risk Analysis of Online Banking 107

models. The results enable identification of the most efficient banks in terms of the particular variables selected through DEA. This information is then further analyzed in terms of multivariate linear regression, which enables significance and correlation to be seen across variables. The combination of models applied to banking risk management issues (in this case, online banking) can provide useful tools to benchmark banking operations and to identify opportunities for improvement in those operations.

11

Economic Perspective

The traditional economic view An early economic view of risk was addressed by Frank H. Knight,1 who focused on the difference between uncertainty (a domain evading accurate measurement: ‘cases of the non-quantitative type’) and risk: (‘a measurable uncertainty, or “risk” proper’). Risk applied to cases where knowledge is available about future outcomes and their probabilities, while uncertainty applies to cases where there is knowledge about future outcomes but not about their probabilities. Risk was viewed as important, drawing upon Courcelle-Seneuil’s view2 that profit is due to the assumption of risk, compatible with von Thünen’s view3 that profit was, in part, payment for certain risks. This was also expressed by F. B. Hawley,4 who argued that risk-taking was the essential function of the entrepreneur. Knight viewed risk as the objective form of the idea, and reserved subjective elements to uncertainty. Since risk was measurable, those things to which it applied had statistical distributions available. Ganegoda and Evans (2012)5 proposed a framework for uncertainty assessment in financial markets. This framework is displayed in Table 11.1. Ignorance covers cases where there is no reliable evidence. Ambiguity covers cases where subjectivity reigns, allowing different humans to interpret the same term differently;such cases arise frequently in banking and investment. Ganegoda and Evans use Li’s Gaussian copula model6 to price collateralized debt obligations as a case in point. Although Li’s model (which sought to identify the risk of a CDO by a correlation metric) seemed to work quite well before the 2008 crisis, it failed under the conditions of stress that that crisis entailed. Much financial forecasting is actually ambiguous in nature, because while we gather statistics on past performance, we bet on future outcomes, and if the underlying conditions of the future vary a great deal from those of the past (and things are always somewhat different over time), statistical assumptions need to be considered in light of changed conditions. The difference 108

Economic Perspective

109

Table 11.1 Realms of uncertainty Label

Realm

Example

Ignorance

Future events are unknown

Value of real estate in a war zone

Ambiguity

Future events vaguely defined

Credit rating of asset-backed securities

Uncertainty

Events known but probability distribution is not

Risk of natural disaster High-degree unique events (fines)

Risk

Events and associated probability distribution known

Most insurance, market risk for banks

Source: Based on Ganegoda and Evans (2012).

between uncertainty and risk is somewhat clearer, although the boundary between uncertainty and ambiguity is fuzzy. The definition of ambiguity would emphasize an infinity of possible outcomes, while uncertainty could be reserved for the realm of definable outcomes. The conventional economic and financial theory of risk is represented by Harry Markowitz’s7 definition of risk as variance. By analyzing the statistical properties of investment alternatives, wise investors can minimize their risk by diversifying, especially across investment alternatives with low or negative correlations. Risk introduces a second criterion of investment along with profit. This leads to the concept of an efficient frontier, the set of investment alternatives which have the lowest variance for a given return, or conversely the highest return for a given variance. Related ideas of William Sharpe (1970) lead to the capital asset pricing model (CAPM), evaluating investment alternatives in terms of risk and return relative to the market as a whole. In CAPM, the riskier a stock, the greater the profit potential (variance would thus be opportunity). Efficient market theory8 posed that the market price incorporated perfect information, with random variations about an accurate price. This approach assumes a realm of risk, as defined in Table 11.2. For other realms, other tools are needed. Obviously, research and development would alleviate uncertainty. The argument would be that if something was important to know, a firm should invest up to the value of learning that something, in order to find out – but the catch is that the outcome of research is always uncertain. Scenario analysis using plausible future scenarios has been proposed, using the idea of maximizing expected utility in the worst-case scenario. A variant is stress testing, studying expected performance under extreme scenarios. However, one can always imagine an outcome that would make any decision fail – such as a nuclear holocaust, or a massive asteroid collision. In the case of risk, where outcomes and probability distributions are known, logic trees (decision trees of outcomes and associated probabilities) can apply.

110

Enterprise Risk Management in Finance

Table 11.2 Time

Evolution of risk management Theory

Essence

Relevance

Underpins derivatives, ultimate role of securities markets to efficiently allocate risk Basis for portfolio choices to optimize risk level at given return Indifference theory In perfect market, value Not true, suggesting Modigliani, Miller of company independent need for efficient capital structure and risk of capital structure mitigation through hedging Hedging should be left Markets compensate Capital asset pricing to investors investors for systematic model risk, but not idiosyncratic Sharpe et al. risk – can eliminate latter through hedging Options-pricing model Volatility of a security a Allows risk transfer, companies can price Black, Scholes, Merton key factor in options waiting price Price of a security driven Segmentation of CAPM Arbitrage pricing by a number of factors systematic risk into theory factors – if prices diverge Stephen Ross from expected, arbitrage can bring back into line There is a shareholder Stockholders avoid Underinvestment value in better risk low-risk/low return to problem avoid shifting wealth to management Myers, Smith, Stulz debt holders Allows deeper markets Variations in price over Binomial option time can be used to more for long-dated options pricing model accurately price options and options paying Cox, Ross, Rubinstein dividends Goal of risk management Managers try to control Risk management risk as a strategic set of to ensure firm has cash framework choices available for valueFroot, Scharfstein, enhancing investments Stein

Late 50s, State preference theory Efficient allocation of early 60s Arrow, Debreu resources & risks requires complete set of securities permitting hedging 1952 Mean variance Investors can analyze Harry Markowitz risk & expected return 1958

1960s

1973

1976

1977

1979

1993

Source: Extracted from Buehler et al. (2008).

If there are many branches to logic trees, Monte Carlo simulation is widely used for analysis. The good feature of Monte Carlo simulation is that any distribution assumption (and any systemic property) can be assumed. However, interpreting the output can easily become challenging.

Economic Perspective

111

Buehler et al. (2008)9 provided a review of their view of the evolution of risk management theory. Table 11.2 displays their timeline. Risk has been a fundamental topic of economic theory for centuries, and for many years, risk management was pretty much fully described by insurance and ad hoc hedging. The evolution outlined in Table 11.2 shows the appearance of derivatives which came to be used as a risk management tool. Financial theory has evolved to a much greater level of sophistication. Markowitz gave us the mean-variance approach, defining risk as variance, and providing a tool to identify the efficient frontier where risk and return were traded off. Markowitz’s model also allows correlations to be included, realizing that investment returns are not independent of the returns of other investments. Sharpe (1970)10 extended Markowitz’s work to the capital asset pricing model, where investments are evaluated in terms of risk and return relative to the market as a whole. In this view, the riskier the stock, the greater the expected profit. Thus this leads to the conclusion that risk is opportunity. Efficient market theory views the market price as incorporating perfect information. Prices then vary randomly around their appropriate equilibrium value. However, the complexities of life demonstrated in 2008 that every model leaves some bit of reality out.

The human factor Humans are known to have limited abilities to deal with extreme event probabilities.11 Sometimes, conventional statistical training does not aid this ability. Humans have been found to have biases, some of which affect their decisionmaking when they are trying to manage risks. A common mistake is overconfidence in one’s ability and knowledge. This also applies to experts. The problem of anchoring involves bias in estimations dependent upon initial estimates. Availability involves reliance upon memory of past experience as a guide in estimating event probability. Risk managers are called upon to estimate the likelihood of possible scenarios in stress testing, but they may be biased by recent events, especially those that make the news. Framing is a cognitive bias in which a view of risk depends upon context. For instance, gamblers who are ahead are notably risk averse, while those that are behind are risk seeking (determined to break even). Risk managers may do the same, depending upon company profitability in the recent past. Herding involves a tendency to mimic others, which often is found in investments. Essentially, herding can lead to bubbles, or to bank runs. Reality Humans have always been susceptible to bubbles. Charles Mackay (1841)12 reviewed a number, to include the Dutch tulip mania in the early 17th century,

112

Enterprise Risk Management in Finance

the South Sea Company bubble of 1711–1720, and the Mississippi Company bubble of 1719–1720. Patterson (2010)13 gave Isaac Newton’s famous quote complaining of losing £20,000 in the South Sea Bubble in 1720: ‘I can calculate the motion of heavenly bodies but not the madness of people.’ More recent problems include the London Market Exchange (LMX) spiral. In 1983, excess-of-loss reinsurance was popular, especially with Lloyd’s of London. Syndicates unintentionally paid themselves to insure themselves against ruin.14 These risks were viewed as independent – but in fact they were not, because they involved a cycle of hedging to the same pool of insurers. Hurricane Alicia was very damaging in 1983, and nearly brought down Lloyd’s of London, much of the damage being blamed on the LMX cycle of reinsurance. Black Monday, October 19, 1987 was another critical event, when the stock exchange nearly melted down. Some blamed portfolio insurance, based on efficient-market theory and models implementing computer trading to take advantage of what were viewed as temporary diversions from a fundamental price reflecting value. Yet another recent problem arose with Long Term Capital Management (LTCM), a firm created to benefit from Black–Scholes formulation of the value of derivatives. Lowenstein (2000)15 reviewed the case in depth, beginning with the theoretical model of the value of the new instrument of derivatives, LTCM’s spectacular success, and its ultimate collapse due to positions held in Russian banking in the latter part of 1998. LTCM was bailed out by the Federal government in the US because it was viewed as too big to be allowed to collapse. In 2002, highly popular information technology stocks plummeted. In the 1990s, venture capitalists were highly amenable to throwing money at any proposal suggesting a way to implement computer technology. Stock prices for many new startups skyrocketed, in hopes that each was the new IBM, or Microsoft, SAP, or Oracle. But most were not viable, and this market sector proved to be yet another bubble. Today, we are still suffering the aftereffects of the subprime mortgage collapse. Banks seemingly prospered in a climate of deregulation and merger begun in 1981. Many invested heavily in financial instruments created out of pools, including mortgages, many of which were generated by aggressive marketing to those who were offered homes beyond their ability to pay for. This was not viewed as a problem, as the housing market in most of the United States had little evidence of decline in value, and in some regions (California, Florida, Nevada) were vastly outperforming other types of investment return. However, in 2008 house values proved capable of declining in value, and many of the most aggressive investing banks were stuck with mammoth deficits. Some (not

Economic Perspective

113

all) such banks were bailed out by the federal government, and most credit this action with saving the US economy (if not the world economy). Difficulty in grasping extreme event probabilities has long been noted. Taleb (2007)16 notes that we are trained to consider fair coin flips to have a 0.5 probability of heads as well as tails. He proposes that if we observe 99 consecutive coin flips, statistical training steers us to assume that the next coin flip will have a probability of 0.5 for both heads and tails. Taleb argues that a more pragmatic estimate of events is that something is crooked. Taleb also discussed casino treatment of risk, one of their definite core competencies. Casinos have mechanisms in place (risk management) to assure they don’t go broke. But operating a casino is not completely immune to risk. Taleb related the four biggest losses casinos experienced in recent times: • A tiger bit a member of the Siegfried and Roy entertainment team, costing the casino about $100 million by one estimation. • A contractor was constructing a hotel annex, suffered losses in the project, and sued. He lost the suit, and tried to dynamite the casino. • Casinos (rightfully so) are required to file tax returns with the Internal Revenue Service. An employee charged with this duty failed to perform – not once but over a number of years. When the malfeasance was discovered, the casino was liable, and had to pay a huge fine as its license was at risk. • A casino owner’s daughter was kidnapped. In violation of law, he used casino money to raise the ransom demanded. While this was understandable, the casino was liable. These four examples represent very rare (hopefully) events, with little basis for accurate actuarial calculation. They fall into the scale of ignorance, in that they have outcomes that were probably deemed beyond the realm of likelihood. But firms must operate in environments that include the possibility of meteors hitting the earth and causing the end of life as we know it, of wars, of terrorism, and global warming threatening islands and port cities. Taleb presented the Black Swan problem. Most humans try to be scientific, and learn from their observations and history. But while nobody in Europe had seen a black swan, and had thus assumed they didn’t exist, when they settled Australia they found some, disproving their empirical hypothesis. Taleb also noted fallacies on the part of investors, who assume data is normally distributed. In practice, especially during bubble bursts, fat tails with higher extreme probabilities are often observed. Cognitive psychology can explain some of this. Kahneman and Tversky (2000)17 emphasized human biases from framing, with different attitudes toward risk found during winning and losing streaks. Humans also have been found to overestimate the probability of rare events,

114

Enterprise Risk Management in Finance

such as the odds of the next asteroid impacting the earth, or the risk of terrorists on airplanes. Akerlof and Shiller (2009)18 argued that standard economic theory makes too many assumptions; when human decisions are involved, historical data is not a good predictor of future performance. Risk mitigation ERM seeks to provide means to recognize and mitigate risks. The field of insurance developed to cover a wide variety of risks, external and internal, covering natural catastrophes, accidents, human error, and even fraud. Financial risk has been controlled through hedge funds and other tools over the years, often by investment banks. With time, it was realized that many risks could be prevented, or their impact reduced, through loss-prevention and control systems, leading to a broader view of risk management. Risk tolerance A key concept of risk management is risk tolerance. Risk tolerance bounds the risk a firm is willing to assume. It can be calculated as the maximum amount of surplus a company is willing to lose over a given period for a specific event. It also could express the most the company is willing to lose per year. Yet another definition is to express the probability that a capital adequacy ratio (the ratio of reserves to potential payouts) will fall below a given level. Another key concept is that organizations are in business to cope with risks in their area of expertise (core competence), but should shed risks that are not in this core. Insurance is the most common means to shed these non-core risks. Recent events Doherty et al. (2009)19 sought to explain the underlying causes of the 2008 global financial crisis, concluding that values, incentives, decision processes, and internal controls all played a role. Oversight and control were found lacking in the years prior to 2008, with heavy use of leverage on the part of investment firms that amplified risks, which in turn were hidden by the strong profit performance measurement. A negative view of risk management would be to ensure loss avoidance. A positive view would emphasize risk management as part of value creation. The insurance industry by its nature is focused on risk transfer. During the 2008 and subsequent period, while the insurance industry remained profitable for the greater part, it experienced significant declines in return as well as losses of risk capital. Doherty et al. proposed approaches to managing risk in this market:

Economic Perspective

115

• Solvency management focuses on limiting the probability of failure to levels in line with the organization’s risk tolerance. Solvency costs include lost future cash flow, opportunity costs of liquidation, and costs of regulatory intervention. Demand for insurance also is affected by credit ratings, which are in turn impacted by solvency. Thus the firm’s access to capital markets is affected, as is the firm’s reputation. Managing solvency risk requires balancing actual capital with the capital required to support portfolio risks as a specified risk tolerance. Strategies and tools to deal with unknown risks include: • slack (having extra cash), contingent equity (could be an option to put newly issued preferred shares in the case of stock price falling below a strike level and an event such as reaching a threshold for payment of claims). • mutualization (which apportions losses over a large pool of members). The insured parties end up sharing the losses of the insurer. A related concept is a pari-mutuel market allowing hedgers and speculators to place bets on events. • Profitable growth ensures funding for strategic investments, particularly after major losses. The problem is that when cash flow is short, management tends to pass up promising investment opportunities. Insurance against adverse events provides new capital in those cases. Reinsurance is a macro-level pooling of risks aimed at developing portfolios of diverse risks expected to have low correlations. Insurance operates on the assumption that the following anomalies are not present: • Moral hazard, which arises when the party being insured has prior knowledge that the probability of claim will be much larger than the norm, which is unknown by the insurer. An example is subprime loan initiators in the mid-2000s selling mortgages to parties known to have a high risk of default, and passing this risk on to other investors. • Adverse selection, which is a market process where undesired results arise due to imbalanced information access by contracting parties. An individual might buy cancer insurance if their doctor tells them they have unfavorable symptoms, but they don’t share this information with the insurer. Sometimes this exact circumstance involves the government requiring insurers to provide coverage to all applying parties regardless of pre-existing conditions (regulatory adverse selection). • Transparency reduces investor uncertainty, and can be obtained in part by eliminating non-core activity impact on reported earnings and capital. Market values were contended to be associated with stable earnings. If noncore risks are transferred to the nonrecurring category in corporate earnings reports, investors should receive a more accurate approximation of

116

Enterprise Risk Management in Finance

firm value. The issue of defining what things belong in recurring or nonrecurring categories does involve a level of subjectivity, lending itself to abuse. The basic argument is that honesty pays in the long run.

Conclusions Risk is traditionally modeled as the product of probability times severity. These two dimensions can be considered in different contexts.20 A formulation of expected frequency plotted against expected severity is appropriate for loss portfolios adding independent events. For single random events, it is more appropriate to view probability versus severity through mechanisms such as risk matrices. The conventional risk finance paradigm would suggest three strategies: If expected severity and expected frequency were both high, risk should be avoided through some control. If expected severity was high but expected frequency low, risk transfer through hedging, insurance, or risk mitigation would be appropriate. If expected frequency were high and expected severity low, diversification through pooling through frequency mitigation would be useful. If both expected frequency and expected severity were low, informal diversification would protect against these rare events. Powers et al. noted two problems with this strategic approach: 1. Low frequency/high severity events (ambiguity/uncertainty) call for risk transfer. This assumes that there will always be a counterparty willing to accept the transferred risk. However, in dire times, insurance companies have been observed to reject business. 2. Under conditions of high expected frequency, pooling through diversification is suggested if expected severity is low, while risk avoidance is recommended if expected severity is high. Higher expected severity events are often observed to have fat tails, which inhibit diversification. Powers et al. extended their analysis to the continuous case, where it is not necessary to establish boundaries between high and low expected frequencies or severities. Financial theory has developed a number of tools to supplement insurance and hedging as means of implementing risk management. Derivatives are intended as means to hedge in a more sophisticated manner. But the human factor enters into the equation when customers do not understand what they are purchasing (and possibly salespeople don’t either). Humans have a tendency to be exuberant that has manifested itself over and over with time, creating high variance in market prices that we know as bubbles. Buehler et al. concluded that companies have their own appropriate ratio of debt-to-equity related to the probability of incurring loss. Greater equity

Economic Perspective

117

capital than required will lead to inefficient capital use, as more profits will be needed to maintain its average per-share profitability. Insufficient equity capital risks default or financial stress, as well as limiting the firm’s ability to take advantage of new growth opportunities. Optimal debt level is a function of a firm’s key market, financial, and operating risks. The firm’s ability to mitigate those risks varies. The risks within the firm’s competence to readily deal with should be retained, while the risks that they are less capable of dealing with should be transferred.

12

British Petroleum Deepwater Horizon

Introduction The Macondo well, operated by BP, aided by driller Transocean Ltd. and receiving cement support from Halliburton Co., blew out on April 20, 2010, leading to 11 deaths. The subsequent 87-day flow of oil into the Gulf of Mexico dominated news in the US for an extensive period of time, polluting fisheries in the Gulf as well as the coastal areas of Louisiana, Mississippi, Alabama, Florida, and Texas. The cause was attributed to defective cement in the well. Subsequent studies by regulatory agencies could detect no formal risk assessment by BP.1 Risk management in this context is not traditional insurance risk management, nor the financial risk management that made hedging famous. Rather, risk management in this broader sense means responsibility to recognize hazards and take action to prevent them to the extent possible. The Deepwater oil spill demonstrates such a context, and the disastrous consequences of not accomplishing such risk management. Our economy gets involved in many extraction activities, most of which involve high degrees of risk. In the 19th century, the prevailing attitude was that miners were paid above the market level of wages in compensation for the risks to their lives involved in their work. We still have mining, but the 20th century saw more activity in petroleum extraction, again with high levels of life-threatening events. Government has reflected society’s changing attitudes. This evolution is still going on, but we feel that the trend is for some group to take the attitude that for every catastrophe, ‘we will not rest until that never happens again.’ The Occupational Safety and Health Act of 1972 went far to make workplaces safer. It has taken many years for our society to work out this transition to safer work practices as a requirement, even at the expense of operating cost. (In fact, we remember when one of the initial motivators for firms to accept OSHA was that ‘it pays’ through reduced legal suits and fines. There have been many acts passed in the US to regulate all industry, including all phases of petroleum production. 118

British Petroleum Deepwater Horizon

119

Deepwater horizon The Macondo Mississippi Canyon Block 252 oil well was 5000 feet deep, 40 miles southeast of the Louisiana shore. It was owned by Transocean Ltd., leased to British Petroleum. On April 20, 2010, the well erupted after a blowout resulting in explosion and fire. Eleven workers were killed, another 17 injured. Two days after the explosion, the entire rig sank. Gas and mud from the well had triggered an explosion that sank the platform and cut the well pipe at the sea floor. A containment cap was in place, but failed. The wellhead was compromised, discharging crude at the rate peaking at 9000 barrels per day into the Gulf of Mexico. The oil spill shut down Louisiana seafood.2 The oil spill persisted from April until July, gushing nearly 5 million barrels of oil before it was finally stopped. The response effort was massive, with some 2600 vessels deployed. Shoreline protection efforts included 4.4 million feet of sorbent boom, and there was heavy use of oil dispersants. Silves and Comfort provided a listing of conditions describing BP risk management with respect to Deepwater Horizon (Table 12.1).

Table 12.1 BP risk factors Conditions

Tolerated conditions ensuing in accident

Technical factors

Technological and environmental learning advanced in response and recovery stage; Corporate information sharing grew; Government regulation was reformed.

Outside expertise

Drew on world-class expertise and equipment, but decision making was not highly adaptive; Information sharing satisfactory at best; Public relations failures.

Outside consulting

Consulted marine science community.

Government role

BP had difficulty recognizing interdependency with government.

Interorganizational factors

Spill planning sophisticated; State & Federal officials had plan and system, but not well prepared for discharge a mile under sea level that continued for months.

Socio-technical issues Principal–agent problems between BP & Transocean; Profit/safety conflicts; Overconfidence in sea floor containment cap; Experiment in drilling 1 mile deep, 4 further miles into the sea; Problems using National Incident Management System in parallel with National Contingency Plan and Oil Pollution Act. Source: Extracted from Sylves and Comfort (2014).

120 Enterprise Risk Management in Finance

The Oil Pollution Act of 1990 Under this US act, spillers were liable for much greater amounts than previously, with stiffer civil and criminal penalties. Spillers were required to pay for the cleanup of oil spills and to compensate those economically injured. States were allowed to impose unlimited liability on shippers. A federal fund financed from an existing 5 cent per barrel oil tax was created to cover cleanup and compensation costs that spillers did not cover. Shippers were responsible for drafting worst-case oil spill response plans for quick cleanup. Oil tankers crossing US waters were required to be double-hulled by 2010. Federal oil spill response capability was reinforced by positioning response teams across the country, coordinated by a national command center. The US President was authorized to take control of oil spill cleanup. A multi-agency oil pollution panel was established to coordinate federal research. Recovery factors in Macondo As is normal in highly visible disasters, a great deal of blame was passed on to BP, as well as to the Minerals Management Service (MMS), the regulatory agency most involved. Government regulatory planning was in place, but the mile-deep sea floor well was beyond regulatory experience. While spill planning had existed, BP expertise failed. BP was accused of tolerating conditions resulting in what Perrow3 referred to as ‘normal accidents.’ The reaction after the event, however, was positive in that BP invested a high level of effort, and government regulation was reformed. BP accessed whatever equipment and expertise it could obtain. The technical issues involved are shown in part in Table 12.2. Table 12.2

Factors in recovery

Oil properties

Light and gaseous oil, evaporating quickly in warm climate

Volume

Months of discharge created immense volume, very difficult to stop

Challenges

Protection of migratory water fowl, fisheries, wetlands, shorelines Interference with water freight, tourism, recreation Confidence in oil platform safety jeopardized

Environmental remediation

Skimming In situ burning (to a slight degree) Berming Absorbent booms Animal rescue Massive use of dispersants on sea floor Reliance in natural forces

Source: Extracted from Sylves and Comfort (2014).

British Petroleum Deepwater Horizon

121

The BP oil spill led to a number of socio-technical mitigation and/or prevention measures for the industry. There was more regulation and testing of containment cap technology. There was a moratorium imposed on deep-water drilling in the Gulf. Oil platform drilling and operations were more stringently regulated. Corporate safety culture improvement was induced. Oil companies shared more information on deepwater drilling technology. Oil spiller liability was increased. Risk management factors Borison and Hamm4 addressed risk management in the BP case. They cited the dominant publicly accepted views to be: 1. the disaster was a black swan, which could not have been foreseen; and 2. BP and its regulators were incompetent black sheep. Borison and Hamm considered a third perspective, seeking insight into what better risk management could have done. The uncertainty (ambiguity) of the case makes traditional tools such as VaR analysis inappropriate; while there is a statistical database of oil spills, they are hopefully rare events. So tools such as stress testing were proposed as more appropriate. MMS was responsible for regulation, and had a system in place to deal with oil spills that included tools to estimate likelihood and impact based on historical data. However, MMS was not required to quantify risk of a large spill from the Macondo well, and appears not to have done so. MMS relied on qualitative analysis in the form of developers specifying a worst-case scenario and giving detailed plans of response. Thus BP was aware of a large oil spill, although no probability estimates were recorded. Thus BP did not underestimate the probability of the spill – rather, BP didn’t estimate it at all. When risk analysis is solely qualitative, there is a great deal of personal interpretation. For instance, a ‘worst-case’ could imply a 1 in 100 event, or a 1 in 1,000,000 event. Borison and Hamm suggested a Bayesian analysis instead. Eckle and Burgherr (2013)5 applied Bayesian analysis to oil chain fatalities, using a Poisson distribution for accident frequency and a gamma distribution for accident severity. Monte Carlo simulation modeling was used to combine these factors. Abbasinejad et al. (2012)6 studied Iranian energy consumption since 1983, finding it unbalanced, with excessive use of oil and gas damaging the environment and encouraging a shift to natural gas and electricity. High growth in energy demand was expected, and a Bayesian vector autoregressive forecasting model of energy consumption was applied. Bayesian analysis bases probabilities on judgment rather than on empirical observations. While statistically it

122

Enterprise Risk Management in Finance

is of course preferable to have data as a basis, risk management often involves situations where the relevant data is not available. Judgment can be based upon expert opinion, which is the preferred source. Oil production was given as a case where data was widely available. Oil price, however, depends upon many highly variable factors, including war, piracy, dry wells, pipeline rupture, train wrecks, and drilling company bankruptcy. Bayesian analysis relies upon expert judgment as the basis for estimating the magnitude and/or probability of these factors.

Conclusions The BP spill is only the most recent salient event demonstrating the risk to the environment from energy policy. One alternative is nuclear energy, with far less probability of a spill. However, there is of course the fear of such a spill being far more dangerous than mere petroleum. Our global economy requires a great deal of transportation, and we have developed a complex system relying on petroleum-based energy. There are known relative risks of spills in the system. Drilling in the ocean taps vast reserves, and occurs at a distance from population centers. But as the BP case demonstrates, spills can be dramatic. Conversely, land-based drilling creates a need to transport oil. Pipelines are safer than trains, but political issues make it difficult to create new pipelines while existing (old) railroad track becomes the major shipment technique. We could switch to electric cars and trucks to vastly reduce the use of petroleum for transportation – but the added strain on the electrical generation system would open new opportunities for disaster. The problem is essentially that our economies are complex and interrelated. Borison and Hamm inferred three major lessons from the BP oil spill: 1. Assessment of unusual events is becoming important enough to call for greater rigor. The nature of such events means that we have relied on qualitative assessment. Borison and Hamm say we need something better. We conclude that the nature of such rare events mean that we probably have to live with qualitative assessment, but techniques such as scenario analysis, stress testing, and Bayesian inference can be applied. 2. Borison and Hamm call for greater formal attention to developing and evaluating customized preparations and responses. They accuse the response to the Macondo well event to have been ‘off-the-shelf.’ We would call for having a variety of such off-the-shelf remedies available that can be quickly implemented. Since it is nearly impossible to anticipate unusual catastrophes, a flexible response would seem appropriate.

British Petroleum Deepwater Horizon

123

3. Borison and Hamm call for systematic learning to ensure that risk management improves over time. We think that they are absolutely correct. Understanding the interactions of complex systems is critical in enabling effective response. The BP oil spill demonstrates a common risk context, involving massive engineering undertakings such as seen in construction projects of many kinds. Tools to plan for risk that were suggested were stress testing (war games, what-if scenario analysis, many other names) and simulation using subjective inputs.

13

Bank Efficiency Analysis

Introduction In today’s economy, the banking industry is of great importance. With the availability of new technology and the internet, more and more organizations are entering some aspect of the banking business and this results in intense competition in the financial services markets. Major domestic banks continue to pursue all the opportunities available to enhance their competitiveness. Consequently, performance analysis in the banking industry has become part of their management practices. Top bank management wants to identify and eliminate the underlying causes of inefficiencies, thus helping their firms to gain competitive advantage, or, at least, meet the challenges from others. Traditionally, banks have focused on various profitability measures to evaluate their performance. Usually multiple ratios are selected to focus on the different aspects of operations. However, ratio analysis provides a relatively insignificant amount of information when considering the effects of economies of scale, the identification of benchmarking policies, and the estimation of overall performance measures of firms. As alternatives to traditional bank management tools, frontier efficiency analyses allow management to objectively identify best practices in complex operational environments. Five different approaches, namely, Data Envelopment Analysis (DEA), Free Disposal Hull (FDH), Stochastic Frontier Approach (SFA), Thick Frontier Approach (TFA), and Distribution Free Approach (DFA) have been reported in the literature as methods to evaluate bank efficiency.1 These approaches primarily differ in how much restriction is imposed on the specification of the best practice frontier and the assumption of random error and inefficiency. Compared to the other approaches, DEA is a better for organizing and analyzing data since it allows efficiency to change over time and requires no prior assumption on the specification of the best practice frontier. Thus, DEA is a leading approach for the performance analysis of the 124

Bank Efficiency Analysis

125

banking industry in the academic literature. However, the DEA frontier is very sensitive to the presence of the outliers and statistical noise, which indicates that the frontier derived from DEA analysis may be warped if the data are contaminated by statistical noise. On the other hand, DEA is not good for predicting the performance of other decision-making units. As a result, artificial neural networks (ANNs) have recently been introduced as a good alternative for estimating efficiency frontiers for decision makers (Wang, 2003). 2 The idea of combination of neural networks and DEA for classification and/ or prediction was first introduced by Athanassopoulos and Curram (1996). 3 They treated DEA as a preprocessing methodology to screen training cases in a study of forecasting the number of employees in the healthcare industry. After selecting samples, the ANNs are then trained as tool to learn a nonlinear forecasting model. Costa and Markellos (1997)4 analyzed London underground efficiency with time series data. They explained how the ANNs results are similar to corrected ordinary least squares (COLS) and DEA. However, ANNs offer advantages in decision-making by denoting the impact of constant vs. variable returns to scale or congestion areas. Fleissig et al. (2000)5 employed neural networks for cost functions estimation. They found convergence problems when the properties of symmetry and homogeneity were imposed on the ANNs. Santin et al. (2004)6 used a neural network for a simulated nonlinear production function, and compared its performance with traditional alternatives like stochastic frontier and DEA under conditions of different numbers of observations and noise. Pendharkar and Rodger (2003)7 used DEA for data screening to create a subsample training data set that is ‘approximately’ monotonic, which is a key assumption in certain forecasting problems. Their results indicated that the predictive power of an ANN that is trained on the ‘efficient’ training data subset is stronger than the predictive performance of an ANN that is trained on an ‘inefficient’ training data subset. As two nonparametric models, there are many similarities between ANNs and DEA models such as:8 • Neither DEA nor ANNs make assumptions about the functional form that links its inputs to outputs. • DEA seeks a set of weights to maximize technical efficiency, whereas ANNs seek a set of weights to derive the best possible fit through observations of the training dataset. Bank branch efficiency is a comprehensive measure using various performance aspects with a number of financial variables. This indicates that the relationship between bank branch efficiency and multiple variables is highly complicated and nonlinear. For example, an efficiency improvement for a bank branch from 0.5 to 0.6 might simply be the result of personnel cost reduction.

126 Enterprise Risk Management in Finance

But the improvement of its efficiency from 0.8 to 0.9 could be due to many causes, for example, scale economy, or an increase in several outputs. ANNs have been viewed as a good tool to approximate numerous nonparametric and nonlinear problems. Thus, the banking industry provides good opportunities for the applications of ANNs. To the best of our knowledge, there are no studies using ANNs dealing with bank branch efficiency, but this chapter presents a DEA-NN approach to evaluate the branch performance of a big Canadian bank. The results are also compared with the corresponding efficiency ratings obtained from DEA. The fact that the DEA property of unit invariance is similar to the property of scale preprocessing required by NNs validates the rationale to implement a comparison between pure DEA results and DEA-NN results. Based on this analysis, similarities and differences between ANNs and DEA models are further investigated.

Data envelopment analysis (DEA) DEA is used to establish a best practice group amongst a set of observed units and to identify the units that are inefficient when compared to the best practice group. DEA also indicates the magnitude of the inefficiencies and improvements possible for the inefficient units. Consider n DMUs to be evaluated, DMUj (j = 1,2 ... n) that consume the amounts Xj = {xij} of m different inputs (i = 1, 2, ... , m) and produce the amounts Yj = {yrj} of r outputs (r = 1 , ... , s). The input-oriented efficiency of a particular DMU0 under the assumption of variable returns to scale (VRS) can be obtained from the following linear programs (input-oriented BCC model:9

min

θ , λ , s+ , s −

s.t .

r r z0 = θ − ε ⋅ 1s + − ε ⋅ 1s −

Y λ − s + = Y0

(13.1)

θ X0 − X λ − s − = 0 r 1λ = 1 λ , s+ , s− ≥ 0 where s+ and s− are the slacks in the system. Performing a DEA analysis requires the solution of n linear programming problems of the above form, one for each DMU. The optimal value of the variable θ indicates the proportional reduction of all inputs for DMU0 that will move it onto the frontier which is the envelopment surface defined by the efficient DMUs in the sample. A DMU is termed efficient if and only if the optimal value θ* is equal to 1 and all the slack variables are zero. This model

Bank Efficiency Analysis

127

allows variable returns to scale. The dual program of the above formulation is illustrated by:

max µ ,ν

s.t .

w0 = µ T Y0 + u0 v T X0 = 1 r µ T Y − v T X + u0 1 ≤ 0

r −µ ≤ − ε • 1 r − vT ≤ − ε • 1 free

(13.2)

T

u0

r If the convexity constraint (1λ = 1) in (13.1) and the variable u0 in (13.2) are removed, the feasible region is enlarged, which results in the reduction in the number of efficient DMUs, and all DMUs are operating at constant returns to scale (CRS). The resulting model is referred to as the CCR model.10 DEA has a rich literature base of over 3000 papers and several books for those who require detailed information on this technology. In summary, each DEA model seeks to determine which of the n DMUs define an envelopment surface that represents best practice, referred to as the empirical production function or the efficient frontier. Units that lie on the surface are deemed efficient in DEA, while those units that do not are termed inefficient. DEA provides a comprehensive analysis of relative efficiencies for multiple input-multiple output situations by evaluating each DMU and measuring its performance relative to an envelopment surface composed of other DMUs. Those DMUs are the peer group for the inefficient units known as the efficient reference set. As the inefficient units are projected onto the envelopment surface, the efficient units closest to the projection and whose linear combination comprises this virtual unit form the peer group for that particular DMU. The targets defined by the efficient projections give an indication of how this DMU can improve to become efficient.

Neural networks Neural networks provide a new way for feature extraction (using hidden layers) and classification (e.g., multilayer perceptrons). In addition, existing feature extraction and classification algorithms can also be mapped into neural network architectures for efficient (hardware) implementation. Backpropagation neural network (BPNN) is the most widely used neural network technique for classification or prediction.11 Figure 13.1 provides the structure of the backpropagation neural network.

128 Enterprise Risk Management in Finance

W11

Z1

W12

Z2

V1

X1

V2

Wn2 ŷ

Wn1 Vm

W1m Xx Wnm Zm

Input layer Figure 13.1

Hidden layer

y–ŷ

Output layer

Backpropagation neural networks

With backpropagation, the related input data are repeatedly presented to the neural network. The output of the neural network is compared to the desired output and an error is calculated in each iteration. This error is then backpropagated to the neural network, and used to adjust the weights so that the error decreases with each iteration and the neural model gets closer and closer to producing the desired output. This process is known as training. When the neural networks are trained, three problems should be taken into consideration. First, it is very challenging to select the learning rate for the nonlinear network; if the learning rate is too large, it leads to unstable learning, but on the contrary, if the learning rate is too small, it results in incredibly long training iterations. Second, settling in a local minimum may be good or bad, depending on how close the local minimum is to the global minimum and how accurate an error is required. In either case, backpropagation may not always find the correct weights for the optimum solution. We may reinitialize the network several times to guarantee the optimal solution. Finally, the network is sensitive to the number of neurons in its hidden layers; too few neurons can lead to under-fitting, however, too many neurons can cause overfitting. Although all training points are well fit, the fitting curve takes wild oscillations between these points. In order to solve these problems, we preprocess the data before training. The scale of data values is bounded to 10 and 100 by dividing by a constant value, such as 10 or 100. The weights are initialized with a random decimal fraction ranging from –1 to 1. Moreover, there are about 12 training algorithms for BPNN. After preliminary analyses and trial,

Bank Efficiency Analysis

129

we chose the fastest training algorithm, the Levenberg–Marquardt algorithm, which can be considered as a trust-region modification of the Gauss–Newton algorithm.

The energy we use One hundred and forty-two branches of a big Canadian bank in the Toronto area were involved in the analysis. The data covered the period October to December, 2001. Summary statistics for the inputs and outputs are reported in Table 13.1. From the table, no consistent trend in the data was found over the time horizon of analysis. There are no significant variations in terms of deposits and loans. Bank branch efficiency analysis Comparing the performance of the ANN for both efficient and inefficient training data subsets in the healthcare forecasting problem, Pendharkar and Rodger (2003)12 indicate that the predictive performance of an ANN that is trained on the efficient training data subset is higher than the predictive performance of an ANN that is trained on the inefficient training data subset. Therefore, DEA-efficient branches are all selected as training data in building a NN for branch efficiency analysis. Troutt et al. (1995)13 suggest that training

Table 13.1 Summary statistics of data Other general Personnel expenses

October 2001

Average Standard Deviation Min Max

Average November Standard Deviation 2001 Min Max Average Standard December Deviation 2001 Min Max

55,222 39,284

38,239 36,023

9,323 401,584

4,100 337,833

52,427 38,719

29,015 29,467

13,286 413,439

5,032 283,236

54,156 39,897

27,654 26,958

14,206 433,591

3,029 264,904

Deposits

Loans

Revenues

98,321,387 123,239,727

150,972 93,409

2,912,171 6,456,600 535,562,721 1,147,686,344

25,666 918,611

90,967,688 59,801,588

91,887,319 60,326,161

98,261,715 123,526,412

163,285 96,536

2,821,525 6,471,055 542,933,642 1,159,694,084

25,618 956,827

92,382,692 61,138,567

98,350,514 123,919,287

159,142 94,745

2,832,913 6,549,843 558,736,191 1,177,876,733

33,409 889,346

130 Enterprise Risk Management in Finance

data for nonparametric models should be at least ten times the number of input variables. Since we have five inputs in the ANNs, a minimum of 50 training examples was necessary to have reasonable learning of connection weights. Since we had less than ten efficient branches (efficiency score is 1) for each training, and a minimum of 50 branches for training was desirable, we used a grouping technique by roughly pre-specified cutoff efficiency threshold values of 0.98, 0.8 and 0.5. Thus, we obtained sufficient branches in our ‘efficient’ set. Note that the word efficient in the DEA context means DMUs with an efficiency score of 1. Since in our case the ‘efficient’ set does not only include all DMUs with an efficiency score of 1, we have used quotation marks to indicate that the word ‘efficient’ has a slightly different meaning from the DEA context. The same logic applies for the word inefficient (efficiency score < 1). The performance of ‘efficient’ and ‘inefficient’ branches was then tested on the entire dataset so that industry efficiency could be predicted and analyzed. The neural network was trained for each month using different combinations of subsets. A computer program was written using Matlab language as well as the neural network add-in module of Matlab. Table 13.2 presents the parameters of the estimated neural networks in the algorithm. The estimated neural network incorporates four tanh hidden units, and the Levenberg–Marquardt algorithm is employed for training. After training, we obtained three estimated neural networks for each month, the best two of which were used for the branch efficiency prediction for each month. The best two estimated neural networks are denoted as DEA-NN1 and DEA-NN2, using training data subset S1◡ S2◡ S3 or S1◡ S2◡ S4, respectively. The results are consistent with Pendharkar et al.’s (2003) findings that it is better to train ANN on an efficient training data subset in order to improve the predictive performance of an ANN.

Table 13.2 Estimated neural network parameters Concept

Result

Data pre-processing

Input/Raw data/1000 (Scale preprocessing) [10, 500] Output/ efficiencies invariant [0,1] 5–10–1 tanh/linear Levenberg–Marquardt 20000 0.985 0.6 0.0001

Network architecture Activation function: hidden/output Algorithm Epochs (max.) R2 learning rate mean square error

Bank Efficiency Analysis

131

Based upon the best two estimated neural networks, the branch efficiencies are calculated. The results are shown in Tables 13.3 and 13.4. Table 13.3 gives results of number of branches corresponding to each efficiency interval for three months. We grouped dataset SS2 into four categories depending on the efficiency value intervals (0.98, 1], (0.8, 0.98], (0.5, 0.] and (0, 0.5], then did a statistics analysis of DEA-NN efficiencies. Table 13.4 shows that the efficiency scores for some DMUs are larger than 1 in the DEA-NN model, which is not allowable in the DEA context. This occurs in the DEA-NN model since NNs actually generate a stochastic frontier based upon the ‘efficient’ DMUs due to the statistical and probabilistic (and thus varying) properties embedded in NNs. Neural network models have the ability to approximate complex nonlinear functions in a semi-parametric fashion. We observe that the bank branch performance is very close in the three-month period, since there is no significant change in the bank’s policy and the economic conditions during the examined period. Table 13.3 Number of branches corresponding to each efficiency interval

No. of branches by CCR in October No. of branches by DEA-NN1 in October No. of branches by DEA-NN2 in October No. of branches by CCR in November No. of branches by DEA-NN1 in November No. of branches by DEA-NN2 in November No. of branches by CCR in December No. of branches by DEA-NN1 in December No. of branches by DEA-NN2 in December

(0.98,1]

(0.8,0.98]

(0.5,0.8 ]

(0,0.5]

5 1 5 9 9

11 20 11 26 31

115 115 113 78 77

11 6 13 29 25

2

27

95

18

9 2 6

26 47 26

78 62 89

29 31 21

Table 13.4 Statistical results corresponding to each efficiency interval

Statistic results by CCR of October Statistic results by DEA-NN1 of October Statistic results by DEA-NN2 of October Statistic results by CCR of November Statistic results by DEA-NN1 of November Statistic results by DEA-NN2 of November Statistic results by CCR of December Statistic results by DEA-NN1 of December Statistic results by DEA-NN2 of December

Max

Min

1.00 1.01 1.02 1.00 1.04 0.99 1.00 1.04 1.17

0.38 0.47 0.40 0.21 0.28 0.40 0.21 0.36 0.26

Standard Mean Median Deviation 0.66 0.67 0.66 0.66 0.67 0.66 0.66 0.68 0.68

0.66 0.66 0.65 0.64 0.63 0.68 0.64 0.69 0.69

0.12 0.12 0.13 0.18 0.18 0.17 0.18 0.18 0.17

132 Enterprise Risk Management in Finance

Inefficient branches can improve their performance by mimicking the practices of their efficient reference set. Furthermore, even a small improvement can result in large monetary savings. However, it is difficult to improve a branch from 0.90 to 0.95, while it is relatively easier to implement practices that will improve the branch efficiency score from 0.6 to 0.7, which can result in rapid improvement and substantial cost savings. This concept can be demonstrated using the DEA-NN2 model as an example. Currently only 2 of the 142 branches are ‘efficient’ in the DEA-NN2 model. The remaining branches are distributed as follows: Table 13.5 Efficiency score distribution Efficiency score intervals No. of branches by DEA-NN2 in November

(0.98,1]

(0.8,0.98]

(0.5,0.8]

(0,0.5]

2

27

95

18

If the four least efficient branches could improve their efficiency rating to the 0.5– 0.8 range while all the other branches remained the same, the branches involved could save 18% of their costs on average (from Table 13.6). The DEA-NN approach can provide guidance for inefficient branches to improve their performance to any efficiency rating that management thinks necessary. To verify the rationality of our proposed DEA-NN approach, a regression analysis between the efficiency result achieved by our current DEA-NN approach and that by the normal DEA model is conducted. In the regression, a unity (1) slope, a zero value of intercept and unity R2 coefficient indicate that our current DEA-NN results provide a good estimation for the DEA result. The regression results based on three months’ data are shown in Tables 13.5, 13.6, and 13.7, where slope, intercept and r-squared coefficient are calculated. The predicted efficiency has a strong correlation with that calculated by DEA, which indicates that the predicted efficiency is a good proxy to classical DEA results. Table 13.6 Implication of slight efficiency improvement on branch costs

Transit no. 21 115 1 34 Total

Current score

Improved score

Current expenses (p.a.)

Difference (p.a.)

0.27 0.39 0.45 0.48

0.6 0.6 0.6 0.6

$80,295 $41,518 $90,721 $157,170

$26,554 $8,652 $14,053 $18,483

18%

133

Bank Efficiency Analysis

From Table 13.7, the DEA-ANN results are highly correlated with those obtained in the DEA CRS techniques. Short-term efficiency prediction For short-term efficiency prediction, another neural network (DEA-NN3) is applied. We used the October data set for training and DEA-NN3 is then applied to the November and December datasets to predict the bank branches’ efficiency ratings. Results are shown in Tables 13.8 and 13.9 respectively. Postprocessing the calculated efficiencies is accomplished by regression analysis between DEA-NN results and CCR DEA results E2. On the whole, the predicted efficiency has similar correlation with that calculated by DEA, especially the DEA-NN3 results for November with an r-squared coefficient of 0.71, which indicates that the predicted efficiency is to some extent a proxy to classical DEA results. Table 13.7 Regression analysis for branch efficiency prediction using October data Parameter DEA-NN1 for October DEA-NN2 for October DEA-NN1 for November DEA-NN2 for November DEA-NN1 for December DEA-NN2 for December

Slope

Intercept

R 2 coefficient

0.88 0.98 0.90 0.85 0.81 0.64

0.09 0.01 0.08 0.10 0.14 0.26

0.92 0.95 0.89 0.91 0.80 0.67

Table 13.8 Number of branches in each efficiency interval Efficiency score interval

(0.98,1]

(0.8,0.98]

(0.5,0.8 ]

(0,0.5]

9 16

26 38

78 84

29 4

9 9

26 38

78 84

29 11

Factor

Max Min

Mean

Median

Standard Deviation

Statistic results by CCR of November Statistic results by DEA-NN3 of November Statistic results by CCR of December Statistic results by DEA-NN3 of December

1.00 1.33 1.00 1.31

0.66 0.77 0.66 0.74

0.64 0.76 0.64 0.73

0.18 0.17 0.18 0.16

No. of Branches by CCR of November No. of Branches by DEA-NN3 of November No. of Branches by CCR of December No. of Branches by DEA-NN3 of December

Table 13.9 DEA-NN3 results

0.21 0.27 0.21 0.26

134 Enterprise Risk Management in Finance

Table 13.10 Regression analysis for short-term efficiency prediction Parameter DEA-NN3 for November DEA-NN3 for December

Slope

Intercept

R 2 coefficient

0.68 0.53

0.32 0.39

0.71 0.56

Table 13.11 Comparison of best-practice branches November

December

DEA efficient

DEA-NN3 efficient

DEA efficient

DEA-NN3 efficient

#3, #13, #31, #40, #49, #64, #81, #92

#3, #49, #64, #40, #81, #92, #104, #131, #93,#110, #16, #29, #135, #91

#3, #4,, #49, #64, #81, #91,#127

#3, #49, #64, #81, #93, #91,#131, #110

Table 13.11 presents the comparison of best-practice branches by DEA and DEA-NN3 in November and December respectively. It can be seen that DEA-NN always has more efficient units on the frontier, since neural networks have the flexibility to solve complex problems where the main information, or ‘knowledge,’ lies implicitly in the data. More good performance patterns in the ‘efficient’ (but not pure DEA-efficient) region are explored so that these inefficient DMUs by DEA are termed efficient by NNs. Neural networks have the ability to approximate complex nonlinear functions in a semi-parametric fashion and provide the main basis for adaptive learning systems. Therefore, by capturing performance patterns and self-learning, neural networks can always generate more efficient units on the frontier.

Conclusions This chapter presents a DEA-NN study applied to the branches of a big Canadian bank. The results are comparable to the normal DEA results. However, the DEA-NN approach produces a more robust frontier and identifies more efficient units since more good performance patterns are explored. Furthermore, the DEA-NN approach identifies those less-than-optimal performers, and suggests areas in which their performance can be improved to attain better efficiency ratings. We conclude this section with a comparison of the two methodological approaches in Table 13.12. In summary, the neural network approach requires no assumptions about the production function (the major drawback of the parametric approach) and is highly flexible.

Bank Efficiency Analysis

135

Table 13.12 Comparison of DEA and DEA-NN to efficiency measurement

DEA

Neural Network

Similarities Nonparametric No assumptions about the functional form that links its inputs to outputs Optimal weights to maximize the efficiency Unit and scale invariant

Nonparametric No assumptions about the functional form that links its inputs to outputs Optimal weights to derive the best possible fit Scale preprocessing

Differences Medium assumptions about functional form and data Medium flexibility Many theoretical studies/applications on efficiency Low cost of software, estimation time

Low assumptions about functional form and data High flexibility Few theoretical studies High cost of software, estimation time

14

Catastrophe Bond and Risk Modeling

Introduction On May 12, 2008, the Wenchuan earthquake occurred in Sichuan province of China, killing at least 69,000. This great disaster caused widespread damage to the infrastructure and huge economic losses to Chinese society. The loss to the Chinese insurance sector was in excess of 65 million RMB about 70 days after the quake. Figure 14.1 displays insurance claim payoffs relative to the days elapsing after the Wenchuan earthquake. Catastrophe (cat) events such as the Wenchuan earthquake and the Swine Flu epidemic of recent years have motivated (re)insurance companies to create many cat risk instruments in order to hedge high risk exposures from natural disasters. The first cat instrument, a cat equity put option, was issued to offer the cat option owner the right to issue convertible preferred shares at a fixed price. This instrument was issued by RLI Corp. in 1996. Since then, the contingent capital market has grown very rapidly due to the increase of unanticipated catastrophic events. Therefore, there is a great demand for insurance and reinsurance companies to appropriately price contingent instruments such as cat bonds.

Catastrophe risk instruments Catastrophe bonds, or cat bonds, are the most common type of cat risk-linked securities. Cat bonds have complicated structures, and refer to a financial instrument devised to transfer insurance risk from insurance and reinsurance companies to the capital market. The payoff from cat bonds is dependent on the qualifying trigger event(s) such as natural disasters, such as earthquakes, floods and hurricanes, or manmade events such as fire, explosions and terrorism. We will review the modeling approaches of cat bonds as follows.

136

Catastrophe Bond and Risk Modeling

137

60 50

Total

Payoff

40 30 Property insurance 20 10 0

Life insurance

0

10

20

30

40

50

60

70

t (in days) Figure 14.1 Insurance payoffs (million RMB) for the Wenchuan earthquake (from www. circ.gov.cn)

Traditional derivative pricing approaches use Gaussian assumptions, but these are not appropriate when applied to instruments such as cat bonds because of the properties of the underlying contingent stochastic processes. There is evidence that catastrophic natural events have (partial) power-law distributions associated with their loss statistics.1 This is not compatible with the traditional log-normal assumption of derivative pricing models. There are also well-known statistical difficulties associated with the moments of power-law distributions, thus rendering it impossible to employ traditional pooling methods and consequently the central limit theorem. Several studies have examined pricing models with respect to catastrophe derivatives such as cat bonds. Geman and Yor2 analyze catastrophe options with payoff (L(T)−K) + where L(T) is the aggregate claim process modeled by a jumpdiffusion process.3 Dassios and Jang4 used a doubly-stochastic Poisson process for claim processes to price catastrophe reinsurance contract and derivatives. Jaimungal and Wang5 studied the pricing and hedging of catastrophe put options (CatEPut) under stochastic interest rates with a compound Poisson process. In contrast, Cox and Pedersen6 priced a cat bond under a term structure model, together with an estimation of the probability of catastrophic events. Lee and Yu7 adopted a structural approach to value the reinsurance contract, using the idea of credit risk modeling in corporate finance.8 This allows the reinsurer to transfer the risk to the capital market via cat bonds and, in effect, to reduce the risk of the reinsurer’s default risk. Since the payments from cat bonds cannot

138 Enterprise Risk Management in Finance

be replicated by the ordinary types of securities available in financial markets, the pricing has to be done using an incomplete market model. The most important feature of cat bonds is their conditional payment. Trigger conditions are generally divided into three categories: indemnity-based trigger conditions, index-based trigger conditions and parametric trigger conditions. An indemnity trigger involves the actual losses of the bond-issuing insurer. It was very popular when the catastrophe bond market emerged. An industry index trigger involves an index created from property claim service (PCS) loss estimates. A parametric trigger is based on quantitative parameters of the catastrophe event, for example earthquake magnitude, central pressure, wind pressure, wind speed, hurricane rainfall, and so on. This chapter focuses on the choice of catastrophe loss model.

Loss model The catastrophe loss model is very important to catastrophe derivatives pricing. Table 14.1 presents various loss models used in the literature. There are three sets of popular models that are widely discussed in recent catastrophe bond pricing literature: the compound Poisson model, the jump-diffusion model and the double exponential jump-diffusion model.9 Demonstration of computation Earthquake loss data obtained from the China Statistics Yearbook is presented in Table 14.3. From Table 14.2, we can see that the biggest earthquake was in 2008 in Sichuan province. In this earthquake catastrophe, although the loss claims were mainly from the government and public endowment, insurers did claim high losses. For example, according to some government surveys (www. circ.gov.cn), the excluded liability of earthquake risk in the clauses of some life insurances was waived by most insurers. Obviously, for modeling catastrophe loss, log-normal performs better than normal distribution. This is consistent with the majority of published research.13 Next, we analyze the statistical characteristics of the logarithm of catastrophe loss: mean reversion and fat tail. We use the presence of an autoregressive (AR) feature to test mean reversion in the data. In linear processes with Table 14.1 Catastrophe loss models Model

Literature

Compound Poisson model Jump-diffusion model Double exponential jump-diffusion model

Vaugirard (2003);10 Jaimungal and Wang (2005) Cox(2000); Geman and Yor (1997) Zhu(2008);11 Chang and Hung(2009)12

Catastrophe Bond and Risk Modeling

139

Table 14.2 Chinese earthquake loss data, 1966–2008 Year 1966 67,68 1969 1970 71,72,73 1974 1975 1976 77,78 1979 80,81,82 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993, 94 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 06,07 2008

Frequency

Magnitude

Loss (10,000 RMB)

1 0 2 1 0 1 1 2 0 1 0 1 0 2 1 0 1 4 3 0 1 0 1 3 2 5 1 2 1 0 6 3 3 0 1

7.2 NA 6.4 7.7 NA 7.1 7.3 7.4,7.8 NA 6.0 NA 5.9 NA 5.0,7.4 5.4 NA 7.6 6.7,6.6,6.1,5.4 5.1,6.9,6.2 NA 5.0 NA 6.5 7.0,6.9,6.4 6.6,6.6 6.2,6.6,5.3,6.2,5.1 5.6 6.5,5.5 5.2 NA 6.8,6.2,5.9,6.1,5.5,6.1 5.9,5.6,5.0 6.2,5.3,5.7 NA 8.0

210300 0 21460 64380 0 19287 173259 2861420 0 50346 0 55146 0 16555 24330 0 330825 125005 55827 0 9404 0 45692 242480 59198 88924 7524 59747 25583 0 231980 36197 105681 0 26432976

normal shocks, this amounts to checking stationarity. If this is true, this can be a symptom for mean reversion, and a mean reverting process can be adopted as a first assumption for the data. Mean reversion cam be defined as the property of always reverting to a certain constant as time passes. This property is true for an AR(1) process if the absolute value of the autoregression coefficient is less than one, that is, |α| < 1. Since for the AR(1) process |α| < 1 is also a necessary and sufficient condition for stationarity, testing for mean reversion is equivalent to testing for

140

Enterprise Risk Management in Finance

stationarity. Note that in the special case where |α| = 1, the process behaves like a pure random walk with constant drift. There is a growing number of stationarity statistical tests available and ready for use in many econometric packages. The most popular are: the Dickey and Fuller (DF) test; the Augmented DF (ADF) test; the Phillips–Perron (PP) test (PP); the Variance Ratio (VR) test; and the Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test. We choose the ADF test due to its robust features. Computation is shown in Table 14.3, indicating a t-statistic value of −4.123 and probability value of 0.0021. This means the earthquake loss data pass the ADF test, indicating mean reversion. To test for fat tails we first use a graphical tool called QQ-plot, which compares the tails in the data with those from the Gaussian distribution. The QQ-plot gives immediate graphical evidence of the possible presence of fat tails. Further evidence can be obtained by considering the skewness and excess kurtosis values in order to see how the data differ from the Gaussian distribution. Computations indicate skewness and excess kurtosis values of 4401 and 23 respectively, suggesting the earthquake loss data do have fat tails. Parameter estimation Our parameter estimation is based on Markov chain Monte Carlo (MCMC) approaches, which are a class of algorithms for sampling from probability distributions based on constructing a Markov chain that has the desired distribution as its equilibrium distribution. MCMC methods are particularly well suited for financial pricing applications with stochastic process problems, for several reasons. First, state variables solve stochastic differential equations, which are built from Brownian motions, Poisson processes, or other i.i.d. shocks. Therefore, standard tools of Bayesian inference can be directly used here. Second, MCMC is a unified estimation procedure which simultaneously estimates parameters and latent variables. MCMC directly computes the distributions of the latent variables and parameters given the observed data. This is a strong alternative to the usual approach of applying approximate filters or latent variable proxies. Finally, MCMC is based on conditional simulation without any optimization. MCMC provides a strategy for generating samples x0:t, while exploring the state space using a Markov chain mechanism. This mechanism is constructed so that the chain spends more time in the most Table 14.3 Results of ADF testing

Augmented Dickey–Fuller test statistic Test critical values 1% level 5% level 10% level

t-Statistic

Prob.

−4.123695 −3.571310 −2.922449 −2.599224

0.0021

Catastrophe Bond and Risk Modeling

141

important regions. The MCMC approach in this chapter uses random walk MH (Metropolis–Hastings) algorithm which assumes uniform distribution in [−0.5, 0.5]. Parameter vector θ is updated by following θ′ = θ + τε, where ε is an error term, and τ is a harmonic parameter. Variance of error term is rectified by changing τ. For details of approximating parameters, readers can refer to Andrieu and Freitas.14 Codes are written and implemented in Matlab language. Table 14.4 presents the computational result of parameter estimation using all three sets of models. Figures 14.2 and 14.3 depict the fitted curve of a compound Poisson model and different jump-diffusion processes using historical and simulated data Table 14.4 Results of parameter estimation Model

Distribution Parameters estimation

Compound Normal Log-normal Poisson Gamma model Loglogistic Jumpdiffusion model

Normal Log-normal Gamma Loglogistic

λ = 1.1628, μ = 632469, σ = 374451 λ = 1.1628, μ = 10.4256, σ = 1.5701 λ = 1.1628, a = 0.2433, b = 259921 λ = 1.1628, μ = 0.185159, σ = 0.76712 μ = 0.0586, σ = 0.4568, λ = 0.9426, μY = 7.9434, σy = 3.6244 μ = 0.0566, σ = 0.4011, λ = 0.9034, μY = 2.0797, σy = 1.1976 μ = 0.0566, σ = 0.4011, λ = 0.9312, a = 1.1381, b = 10.4235 μ = 0.0566, σ = 0.4011, λ = 0.9034, μY = 0.9524, σy = 2.0759 μ = 0.0968, σ = 2.1346, λ = 1.0136, η1 = 10.4950, η 2 = 6.1976, p = 0.9735, q = 0.0265

Double exponential jump-diffusion process

16 14 12 10 8 6 4 2 0 2 Figure 14.2

3

4

5

6

7

8

9

Frequency histogram of logarithm of earthquake loss with normal fit

142

Enterprise Risk Management in Finance

× 10–6 Loss Normal Lognormal Gamma Log-Logistic

1.4 1.2

Density

1 0.8 0.6 0.4 0.2 0 0.5

Figure 14.3

1

1.5 Data

2

2.5 ×107

Historical vs. simulated data distribution of compound Poisson model

distribution respectively. Figure 14.3 suggests that for compound Poisson models, the loglogistic distribution is the best choice. Figure 14.4 implies that for jump-diffusion models, log-normal distribution is the best choice; however, the double exponential jump-diffusion model fits Chinese earthquake loss data best. These conclusions will be further verified in the next section. To validate our conclusion, we test the goodness of fit by the Kolmogorov– Smirnov (K–S) test, which is based on the absolute value of the maximum difference D max between the cumulative distributions of two data matrices.15 One of the advantages of the K–S test is that it leads to a graphical representation of the data, which enables the user to detect normal distributions. For larger datasets with, for example, a sample size of greater than 40, the Central Limit Theorem suggests that the t-test will produce valid results even in the presence of non-normally distributed data. However, highly non-normal datasets can cause the t-test to produce fallible results, even for large N datasets. In the last example you will see a case where the t-test fails at N = 80. Table 14.5 gives the computation result of the K–S test, where the absolute value of the maximum difference D max and the p value are given. The result of the K–S test supports Figures 14.3 and 14.4. The double exponential jump-diffusion model is the best fit for Chinese earthquake data.

Catastrophe Bond and Risk Modeling

143

×10–7 7 Loss Jump diffusion-Normal Jump diffusion-Lognormal Jump diffusion-Camma Jump diffusion-Log-logistic Double exponential jump diffusion

6

Density

5 4 3 2 1 0

0

0.5

1

1.5

2

Data Figure 14.4

2.5 ×107

Historical vs. simulated data of stochastic process

Table 14.5 Results of K–S test Model

Distribution

Compound Poisson model

Normal Log-normal Gamma Loglogistic

Jump-diffusion model

Normal Log-normal Gamma Loglogistic Double exponential jump-diffusion model

Dmax of K–S test

P value of K–S test

0.3323 0.1817 0.2541 0.1778

0.2832 0.7041 0.5651 0.7415

0.1531 0.0946 0.1845 0.1124 0.0951

0.6718 0.8092 0.7146 0.7485 0.8356

Error analysis This section provides error analysis in order to validate solutions from prior material by running simulations. The goodness of fit can also be tested, following the error analysis. We first simulate 10,000 paths of earthquake loss data using the Monte Carlo simulation. Then we calculate averages of 10,000

144

Enterprise Risk Management in Finance

Table 14.6 Error analysis Model

Distribution

Compound Poisson

Normal Log-normal Gamma Loglogistic

Jump-diffusion

Normal Log-normal Gamma Loglogistic Double exponential jump-diffusion model

msd_Error

avg_Error

max_Error

5.39623 0.64278 1.82371 0.31423

2.4758 0.8751 1.3045 0.6026

3.64896 0.78545 1.54816 1.59713

2.45201 0.52460 2.24742 0.28252 0.13571

1.52306 0.19821 1.0564 0.8991 0.1024

1.62546 1.57233 1.81238 1.61375 1.2963

paths to estimate loss data. Finally, we compute their mean square deviation, mean absolute error and maximal absolute error. The computational results are shown in Table 14.6. As can be seen from Table 14.6, the mean square deviation value and absolute average error value of the double exponential jump-diffusion model are 0.13571 and 0.1024 respectively, which are less than the corresponding values of other models. Second, the maximal absolute error value in the double exponential jump-diffusion model is 1.2963, close to the value in loglogistic jump-diffusion model, which is 0.8991. These values suggest that the double exponential jump-diffusion model is the best fit for Chinese earthquake loss data.

Conclusions In this chapter, we have reviewed the state-of-the-art approaches in modeling catastrophe losses for cat bond modeling and pricing. We tested Chinese earthquake loss data using three models: the compound Poisson model, the jumpdiffusion model and the double exponential jump-diffusion process, where normal, log-normal, gamma and loglogistic distributions were employed for comparison. Markov chain Monte Carlo (MCMC) was used for parameter estimation, and Monte Carlo simulation was employed to generate simulated data for error analysis. Results indicate that for compound Poisson models the loglogistic distribution is the best choice, while for jump-diffusion models the log-normal distribution is the best. Results also suggest that the double exponential jump-diffusion model is the best fit for Chinese earthquake loss data.

15

Bilevel Programming Merger Analysis in Banking

Introduction In modern organizations, component elements members are mutually dependent on a common set of finite resources.1 These organizational resources include funds, personnel, time, effort, and information.2 As a result, organizations have been described as large pools of scarce shared resources, for which component elements (subgroups) compete.3 MEI, a global provider of trade promotion management solutions, surveyed 52 consumer packaged goods (CPG) manufacturers in May 2011, finding that ‘Trade promotions budgets do not grow and IT budgets are still clamped down, yet these organizations somehow need to improve promotion effectiveness. They are no longer concerned with streamlining the deduction reconciliation process, but they do want better visibility into where their scarce dollars are being spent.’4 Due partly to pressure from competition and shareholders, many corporations, including banks and other financial institutions, seek ways to rearrange their organizational structure and to widen their geographical reach and product variety. These changes often aim to improve efficiency through potentially higher economies of scale and wider scope. This chapter aims to examine how such merger performance is gauged in the presence of scarce shared resources, using a banking organization as an example. In this evaluation, we concentrate on the withinfirm competition for the common resources. Mergers involve a series of decision processes at different stages of M&A activity. Jemison and Sitkin (1986)5 identified four impediments to effective decision-making during M&A: activity segmentation, escalating momentum, ambiguity, and misapplication of acquiring company systems in the acquired company. They emphasized the complexity and ambiguity present in the M&A process, and pointed out that activity segmentation helps executives manage that complexity. In this context, in order to gain a competitive advantage, bank merger decision-makers should identify and benchmark their true managerial 145

146 Enterprise Risk Management in Finance

efficiency with respect to network-merging and restructuring processes such as the merging of bank branches.6 For example, when United Overseas Bank had to rationalize its operations from retail to wholesale banking in late April 2005, 66 out of its 67 branches were merged with Banco de Oro. This chapter offers an efficiency evaluation of financial operations viewed as a series of supply chain operations. We take a view of mergers differing from traditional inter-organizational concerns by considering the context of a single hierarchical firm. In our study, a merger refers to a combination of operations of different parallel business units, each with two levels of decision-making, with occasionally conflicting objectives. The leader at the upper level of operations and the follower at the lower level seek to optimize their individual objectives, and make their own set of decisions. The hierarchical process means that the leader sets the value of their decisions first and then the follower reacts, bearing in mind the selection of the leader. The goal of the leader is also to optimize their specific objective while incorporating the reaction of the follower to their course of action. For example, we consider an investment bank which operates in both a primary and secondary capital market. The bank deals with transactions in long-term instruments with maturity longer than one year, such as corporate debentures, government bonds and preference shares. Banks receive payments based on these long-term instruments from the primary capital market and sell them in the secondary capital market, which provides liquidity and marketability. This investment banking operation can then be viewed as a supply chain where the primary and secondary markets are upstream and downstream chain members respectively. Within the same bank, the front office is responsible for the secondary capital market business. The middle office or marketing division is responsible for collecting loans from the primary market. The two divisions must compete for scarce resources, for example, a budget for marketing activity and IT maintenance. From this viewpoint, M&A in the banking industry requires a framework that takes into account both performance in the primary and secondary markets and the competition for common resources among the different subsystems.

A conceptual banking chain with constrained resources A typical serial (supply) chain includes a stream of processes (operational activities) of goods and services that starts with the customer order, goes from raw materials through the supply and production stages, and ends with the distribution of products to the customer.7 We next discuss and analyze banking functions with constrained resources from this serial chain perspective. Banking activities can be roughly divided into the two markets mentioned earlier: the primary market and the secondary market. In Wu and

Bilevel Programming Merger Analysis in Banking

147

Birge (2012),8 the primary-market operations are to initiate mortgage loans which are then delivered to residential and commercial borrowers, all with associated costs and consumed resources. The secondary-market business operations include selling the mortgage loans obtained from the primary market to investors as whole loans, or pooled as mortgage-backed securities. In contrast to Wu and Birge, the conceptual banking chain model here includes other operations, such as IT-intensive or industry-specific products where sub-chains compete for necessary resources.9 Examples of players in both the primary and secondary markets might include Canadian Imperial Bank of Commerce (CIBC) and Air Canada as examples of industry-specific players involved in the selling (CIBC as an initiator) and buying of assetbacked commercial paper. Figure 15.1 displays a conceptual model of the market-level banking process with constrained resources from a supply chain perspective. For any investment bank branch that manages business in both markets, the physical function of the chain is entirely focused on the conversion of a primary market sale (the act of originating a residential or commercial loan) into a secondary market sale (the act of selling and delivering that loan to a capital markets investor). To evaluate the potential gains of a merger involving banking institutions and their divisions, we analyze the banking chain merging problem as a serial (supply) chain including a leader–follower relationship. In the following sections, we build this model and illustrate its applicability. We note that the leader in the model can be at either the upstream or the downstream level, and

Finite resources, e.g., IT Budget

Personnel

Others

Profit

Personnel Primary market business Borrower selection; Underwriting

Loans Information

Secondary market business Securitization; Derivative trading

Loan recovery

Information flow Product flow Sales, risks, availability, stock, customer satisfaction

Figure 15.1

Supply chain model of the banking process with constrained resources

148

Enterprise Risk Management in Finance

the follower can be correspondingly positioned at the downstream or upstream level. This yields two possible game structures: • The bank puts more emphasis on the primary-market business than the secondary-market business. The primary-market business unit seizes constrained resources to reach its goal, and the secondary-market business unit attempts to maximize its profits, defined and solved as the follower. We denote this as an upper-leader (UL) game. • The bank places greater emphasis on the secondary-market business than on the primary-market business. The secondary-market business unit serves as a leader, and the primary-market business unit serves as a follower in the Stackelberg game. We denote this as a lower-r (LL) game. The problem for the leader at the lower level is to find an optimal resource allocation and weighting scheme that maximize the total profit subject to the follower’s optimal strategy. The follower at the upper level attempts to find an optimal weighting scheme to maximize its total profit given the resource allocating scheme determined by the leader.

Mathematical model A bilevel programming problem is a hierarchical optimization problem consisting of two levels when the constraints of an optimization problem are also determined by the other optimization problem. The upper level, which is also termed the leader’s level, is dominant over the lower level which is also seen as the follower’s level. The leader makes the choice first, to optimize his objective function. Observing the leader’s decisions, the follower makes their own decisions which in turn affect the leader’s strategy. A bilevel linear programming problem (BLP) given by Bard (1998)10 is formulated as follows: min F( x, y ) = p1T x + q1T y x

s.t .

A1 x + B1y ≤ b1

min f ( x, y ) = p2T x + q2T y

(I)

y

s.t .

A2 x + B2 y ≤ b2

where x ϵ Rn, y ϵ Rm refer to the decision variables corresponding to the upper and lower level respectively, p1, p2 ϵ Rn, q1, q2 ϵ Rm, b1 ϵ Rc, b2 ϵ Rd, A1 ϵ Rc×n, B1 ϵ Rc×m, A2 ϵ Rd×n, B2 ϵ Rd×m, and T denotes transpose. Existing work such as Hansen et al. (1992)11 and Bard (1998)12 provides methods to transform the linear bilevel programming problem into a single-level programming problem, which is standard mathematical programming and relatively easy to solve.

Bilevel Programming Merger Analysis in Banking

149

Data envelopment analysis (DEA) is a linear programming methodology to measure the efficiency of multiple organizations and indicate the differences between the inefficient ones and the best-practice ones. DEA is a widely used technique to evaluate the performance of various organizations in public and private sectors. In DEA, the organization is also called a decision-making unit (DMU). Generically, a DMU is regarded as the entity responsible for converting inputs into outputs. For example, banks, supermarkets, car makers, bank branches etc. can all be deemed as DMUs. Consider n DMUs that use a vector of p inputs: xi = (xi1, ... ,xip) to produce a vector of q outputs yi = ( yi1, ... ,yiq). The profit efficiency for DMU j can be evaluated based on a linear programming model proposed by Cooper et al. (2000)13 q

max

∑d

p

T r

r =1

s.t .

y% jr − ∑ csT x% js s =1

n

∑λ x

i ir

≤ x% jr

( r = 1,..., p ),

≥ y% js

( s = 1,..., q ),

(II)

i =1 n

∑λ y

i is

i =1

λi ≥ 0

(i = 1,..., n ),

where (x~j1, ... ,x~jp, y~j1, ... ,y~jq) are decision variables and c = (c1, ... ,cp) and d = (d1, ... ,dq) are the unit price vectors attached to the input x~ = (x~j1, ... ,x~jp) and output y~j = (y~j1, ... ,y~jq) vectors respectively, λ = (λ1, ... ,λn) is a nonnegative multiplier used to aggregate existing production activities. Based on an optimal solution (x*j1, ... , x*jp, y*j1, ... , y*jq) of the above model, the profit efficiency of DMU j (PEj) is computed as follows: q

∑d y r

PEj =

p

jr

r =1 q

− ∑ cs xjs s =1 p

(III)

∑ dr y − ∑ cs xjs* * jr

r =1

s =1

where yj = (yj1, ... ,yjq), xj = (xj1, ... ,xjq), are the vectors of observed values for q

DMU j. Under the positive profit assumption, i.e.,

∑d r =1

p

T r

y jr − ∑ csT xjs > 0 , we s =1

have 0 < PEj ≤ 1, and DMU j is profit-efficient if and only if PEj = 1. The bilevel programming and DEA model are combined to create an integrated bilevel programming-DEA model to evaluate the performance of a hierarchical system and its sub-levels under two game situations. The NP-hard bilevel-programming-DEA model is reformulated into at standard linear-bilevel

150 Enterprise Risk Management in Finance

programming form, which can then be easily transformed to a more tractable single-level programming problem.

Merger evaluation Evaluation of a merger with bilevel structure can be accomplished in two stages. First, a firm is evaluated using the average input bundle of existing production. Second, the average firm production is doubled in scale to reach the merged firm production. Then the performance of the two (merged and virtual firm) are compared using the average input bundle.14 The first stage is called the harmony effect and the second is called the scale effect. The harmony effect is useful because, if firms shared the combined input equally and used the identical average bundle, each would produce this higher level of output,15 assuming the relationships within the DEA framework. To analyze the potential gains from a merger of n bilevel systems, the following five steps are proposed to compute the harmony efficiency, scale efficiency, and merger efficiency.16 We use the CRS assumption as a demonstration example. Similarly, the VRS assumption can also be employed to analyze the post-merger returns to scale. Step 1: Solve the bilevel programming DEA problem for each DMU, using integrated bilevel programming-DEA model to obtain optimal solutions. Using nonnegative multipliers, solutions of integrated bilevel programming-DEA give the best-practice frontier through a convex combination of existing production activity. Step 2: For each variable, compute the average slack-adjusted input-output bundle, and the profits of leader and follower are computed as the average input-output bundle. Step 3: Solve the bilevel programming DEA problem with the average inputoutput bundle to generate efficiency values and record the corresponding optimal input-output bundle for the leader, follower and system. Step 4: Compute the total (slack-adjusted) input and output bundles of n systems, and the profits of leader, follower and system using the total inputoutput bundle are computed. Step 5: Solve the bilevel programming DEA problem with the total input-output bundle. With the solution, we can compute merger efficiency, harmony and scale efficiency of the whole bilevel system as well as for the subsystems. If the computed merger efficiency is greater than 1, then n merger members will benefit from the potential profit generated by the merger; otherwise, it would be more efficient to keep these units separate.

Bilevel Programming Merger Analysis in Banking

151

Incentive incompatibility problems occur in merger evaluation from two primary sources: First, there is a possibility of double marginalization, which occurs when both upstream and downstream firms (or firm divisions) have monopoly power and each firm reduces output from the competitive level to the monopoly level, creating two deadweight losses.17 Second, the leader can use their stronger market power to seize more resources than the follower receives. If this is achieved by sacrificing the follower’s benefit, the follower will not participate in the merger. Therefore, there is a need to revise the strategy to yield an incentive-compatible strategy. To motivate an inefficient follower to participate in the merger activity, the leader promises to share at least α% of their profit with the follower. Therefore, the total profit that the follower would receive is at least their actual profit plus α% of the leader’s profit. The total profit of the system remains unchanged under this profit-sharing strategy. This profit-sharing strategy adds an incentive-compatible constraint to the bilevel programming DEA model. A numerical example for incentive incompatibility To illustrate the incentive compatibility issue, we consider a hypothetical example with eight bilevel systems to be merged. Each system consists of the leader and the follower. For the leader, we employ three inputs (two direct inputs X D1 and one shared input X1) and three outputs (two direct outputs Z1 and one intermediate output Y ). For the follower, we utilize four inputs (two direct inputs X D2, one shared input X2 and one intermediate input Y ) and two direct outputs Z2. The raw data is shown in the table of the online supplementary appendix. A premerger analysis suggests that the leader’s gain can be improved by around 25% of the observed profits, while the follower can gain nothing since their relative efficiency scores are 1. This implies that a potential merger activity might not be favored by the followers. We therefore employ the profitsharing strategy to treat this problem. Under the first situation with the CRS assumption, to find a suitable α for the leader in this strategy, we must calculate Table 15.1 Input and output data for the 8 branches in the numerical example Branch DMU1 DMU2 DMU3 DMU4 DMU5 DMU6 DMU7 DMU8

XD1 2.5 7 3 9 2.3 7.4 3.5 8.8

X1

13 4 12 13.4 7 9.8 18 4.6 12.5 5 11.7 14 7.5 10 17.9 5

Z1 35 76 52 63 33 73 57 60

Y 60 53 42 71 62 50 45 70

30 55 40 70 35 53 38 72

X D2 1.5 5.6 4 8.8 1.6 5.8 4 9

12 13 15.4 11.2 12.3 13 15.6 11.5

X2 16 6.6 10.2 15.4 15 6 10 15

Z2 55 87 65 78 52 85 69 75

65 45 56 89 65 42 56 90

PL

PF

105.5 151.6 114.2 172.4 110.2 142.9 119 170.3

60.5 51.8 51.4 61.6 53.1 49.2 57.4 57.5

152 Enterprise Risk Management in Finance

the profit efficiency values for both leaders and followers, and then change the value of the profit-sharing parameter α from 0 to 0.1. A numerical analysis suggests that both leaders and followers are better off when α « (0, αˆ), where αˆ ≈ 0.01.

Case study: banking chain illustration This section conducts a banking chain merger efficiency analysis using our proposed approach. The Canadian banking industry experienced an increasingly dynamic market environment due to a change in the legislative regime of the Canadian government in the early 1990s. Benefiting from new and cost-effective technology, Canadian banks have in many ways increased performance measurements and reduced operating costs. They have maintained or even increased the quality of their services while expanding to a broader customer base in order to be more competitive in the global banking market. For example, GIS-based technologies have been employed by Canadian banks for merger evaluation, particularly for derivation of market boundaries and market share estimation.18 Both negative and positive effects of mergers need to be taken into consideration along with uncertainties and risks stemming from multiple sources. Gauging the potential gains from mergers, and the decomposition of these gains into harmony and scale effects, provide support for decisions made by banks on whether to green-light a merger with its underlying conditions. The hierarchical structure of the banking chain is similar to Figure 15.1. For clarity of exposition, in the bilevel banking chain model we extract two direct inputs (personnel costs, and other expenses); one intermediate output (loans); and two final outputs (profit, and loan recovery). We consider the annual IT budget as a constrained input resource to support computers, required software and systems, data network rentals, and any maintenance and repair. The large Canadian bank we consider has the greatest market share in the Canadian e-banking business, which relies heavily on IT support. Personnel costs include salaries and benefit payments for full-time, part-time, and contract employees. Other expenses are costs other than personnel and IT budget, such as marketing and advertising expenditures, training and education costs, and communication expenses. Loans are composed of credit notes issued to individual customers. Thirty branches of this bank in Ontario are selected to consider mergers of the branches as a form of intra-firm reorganization. The potential for performance improvement in these branches provides the incentive to consider gains from merging independently operated branches with other banking processes. We chose branches from the same bank for two primary reasons. First, we expect high relative efficiency levels because of the similarity of the technologies and locations and the cooperation among the branches. Second, existing work from

Bilevel Programming Merger Analysis in Banking

153

Table 15.2 Raw data for 30 branches XD1

X1

Other Personal expense cost (×105) Branch (×105)

IT budget (×105)

DMU1 DMU2 DMU3 DMU4 DMU5 DMU6 DMU7 DMU8 DMU9 DMU10 DMU11 DMU12 DMU13 DMU14 DMU15 DMU16 DMU17 DMU18 DMU19 DMU20 DMU21 DMU22 DMU23 DMU24 DMU25 DMU26 DMU27 DMU28 DMU29 DMU30

0.133 0.169 0.24 0.159 0.156 0.18485 0.5642 0.12 0.198 0.198 0.137 0.297 0.131 0.125 0.138 0.144 0.076 0.155 0.14 0.126 0.12843 0.059 0.057 0.141 0.146 0.196 0.105 0.121 0.127 0.165

71.3 107.1 122.4 41 36.3 40.9 91.8 123.5 182.1 191.5 302.8 544 87.4 691.8 458 124.1 45 589.2 713.8 97.3 229.4 44.4 50.8 37 39.5 268 78.1 87.2 175.7 193.9

1.5 1.7 2.35 1.1 2.11 1.33 0.6 0.71 1.2 1.2 2 3.8 0.5 3.7 4 1.1 0.53 3.45 3.82 1.28 1.36 0.55 0.57 0.98 1.04 2.06 0.67 1 0.106 1.72

Y

X2

Z2

IT Loan Loan budget Profit recovery (×105) (×105) (×105) (×105) 1447.8 1950.2 2095.2 1364.4 1390.2 1520.6 8118.6 1144.1 1742.5 1742.5 3153.7 4517.7 1434.2 3249.1 2622 1749.3 951.2 4246.9 3915.8 1898.7 1876.5 754.6 759.5 1690.6 1726.4 3643 1158.1 2220.7 2067 2132.5

2.5 2.3 1.65 2.9 1.89 2.67 3.4 3.29 2.8 2.8 2 0.2 3.5 0.3 0 2.9 3.47 0.55 0.18 2.72 2.64 3.45 3.43 3.02 2.96 1.94 3.33 3 3.894 2.28

523.2 534 536.3 495.4 521.1 523.7 610.3 519.9 527.4 527.4 442.9 386 517.7 564.8 402 524.3 506.7 600.2 372.5 524.3 487.1 515.3 512.3 523.3 526.3 560.1 512 524.8 525.3 493.5

1427.7 1923.3 2066 1324.8 1365.2 1496.3 8005.2 1126.9 1712.9 1712.9 2980.6 4300.9 1412.7 3070.4 2283.8 1728.3 932.18 4026.1 3559.5 1870.4 1805.2 744.79 749.78 1658.5 1697.1 3577.4 1143 2158.5 2042.2 2030.1

PL (×105)

PF (×105)

1374.9 1841.2 1970.2 1322.1 1351.6 1478.2 8025.6 1019.8 1559 1549.6 2848.8 3969.6 1346.2 2553.5 2159.9 1624 905.59 3654.1 3198 1800 1645.6 709.59 708.07 1652.5 1685.7 3372.7 1079.2 2132.4 1891.1 1936.7

500.6 504.8 505.4 452.9 494.2 496.7 493.5 499.4 495 495 267.8 169 492.7 385.8 63.84 500.4 484.2 378.8 15.98 493.3 413.2 502 499.1 488.2 494 492.6 493.6 459.6 496.6 388.9

the theory of transfer effects suggests that merger performance is higher when business units operate in similar industries, due to a positive transfer effect; the merger performance could suffer if it was made in a different industry, where a negative transfer effect may occur.19 In general, similar organizational cultures may enhance the chances of gains from operational mergers. Considering this situation, we expect potential gains from mergers, in particular from the harmony effect due to new market conditions, and environmental regulations may create new potential demand for the extension

154

Enterprise Risk Management in Finance

of services. The qualifications and structures of the extension offices may adapt more slowly than more established offices due to union restrictions, recruitment difficulties, etc. We set the output price or weight vectors to unity, which means that all output variables are equally weighted in the computation. Program codes in the Matlab language were generated, and the computation was performed on a PC Pentium at 1.79GHz with 1.99GB of RAM under Windows XP. Codes were written in Matlab 7.1 Release 14, and required 11 minutes of computation time to solve the models. To conduct a pre-merger analysis, we investigate the profit efficiency of the system, the leader, and the follower for the 30 existing branches and to assess potential gains. The models are implemented under both the CRS and VRS assumptions. Table 15.3 gives the mean profit efficiency, the standard deviation of the profit efficiencies, the number of fully efficient bank branches, and the lowest and highest profit efficiency among all bank branches. From the summary of different estimations, the level of profit inefficiency (1-PE) of the bank from the Great Toronto area is 20–30% in most specifications. The interpretation of this result is that if everyone learned best practices, total profits could be improved by 20–30% without changing the organization of the bank. We can make three observations from the efficiency values in Table 15.3. First, the DEA efficiency increases from the CRS assumption to VRS assumption, which is consistent with existing DEA literature20 (and naturally follows from the reduced constraint set). Second, there exist substantial inefficiencies for most bank branches both in the full chain and its members. Moreover, if we assume a CRS technology, we see no branch that is profit efficient as a system though some branches are profit-efficient in their leader operations (e.g., Branch 7, 11 and 28) and others are profit-efficient in their follower operations (e.g., Branch 22 and 23). If we assume a VRS technology instead, three bank branches, i.e., Branch 2, 22 and 23, are profit-efficient in the system, the leader, and the follower. This validates Proposition 1, namely, that the system is profit-efficient only if both the leader and the follower are profit-efficient. Third, profit efficiency scores of the system are close to those of the leader but

Table 15.3 Profit efficiency values Statistics

PEL-CRS

PEF-CRS

PES -CRS

PEL-VRS

PEF-VRS

PES -VRS

Mean Standard deviation Minimum Maximum #{E=1}

0.727833 0.177358 0.432 1 3

0.565767 0.202374 0.029 1 2

0.7017 0.151079 0.467 0.963 0

0.798533 0.189591 0.456 1 8

0.861167 0.256551 0.036 1 3

0.809433 0.162195 0.524 1 3

Bilevel Programming Merger Analysis in Banking

155

quite different from those of the follower. This seems to suggest that the system profit efficiency is mainly affected by the leader. To examine how resource allocation impacts the firm’s performance, it is necessary to align resource allocation to a company’s bottom-line business goals. For this assessment, banking operations consist of two main valueadded activities in two market. Under the UL game structure, the upstreamlevel sub-division has stronger market power and decides on the amount of resources it needs to maximize efficiency. Hence, we identify stage 1 as a resource allocation-related value-added activity under the UL game scenario. In contrast, stage 2 is related to resource allocation and value-added activity under the LL game scenario. This assumption, although possibly more restrictive than in reality, may be reasonable and necessary because of the lack of information about how management can allocate and spend annual IT budgets. Under the UL game scenario and the VRS assumption of the 30 branches, three achieved overall efficiency, eight were rated efficient in stage 1 (resource allocation-related activity), and three were rated efficient in stage 2. The results show five branches ass rated efficient in the resource allocation-related activity (stage 1) without achieving overall efficiency. The following classification of branches aims to provide a means for management to better evaluate their resource allocation-related operations.21 • Type 1 represents branches efficient in the resource allocation-related valueadded activity, but which cannot achieve overall efficiency. Branches 7, 18, 19, 24, and 28 belong to this classification. • Type 2 represents branches which achieved overall efficiency in spite of inefficiency in the resource allocation-related activity. None of the branches belongs to this classification. • Type 3 represents branches which are efficient both in resource allocation and overall. Branches 5, 22, and 23 belong to this classification. • Type 4 represents branches which are inefficient both in the resource allocation and overall. Branches 1–4, 6, 8–17, 20, 21, 25–27, 29, and 30 belong to this classification. Different types of branches can now be studied in more detail to identify the reasons for performance differences. To explain the systematic differences between different types of branches, management may refer to factors such as the type of system used, management practices, and implementation procedures. Similar to Wang et al. (1997)22 an interesting observation from this classification is that inefficiency in resource allocation always leads to overall inefficiency. Moreover, an overall inefficiency seems to imply efficiency in resource allocation-related activity.

156

Enterprise Risk Management in Finance

We now compare the proposed approach with existing efficiency evaluation tools: normal DEA and stochastic frontier analysis (SFA). In the SFA framework, we estimate linear and log-linear specifications of the mean structure and truncated normal distribution for the inefficiency error term. In the normal DEA framework, we compute efficiency values for the scale assumptions CRS (CCR) and VRS (BCC). These existing tools simply consider a firm as an input-output black box. That is, they are not concerned with the inner intermediate inputs/ outputs or any decision hierarchy or game structure. For each estimation method, Table 15.4 shows the mean Farrell efficiency, the standard deviation of the Farrell efficiencies, the number of fully efficient bank branches, and the lowest and highest Farrell efficiency among all bank branches. From Table 15.4, the level of profit inefficiency (1-PE) in the bank from the Great Toronto area is 0–10% in most specifications if we ignore inner activities (e.g., loan collection) among the leader and follower business units, which results in significant overestimates of bank performance. We further compute Pearson correlation coefficients and significant values in Table 15.5. The correlation analysis shows that PE-CRS and PE-VRS generate highly correlated results, with a Pearson correlation value of 0.842, but do not seem to produce a correlated result with existing DEA and SFA. We also note that the Spearman correlation between the individual efficiencies calculated in the SFA linear model and the CCR model are 0.127, while it is −0.031 in the SFA log-linear model. The Spearman correlation between BCC DEA and SFA cannot be computed, because the BCC efficiency score is unity, suggesting no discriminating power. In general, then, these models do not suggest agreement in the individual evaluations. This implies that existing models analyzed here cannot directly be used as authoritative efficiency analysis models for bank branches with bilevel structures. Traditional DEA or SFA provide a measure of overall performance, but the overall efficiency measure derived cannot provide insight into the efficiencies of sub-systems (either the leader or the follower) operations and their importance in realizing final outputs of both pre-merger and post-merger entities. In the analysis of specific merger cases,

Table 15.4 A comparison with existing DEA and SFA SFA SFA Linear loglinear

Statistics

PE-CRS

PE-VRS

CCR

BCC

Mean Standard deviation Minimum Maximum

0.7017 0.151079

0.809433 0.162195

0.9829 0.030204

1 0

0.99167 0.000691

0.915705 0.087243

0.467 0.963

0.524 1

0.889 1

1 1

0.989663 0.993046

0.658106 0.988997

Bilevel Programming Merger Analysis in Banking

157

Table 15.5 Correlations

PE-CRS PE-VRS CCR

SFA SFA BCC LINEAR LOGLINEAR

PE-CRS

Correlation 1.000 Sig. (2-tailed) . PE-VRS Correlation .842** 1.000 Sig. (2-tailed) .000 . CCR Correlation .483** .544** 1.000 Sig. (2-tailed) .007 .002 . BCC Correlation . . . NA Sig. (2-tailed) . . . SFA LINEAR Correlation .099 .222 .127 Sig. (2-tailed) .602 .238 .502 SFA OGLINEAR Correlation .011 .113 −.031 Sig. (2-tailed) .953 .553 .873

. . 1.000 . .917** .000

1.000 .

Notes: ** Pearson Correlation is significant at the 0.01 level (two-tailed); NA cannot be computed because at least one of the variables is constant.

it is then important to open the black box and develop good underlying production models of the technology accommodating inner intermediate inputs/ outputs and game structure. Post merger To respect the fact that most branches favor merging two adjacent branches, we examine potential profit gains by merging two branches at a time. This leads to a total of 435 possible mergers involving two branches. Therefore, the relative profit efficiency of these 435 possible mergers is computed with reference to the original DMU by our bilevel programming-DEA model. We tested the merger gains from all of these combinations using both the CRS and VRS bilevel DEA chain merger models. We examined two kinds of merger activities: a merger of individual subchain (either leader or follower) members and a chain merger. Table 15.6 gives the computational statistics under both the CRS and VRS assumptions: the number of the efficient and coordinated (using profit-sharing strategy) mergers, and the average merger efficiency scores Em. Tables 15.7 and 15.8 present merger efficiencies of the top ten most promising mergers under both CRS and VRS assumptions respectively. If we believe in the estimated CRS technology, under the 1st situation with UL game structure, we see that at an overall scale the average potential profit gains in the (435–12) bilevel system mergers is 2.2% (=1.022–1) while this number decreases to 0.7% (=1.007–1) under the second situation, with LL game structure. Indeed, in the detailed results, 100 out of the 225 effective system

158 Enterprise Risk Management in Finance

Table 15.6 Statistics under both CRS and VRS assumptions CRS

VRS

Measure (>100%)

Number

_ Em

The UL game Efficient mergers for the leader Efficient mergers for the follower Efficient mergers for the whole system Coordinated efficient mergers

218 191 225

1.023 1.02 1.022

163

1.02

The LL game

299

1

249

1.035

113

1.263

335

1.007

112

1.194

193

1.006

3

1.2

Game

Efficient mergers for the new leader Efficient mergers for the new follower Efficient mergers for the whole system Coordinated efficient mergers

Number

_ Em

213 13 213

1.779 1 1.61

7 70

1.45 1

Table 15.7 Top ten promising mergers under UL game structure RTS

CRS

VRS

Merger

ELm

HL

SL

EFm

HF

SF

ESm

HS

SS

22,23 2,22 2,23 22,29 16,22 16,23 4,22 17,23 17,22 13,23 5,24 5,27 20,22 1,5 22,30 24,27 17,30 1,17 23,29 16,22

1.327 1.188 1.186 1.173 1.155 1.147 1.137 1.132 1.129 1.106 4.111 4.084 3.867 3.715 3.48 3.398 3.311 3.213 3.168 3.136

1.327 1.188 1.186 1.173 1.155 1.147 1.137 1.132 1.129 1.106 1 1.237 1.295 1.212 1.261 1.11 1.116 1.002 1.137 1.088

1 1 1 1 1 1 1 1 1 1 4.111 3.3 2.985 3.064 2.76 3.063 2.966 3.208 2.786 2.882

1.138 1.051 1.056 1.061 1.005 1.055 1.001 1.076 1.076 1.019 0.996 0.997 0.997 0.997 0.998 0.996 0.998 0.999 0.997 0.997

1.138 1.051 1.056 1.061 1.005 1.055 1.001 1.076 1.076 1.019 1 1 1 1 1 1 1 1 1.001 1

1 1 1 1 1 1 1 1 1 1 0.996 0.997 0.997 0.996 0.998 0.996 0.998 0.999 0.997 0.997

1.253 1.146 1.147 1.138 1.109 1.119 1.095 1.112 1.111 1.081 3.334 3.335 3.087 3.1 2.947 2.851 2.836 2.655 2.639 2.629

1.253 1.146 1.147 1.138 1.109 1.119 1.095 1.112 1.111 1.081 1 1.18 1.215 1.164 1.205 1.085 1.092 1.001 1.104 1.068

1 1 1 1 1 1 1 1 1 1 3.333 2.827 2.54 2.662 2.446 2.628 2.596 2.652 2.39 2.463

pairs have an improvement potential of more than 5% under the first situation, with UL game structure, and 128 out of the 335 effective system pairs have an improvement potential of more than 2% under the second situation, with LL game structure.

Bilevel Programming Merger Analysis in Banking

159

Table 15.8 Top ten promising mergers under LL game structure RTS

Merger

ELm

HL

CRS

15,29 21,29 15,23 14,23 15,17 28,29 15,24 21,25 13,21 12,16

1.047 1.027 1.002 1.001 1.001 1.001 1.001 1.001 1.001 1.001

1.047 1.027 1.002 1.001 1.001 1.001 1.001 1.001 1.001 1.001

VRS

21,27 4,21 28,30 3,29 21,28 13,21 7,21 20,21 15,19 4,30

1.008 1.008 1.007 1.007 1.007 1.007 1.006 1.006 1.005 1.005

1.708 1.977 1.819 1.999 1.755 1.802 2.481 1.773 1.603 2.033

SL

EFm

HF

1 1 1 1 1 1 1 1 1 1

1.033 1.019 1.002 1.031 1.002 1.001 1.002 1.001 1.001 1.009

1.033 1.019 1.002 1.031 1.002 1.001 1.002 1.001 1.001 1.009

0.59 0.51 0.554 0.504 0.574 0.559 0.406 0.567 0.627 0.494

0.961 0.818 0.858 1.004 0.887 0.682 0.384 0.804 0.78 0.627

0.872 0.928 0.944 1.177 0.938 0.818 0.973 0.795 0.916 0.934

SF

ESm

HS

SS

1 1 1 1 1 1 1 1 1 1

1.001 1.001 1.004 1.111 1.005 1.003 1.006 1.005 1.004 1.036

1.001 1.001 1.004 1.111 1.005 1.003 1.006 1.005 1.004 1.036

1.101 0.881 0.909 0.853 0.946 0.833 0.395 1.01 0.851 0.671

0.969 0.847 0.883 1.005 0.908 0.733 0.435 0.836 0.811 0.685

1.024 0.946 1.092 0.776 1.089 0.81 1.326 0.758 1.077 0.843 0.974 0.752 1.097 0.397 0.95 0.88 1.01 0.803 1.103 0.621

1 1 1 1 1 1 1 1 1 1

If we assume a VRS technology instead, the corresponding results are given in the last two columns of Table 15.6. In the VRS calculations, 12 mergers under the first situation, with UL game structure, have improvement potential. There were no mergers under the second situation, with LL game structure. These merged branches are outside the technology determined by the 30 bank branches. The explanation is that when two branches are merged, they become very large compared to the existing branches (with a similar mix of resources and services) and consequently are above the estimated optimal scale size for this mix. Therefore, the existing best-practice production does not seem to suggest that the resulting production plans are feasible. As shown in Table 15.6, considerable potential gains are observed from merging the bilevel system. If a CRS technology is assumed, we see that 51.7% (=225/435) of all possible chain merger scores are larger than one under the first situation, with UL game structure, and 77% (=335/435) of merger scores are larger than one under the second situation, with LL game structure. Under the VRS assumption, Table 15.6 indicates that only 49% and 25.7% of merger scores in two game structures are larger than unity, and the gains from merging (as opposed to the gains from individual improvements) are considerably less.

160 Enterprise Risk Management in Finance

To examine how a profit-sharing strategy solves the incentive incompatibility problem, in Table 15.6 we show the number of coordinated mergers using profit-sharing strategy and their respective mean profit efficiency scores under both CRS and VRS assumptions. From Table 15.6, it can be seen that the number of the coordinated efficient mergers under the CRS assumption is greater than that under the VRS assumption in both of the game situations. In particular, 44.3% of the total 435 mergers, which is 193 mergers, are coordinated efficient under CRS in the second situation. In the first situation, with UL game structure, 218 mergers are found to be efficient from the leader’s perspective, and 74.8% of them, 163 mergers, are coordinated efficient. The mean efficiency scores of the leader in general decreases, but the mean efficiency scores of the follower improve. To further examine the most promising mergers, Tables 15.7 and 15.8 provide the top ten most promising mergers under both the CRS and VRS technologies for two game structures where we report merger efficiency, the harmony effect, and the scale effect for both leaders and followers. We make three observations from these results. First, both tables illustrate that there are potential gains from mergers. Second, with CRS, the scale effects SL , SF and SS are unity, as there is no gain in resizing with constant returns to scale. Under the VRS technology, both Tables 15.7 and 15.8 indicate that the harmony effect favors mergers while the scale effects of the followers appear to work against mergers. Third, if a CRS technology is assumed, all ten mergers are coordinated efficient under two game structures, since both leaders and followers yield potential gains from mergers. However, if a VRS technology is assumed, all ten mergers are non-coordinated since followers yield efficiency scores that are smaller than unity. From Table 15.6, we see that under the UL game structure and VRS technology, seven mergers are found to be coordinated efficient. In the second situation, with LL game structure, where the leader and the follower are interchanged, under the CRS assumption 299 mergers are found to be efficient from the new leader’s perspective, and 193 mergers are found coordinated efficient. Under the VRS assumption, 70 mergers are found efficient from the leader perspective, but only three of these mergers are found to be coordinated efficient.

Managerial insights A multi-methodological approach is developed to evaluate the potential gains of a banking operations merger. This multi-methodological approach incorporates both analytical model development, such as bilevel programming and DEA, and a case study based on real banking operations data. The theory and the case study results for an intra-firm analysis of merger possibilities are

Bilevel Programming Merger Analysis in Banking

161

meant to provide management with a tool to identify potential improvement areas resulting from combining units within a common supply chain subject to constrained resources. The quantitative theory development and the case study have a two-way relationship: the case study is used for understanding the metrics and their relationships in quantitative analysis. On the other hand, quantitative theory development is used to understand the phenomena observed in the case study. The case study examines a phenomenon in the natural setting of banking operations merger, employing multiple methods of data collection from several sources. Based on this theory, management can find possible rewarding new alignments with incentive-compatible merger activities. The demonstrative case study shows how the models can provide management with information about potentially promising merger cases that respect constrained resources and sub-unit incentives. Existing work 23 suggests that various internal or external factors influence a firm’s decision to become an ‘acquirer’ or a ‘target.’ Potential internal factors involve economic factors (e.g., the financial profile of the firm) or noneconomic factors (e.g., managerial motives). Potential external factors include macro or industrial conditions such as growth, capacity utilization, market share, regulations, antitrust policies, tax structure etc. The case study illuminates internal factors of interest to management. These factors may dominate external factors in this analysis for two reasons: first, the DEA approach by default assumes that all entities under evaluation are homogeneous, which implies that the external factors are most likely to affect all the merger entities in a similar manner. Second, the external factors may be a function of internal factors which are included in the two-stage system of the study. Based on the computations here, we can recommend potential areas to combine to save IT, facility, and other costs, and to ensure that the firm’s internal supply chain activities operate more efficiently.

Conclusions Evaluation of potential gains from merging firms with chain operations subject to constrained resources is a multi-criteria decision problem for many organizations. In some situations, the evaluation of the performance of chain operations involves factors which simultaneously play the role of inputs and outputs. A concurrent consideration of multiple criteria complicates the performance evaluation of such merger decisions. Competing business divisions, indeed, have different levels of achievement under multiple criteria. We have studied banking operations with a leader–follower game structure from a series chain perspective, where the primary and secondary markets are upstream and downstream chain members respectively. We described bilevel programming series-chain DEA models to evaluate the potential merger. We

162

Enterprise Risk Management in Finance

also defined merger efficiency concepts for both single units and multiple units under this structure, and developed an approach to solve the NP-hard bilevel programming-DEA model. In addition, we discussed the decomposition of merger efficiency into a harmony effect and a scale effect at both the chain and sub-chain levels. In this framework, we have shown that the supply chain with constrained resources and a leader–follower relationship is efficient if and only if both leader and follower are efficient. We proposed a profit-sharing strategy to address incentive incompatibility problems that might be present in merging firms with such a leader–follower structure. Both leaders and followers benefit under the proposed incentive-compatible strategy. A case study of potential intra-firm banking-chain mergers with limited input resources was also presented to illustrate the proposed approach. Using 435 potential mergers involving branches merging in pairs, the results show significant potential gains from these mergers in banking chains with a leader– follower structure and constrained recourses. The case study demonstrates that bank branches achieve potential gains by conducting intra-firm mergers, which can be incentive-compatible. The findings of the case study also provide insights into the consequences of different pairings of firm entities and the results of different types of M&A deals. This allows a deeper understanding of mergers in the financial sector and its implications on the acquiring banking entities with chain operations.

16

Sustainability and Risk in Globalization

We all are aware that we live in a world of change, and that there are many things going on that appear to be problematic. There are islands in the Pacific that are going underwater (well, they aren’t sinking, but the water level is rising). There are glaciers that are disappearing. Europeans spent a good part of the second millennium AD looking for a Northwest Passage to Asia – but nature appears to be providing one for us in the third. Climate appears to be changing, although weather being what it is, long-range trends are elusive at best. But North Dakota might someday be a winter haven, and Omaha might be a seaport. We need to worry about the environment. We will, of course, argue about how to cope. Some want to ban coal and petroleum use NOW. Others prefer to consider moving uphill. The utilization of high technology for loss-prevention and control systems in natural disasters or fires, accidents and quantitative models in derivatives for insurances and finance has expanded substantially in the past decade. Encouraged by traumatic recent events such as 9/11/2001 and business scandals including Enron and WorldCom,1 risk management has not only developed a control focus, but most importantly it remains a tool to enhance the value of systems, both commercial and communal. Integrated approaches to manage risks facing organizations have been developed along with new business philosophies of enterprise risk management.

Enterprise sustainability Enterprise sustainability is a term addressing the need to consider an organization’s practices in light of economic, social, and environmental impacts.2 Businesses need to maintain their customer base and their brand reputation. This is related to corporate social responsibility and to the need for businesses to contribute toward the solution of social problems, as well as to risk

163

164 Enterprise Risk Management in Finance

management’s aim to identify present and future threats and to devise plans to mitigate or eliminate these risks. All human endeavors involve uncertainty and risk. Mitroff and Alpaslan (2003)3 categorized emergencies and crises into three categories: natural disasters, malicious activities, and systemic failures of human systems. Nature does many things to us, disrupting our best-laid plans and undoing much of what humans have constructed. Events such as earthquakes, floods, fires and hurricanes are manifestations of the majesty of nature. Recent events, including the tsunami in the Indian Ocean in 2004 and Hurricane Katrina in New Orleans in 2005 demonstrate how powerless humans can be in the face of nature’s wrath. Global economic crisis risks are profound and widespread over the last decade. Businesses in fact exist to cope with risk in their area of specialization. But chief executive officers are responsible for dealing with any risk fate throws at their organization. Malicious acts are intentional on the part of fellow humans who are either excessively competitive or who suffer from character flaws. Examples include the Tylenol poisonings of 1982, syringes being placed in Pepsi cans in 1993, the bombing of the World Trade Center in 1993, Sarin gas attacks in Tokyo in 1995, terrorist destruction of the World Trade Center in New York in 2001, and corporate scandals within Enron, Andersen, and WorldCom in 2001. More recent malicious acts include terrorist activities in Spain and London, and in the financial realm, the Ponzi scheme of Bernard Madoff uncovered in 2009. Wars fall within this category, although our perceptions of what is sanctioned or malicious are colored by our biases. Criminal activities such as product tampering or the blend of kidnapping and murder are clearly not condoned. Acts of terrorism are less easily classified, as what is terrorism to some of us is the expression of political behavior to others. Similar gray categories exist in the business world. Marketing is highly competitive, and positive spinning of your product often tips over to malicious slander of competitor products. Malicious activity has even arisen within the area of information technology, in the form of identity theft or tampering with company records. Probably the most common source of crises is unexpected consequences arising from overly complex systems.4 Examples of such crises include Three Mile Island in Pennsylvania in 1979 and Chernobyl in 1986 within the nuclear power field, the chemical disaster in Bhopal, India, in 1984, the Exxon Valdez oil spill in 1989, the Ford-Firestone tire crisis in 2000, and the Columbia space shuttle explosion in 2003. The financial world is not immune to systemic failure, as demonstrated by the Barings Bank collapse in 1995, the failure of Long-Term Capital Management in 1998, and the subprime mortgage bubble implosion leading to near-failure in 2008. The use of electric cars, which are

Sustainability and Risk in Globalization

165

viewed as a solution to global warming, have systemic problems related to the shortage of batteries with sufficient power and duration at reasonable cost, as well as the paucity of recharging stations that would make their use practical. All organizations need to prepare themselves to cope with crises from whatever source. Managers are expected to anticipate everything bad that could happen to their organization, and ideal risk management would have a plan for each possible contingency. It is, of course, a good idea to be prepared. However, crises by definition are almost always the result of things not going according to plan, whether the result of nature, malicious humans, or other systemic features catching us unprepared. We cannot expect to cope with every contingency, but it is prudent to develop sufficient spare resources to deal with the unexpected.

Types of risk In general, Risk is defined as the unknown change in the future value of a system. Risks can be viewed as threats, but businesses exist to cope with specific risks. Thus, if they encounter a risk that they are specialists in dealing with, the encounter is viewed as an opportunity. Risks have been categorized into five groups:5 Opportunities – events presenting a favorable combination of circumstances giving rise to the chance of beneficial activity; Killer risks – events presenting an unfavorable combination of circumstances leading to hazard or major loss or damage resulting in permanent cessation of operations; Other perils – events presenting an unfavorable combination of circumstances leading to hazard of loss or damage leading to disruption of operations with possible financial loss; Cross-functional risks – common risks leading to potential loss of reputation; Business process unique risks – risks occurring within a specific operation or process, such as withdrawal of a particular product for quality reasons. Opportunities should be capitalized upon in most circumstances. Not taking advantage of opportunities leads to the growth of competitors, and thus increased risk; but if opportunities are pursued, enterprise strategy can be modified to manage the particular risks involved. Killer risks are threats to enterprise survival, and call for continuous risk treatment, monitoring, and reporting. The other perils require analysis to assess ownership, treatment, residual risk, measurement, and reporting.

166

Enterprise Risk Management in Finance

Contexts of sustainable risk Risks arise from everything that humans attempt. Life is worthwhile because of its challenges. Doing business has no profit without risk, rewarding those who best understand systems and take what turns out to be the best way to manage these risks. We have addressed risk management as applied to production in the food we eat, the energy we use to live, and the manifestation of the global economy via supply chains. What we eat One of the major issues facing human culture is the need for quality food. Two factors that need to be considered are: human population growth, and threats to the environment. We have understood since Malthus that the human population cannot continue to grow exponentially. Some countries, such as China, have been proactive in seeking to control their population growth. Other areas, such as Europe, are documenting decreases in population growth, probably due to societal consensus. But other areas, which include India and Africa, continue to experience rapid increases in population. This may change as these areas become more affluent (see China and Europe). But there is no universally acceptable way to control human population growth. Thus, we expect to see a continued increase in demand for food. Agricultural science has been relatively effective in developing better strains of crops through a number of methods, including bioengineering and genetic science; this led to what was expected to be a ‘green revolution’ a generation ago. As with all of mankind’s schemes, the best-laid plans of humans involve many complexities and unexpected consequences. North America has developed the means to increase production of food that is free from many of the problems that existed a century ago. However, people in Europe, Australia, Asia, and Africa, are concerned about new threats arising from genetically modified agricultural crops. Many US citizens are concerned about the risks of genetically modified food. This is another example of human efforts to reach improvements that lead to new dangers, or unintended consequences, with great disagreement about what is seen to be the truth. A third factor complicating the food issue is distribution. North America and Ukraine have long been fertile producing centers, generating surpluses of food. This connects to supply chain issues. But the fundamental issue is the interconnected global human system with surpluses in some locations and food scarcities in others. Technically, this is a supply chain issue. But more importantly it is an economic issue of sharing food, which is a series of political issues. The heavy reliance of contemporary businesses on international collaborative supply chains leads to many risks arising from shipping, as well as to other factors such as political stability, physical security from natural disaster, piracy

Sustainability and Risk in Globalization

167

on the high seas, and changing regulations. Sustainable supply chain management has become an area of pressure potentially applied by governing agencies, customers, and the various corporate entities involved within a supply chain. These interests can include industry cartels such as OPEC, regulatory environments such as the Eurozone, industry lobbies as in the sugar industry, and so forth that complicate international business on the scale induced by global supply chains. Water is one of the most abundant assets on the Earth, probably next to oxygen, which chemists know is a related element. Rainwater used to be considered pure. The industrial revolution caused the unintended production of acid precipitation with numerous unanticipated consequences, locally, regionally and globally. Water used to be free in many places; only 30 years ago, in those places paying for water would have been considered the height of idiocy. Managing water is recognized as a major issue in that less than 3% of the world’s water is fresh. Lambooy (2011)6 called for attention to waste water management, management of freshwater consumption, and groundwater control management. Water management is increasingly an economic issue, leading to the political arena. Wherever there is a scarcity of water, this induces political efforts to gain a greater share for each political entity, involving allocations not only of drinking water for cities, but also for irrigation of agriculture lands and even for sustainable levels to enable river navigation. The energy we use The generation of energy in various forms is a major issue leading to political debate concerning tradeoffs between those seeking to expand the extraction of fossil fuels to meet the expanding demands vs. those who seek to help reduce the climate change that causes release of fossil carbon dioxide by working to make the transition to renewable energy and a strong emphasis on reduction of inefficiencies in the entire system. Oil continues to be a major source of energy, but its extraction, refining, transportation, and usage not only causes numerous environmental risks but also causes related catastrophic risks such as oil spills and market risks such as highly fluctuating prices. Increasingly, many nations, and regions within nations, are expanding their efforts to transition to post-fossil-carbon societies by changing to renewable energy-based systems. For example, Ng and Goldsmith (2010)7 developed a conceptual and dynamic programming model to explain the entry behaviors of different types of bioenergy businesses, and to demonstrate that bio energy entry decisions emphasize a basic trade-off involving gains from a commitment to specialized, and correspondingly higher cost assets, and gains from remaining flexible with lower levels of fixed and less specific assets. Meyler et al. (2007)8 developed insights into complex landscapes of risk in which the

168 Enterprise Risk Management in Finance

natural environment and well-being of residents was largely ignored as fuel prices and energy security were debated. The impact of oil exploration in the Mexican rain forest was reported by Santiago (2011):9 urbanization and civilization were highly localized, collapsing for reasons not well understood in Northern Veracruz. Cost risks in alternative energy resources have been studied,10 and financial management using the Pinch concept have been proven to be useful in development of tools at preliminary design stage for rough target setting, alternatives evaluation and decision-making. Risks involved in the impact of gasoline blending11 and of ethanol production and usage.12 Technology choices in process residue handling and in fuel combustion are key, whilst site-specific environmental management tools should address the broader biodiversity issues. Mining is a field which traditionally faces high production risks, such as uncertain supply yield or marginal cost variability. Akcil (2006)13 reported on practices in gold and silver mining in Turkey, identifying the importance of cyanide management in that environment.

Globalization Living and working in today’s environment involves many risks. The processes used to make decisions in regard to these risks must consider both the need to keep people gainfully and safely employed through increased economic activity and to protect the earth from threats arising from human activities. We need to consider that there are many risks, and we have challenges in developing strategies, controls, and regulations designed to reduce the risks while seeking to achieve other goals. Globalization has played a major role in expanding the opportunities for many manufacturers, retailers, and other business organizations to be more efficient. The tradeoff has always been the cost of transportation, as well as the added risk of globalizing. In 2010 the Eyjafjallajökull volcano in Iceland shut down air transportation across most of Europe. Some Europeans spent a full week waiting for some means to travel across Europe. Supply chains were also disrupted, as transportation (logistics) is key to linking production facilities in supply chains; many in Europe found their supermarkets short of fresh fruit and flowers. Supply chains often depend on lean manufacturing, requiring just-in-time delivery of components. These systems are optimized, which means elimination of the slack that would cover contingencies such as volcanic disruption of air transport. Multiple sourcing, flexible manufacturing strategies, and logistics networks capable of alternative routing are clearly needed.

Sustainability and Risk in Globalization

169

On March 11, 2011, an earthquake off the Pacific coast of Japan led to a catastrophic tsunami that destroyed most of a rich area of advanced technology manufacturing. It also severely damaged a nuclear power plant, which at the time of writing still saw damage control efforts. While the worst impact was in terms of Japanese lives, there also was major impact on many of the world’s supply chains. Organizations such as Samsung, Ford Motor Company, and Boeing found production disrupted due to lack of key components from Japan. Japanese plants produced about 20% of the semiconductors used worldwide, and double that for electronic components; for example, Toshiba produced one-quarter of the nano flash chips used, and on March 14, 2011, it had to halt operations due to power outages. Modern supply chains need to develop ways to work around any kind of disruption. Wars of course lead to major disruption in supply chains, but tariff regulations can have an impact as well. In 2002, Honda Motors had to spend $3,000 per ton in airlifting carbon sheet steel to the US after tariff-related supply disruptions. In January 2011, the Volkswagen, Porsche and BMW supply chains in Germany were taxed by surging demand; Volkswagen actually had to halt production due to engine and other part shortages. This was not due to natural disaster or war, or any other negative factor, but rather to booming demand in China and the United States. Lean manufacturing and modern consumer retailing operations require maintenance of supply. Supply chains can offer great value to us as consumers: competition has led to better products at lower costs enabled by shipping (by land and air as well as sea) through supply chains; outsourcing allows producers to access the best materials and process them at the lowest cost. Lean manufacturing enables cost efficiency as well. But both of these valuable trends lead to greater supply chain exposure. There is a need to gain flexibility, which can be obtained in a number of ways: • Use of diversified sources to enable use of alternatives in quick response to disruptions; • Flexible manufacturing strategies allowing options to produce critical products in multiple locations with rapid changeover capability; • Flexible product design to reduce complexity and leverage common platforms and parts, thus reducing exposure to supply disruption; • Global logistics networks to access low cost and low risk through multiple routs and contingency shipping plans. Economically efficient supply chains push the tradeoff between cost and risk. The lowest cost alternative usually is vulnerable to some kind of disruption. Some of the economic benefit from low cost has to be invested in means to enable flexible coping with disruption.

170 Enterprise Risk Management in Finance

Supply chain risk management Supply chain risk management can be described as a systematic, integrated approach to manage all risks facing an organization operating a supply chain. The benefit is the development of means to anticipate, measure, and control risk. The internet allows business to be conducted all over the globe; this presents many new opportunities for organizations to market to new customers, and thus improve their business opportunities. It is interesting to compare the old way of organizing business by vertical integration, made so successful by John D. Rockefeller and Standard Oil, by U.S. Steel, Alcoa, and others. They took the idea of system logistics developed by the military, and applied it to business, taking the approach that if there was any profit to be made in their supply chain, they wanted it. This led to vertical supply chains connecting mines, processing, transportation, and various forms of production to different levels of marketing for massive monopolies. Enforcement of such monopolies was easiest in businesses calling for high capital investment. The modern way of conducting business is quite different. Many of the formerly adversarial relationships of 19th and early 20th century businesses have been replaced by cooperative arrangements among supply chain members. The focus is on being more competitive, and thus emphasizing services related to the products being made. There also is an emphasis on linking together specialists, with a dynamic integration of often reasonably independent entities to work together to deliver goods and services. Goods and services seem ever less distinguishable, making the old dichotomy of operations passé. Global competition, technological change, and continual search for competitive advantage have motivated risk management in supply chains. Supply chains are often complex systems of networks, reaching hundreds or thousands of participants from around the globe in some cases (such as Wal-Mart or Dell). The term has been used both at the strategic level (coordination and collaboration) and the tactical level (management of logistics across functions and between businesses). In this sense, risk management can focus on identification of better ways and means of accomplishing organizational objectives rather than simply preserving assets or avoiding risk. Supply chain risk management is interested in coordination and collaboration of processes and activities across functions within a network of organizations. Supply chains enable manufacturing outsourcing to take advantages of global relative advantages, as well as increase product variety. But there are many risks inherent in this more open, dynamic system. The petroleum supply chain is critical to the world economy. Disruption of petroleum markets has had great financial impact on a number of economies. Downstream users of petroleum products feel at the mercy of upstream

Sustainability and Risk in Globalization

171

providers; but even providers have suffered extreme stress. Nigeria’s export revenues were dramatically impacted by production shutdown from December 2005 to April 2007.14 Risk management in the petroleum industry is managed through hedging, using futures, forwards, options, or swaps. Modeling support in the form of Monte Carlo simulation is often applied. Optimization is sometimes applied as well, but encounters obstacles in the form of fitting distributions (specifically, the fat tail problem). There is an increasing tendency toward an integrated or holistic view of risks. Enterprise Risk Management (ERM) is an integrated approach to achieving the enterprise’s strategic, programmatic, and financial objectives with acceptable risk. The philosophy of ERM generalizes these concepts beyond financial risks to include all kinds of risks beyond disciplinary silos. Contingency management has been widely systematized in the military. Systematic organizational planning recently has been observed to include scenario analysis, giving executives a means of understanding what might go wrong, thus giving them some opportunity to prepare reaction plans. A complicating factor is that organization leadership is rarely a unified whole, but rather consists of a variety of stakeholders with potentially differing objectives.

Global business risks Globalization involves more cross-organizational supply chains. Supply chains involve many risks, which can be categorized as internal (involving issues such as capacity variations, regulations, information delays, and organizational factors) and external (market prices, actions by competitors, manufacturing yield and costs, supplier quality, and political issues).15 Examples of internal failures are not widely publicized, although they certainly exist. Supply chain organizations need to worry about risks from every direction. In any business, opportunities arise from the ability of that organization to deal with risks. Most natural risks are dealt with either through diversification and redundancy, or through insurance, both of which have inherent costs. As with any business decision, the organization needs to make a decision considering tradeoffs. Traditionally, this has involved the factors of costs and benefits. Society is moving toward ever more complex decision-making domains, requiring consideration of ecological factors as well as factors of social equity. Internal risk management is more directly the responsibility of the supply chain organization and its participants. Any business organization is responsible for managing financial, production, and structural capacities. It is responsible for programs to provide adequate workplace safety, which has proven to be cost-beneficial to organizations as well as fulfilling social responsibilities. Within supply chains, there is need to coordinate activities with vendors, and

172

Enterprise Risk Management in Finance

to some degree with customers (through bar-code cash register information providing instantaneous indication of demand). Information systems technology provides a new era of effective tools to keep on top of supply chain information exchange. Another factor of great importance is the responsibility of supply chain core organizations to manage the risks inherent in the tradeoff between wider participation made possible through internet connections (providing a larger set of potential suppliers leading to lower costs) with the reliability provided by long-term relationships with a smaller set of suppliers that have proven to be reliable. Dealing with external risks involves more opportunities to control risk sources. In the past, some supply chains have had an influence on political systems; arms firms like that of Alfred Nobel come to mind, as well as petroleum businesses. While most supply chain entities are not expected to be able to control political risks to include wars and regulations, they do have the ability to create environments leading to labor unrest. But it is expected that supply chain organizations have an even greater influence over economic factors; while they are not expected to be able to control exchange rates, the benefit brought by monopolies or cartels is their ability to influence price. Business organizations also are responsible for developing technologies providing competitive advantage, and developing product portfolios in dynamic markets with product life cycles. The risks arise from competitors’ abilities in never-ending competition. Ritchie and Brindley (2007) viewed five major components of a framework in managing supply chain risk.16 1. Risk context and drivers: Risk drivers arising from the external environment will affect all organizations, and can include elements such as the potential collapse of the global financial system, or wars. Industry specific supply chains may have different degrees of exposure to risks. A regional grocery will be less impacted by recalls of Chinese products involving lead paint than will those supply chains carrying such items. Supply chain configuration can be the source of risks. Specific organizations can reduce industry risk by the way the make decisions with respect to vendor selection. Partner specific risks include consideration of financial solvency, product quality capabilities, and compatibility and capabilities of vendor information systems. The last level of risk drivers relate to internal organizational processes in risk assessment and response, and can be improved by better equipping and training of staff and improved managerial control through better information systems. 2. Risk management influencers: This level involves actions taken by the organization to improve their risk position. The organization’s attitude toward risk will affect its reward system, and mold how individuals within the

Sustainability and Risk in Globalization

173

organization will react to events. This attitude can be dynamic over time, responding to organizational success or decline. 3. Decision makers: Individuals within the organization have risk profiles. Some humans are more risk averse, others more risk seeking. Different organizations have different degrees of group decision making. More hierarchical organizations may isolate specific decisions to particular individuals or offices, while flatter organizations may stress greater levels of participation. Individual or group attitudes toward risk can be shaped by their recent experiences, as well as by the reward and penalty structure used by the organization. 4. Risk management responses: Each organization must respond to risks, but there are many alternative ways in which the process used can be applied. Risk must first be identified. Monitoring and review requires measurement of organizational performance. Once risks are identified, responses must be selected. Risks can be mitigated by an implicit tradeoff between insurance and cost reduction. Most actions available to organizations involve knowing what risks the organization can cope with because of their expertise and capabilities, and which risks they should outsource to others at some cost. Some risks can be dealt with, others avoided. 5. Performance outcomes: Organizational performance measures can vary widely. Private for-profit organizations are generally measured in terms of profitability, short-run and long-run. Public organizations are held accountable in terms of effectiveness in delivering services as well as the cost of providing these services. In normal times, there is more of a focus on high returns for private organizations, and lower taxes for public institutions. But risk events can make their preparations to deal with risk exposure much more important, focusing on survival.

Conclusions Technology has grown rapidly, a characteristic of our advancing civilization. We have seen tremendous gains in computer technology, in technology for war machinery, and in the use of technology to gain strategic innovation. Globalization is made possible by the ability to communicate worldwide over the internet, enabling supply chain operation through the exchange of files and information. This technology has enabled improved production methods and the development of global supply chains. While we all recognize and appreciate these benefits of technology, we have all seen cases where technology rebounds upon us with unexpected consequences. There are many downside risks to technology. Nuclear energy, which

174

Enterprise Risk Management in Finance

provides excellent characteristics with respect to global warming, is widely adopted in Europe and Japan, although viewed very negatively in the United States. Genetically engineered food is viewed in the US as a potential salvation for many starving people, but is viewed as unacceptably risky in Europe and Africa. Chinese manufacturing is considered a very important element in manufacturing survival by most retailers throughout the world, although subcontracting risks have arisen on occasion. Technology provides many valuable tools, but also introduces new risks. The history of risk management has evolved since time immemorial. Levantine and Chinese traders prior to A.D. undoubtedly coped with the risks of sailing in order to trade, as the Egyptians and Babylonians did before them. The coffee house of Lloyds of London developed as a meeting place for the seeds of the insurance industry. But adopting insurance has a cost. Risk is what businesses exist to deal with. Frederick Bernard Hawley (1907)17 declared risk-taking to be the essential function of the entrepreneur, and thus the basis of his income. Frank Knight (1921)18 argued for profit as due to the assumption of risk. Risk management is therefore not the avoidance of risk, but rather avoiding risks that the organization is not competent to cope with, while seeking risk in its area of expertise. A number of psychological-based researchers have emphasized that the role of human preference expands the interest of risk management beyond objective data concerning probabilities to the more complex judgmental forum requiring subjectivity. The works of Kahneman and Tversky (2000)19 have led to many studies in the rich field studying the psychological impact of rational decisionmaking under conditions of uncertainty. We continue to be challenged by the complexities of the interacting natural and social systems which generate the risks that keep us concerned and active.

17

Risk from Natural Disasters

Introduction Natural disasters by definition are surprises, causing a great deal of damage and inconvenience. Earthquakes are among the most terrifying and destructive natural disasters threatening humans. Emergency management has been described as the process of coordinating an emergency or its aftermath through communication and organization for deployment and the use of emergency resources. This chapter provides the state-of-the-art studies of risk and emergency management related to the Wenchuan earthquake that happened in China in May 2008. Natural disasters are the biggest challenge that risk managers face, due to the threats that go with them.1 Natural disasters by definition are surprises, causing a great deal of damage and inconvenience. Some things we do to ourselves, such as revolutions, terrorist attacks and wars; terrorism led to the gassing of the Japanese subway system, to 9/11/2001, and to the bombings of the Spanish and British transportation systems. Some things nature does to us – volcanic eruptions, tsunamis, hurricanes and tornados. The SARS virus disrupted public and business activities, particularly in Asia.2 More recently, the H1N1 virus has sharpened the awareness of the response system world-wide. Some disasters combine human and natural causes – we dam up rivers to control floods, to irrigate, to generate power – and even for recreation, as at Johnstown, PA, at the turn of the 20th century. We have developed low-pollution, low-cost electricity through nuclear energy, as at Three-Mile Island in Pennsylvania and Chernobyl. We have built massive chemical plants to produce low cost chemicals, as at Bhopal, India. Lee and Preston provide a review of high-impact, low-probability events, focusing on analysis of the Eyjafjallajökull volcano.3 This event was representative of “black swans”,4 that is, impossible-to-predict events with a very low likelihood but high costs of mitigation. Other examples include Hurricane 175

176

Enterprise Risk Management in Finance

Katrina in New Orleans, the Japanese earthquake and tsunami of 2011, and the 2003 SARS outbreak. What we don’t do to ourselves in the form of wars and economic catastrophes, nature trumps with overwhelming natural disasters. These disruptions are major components of supply chain systems, which have become key components of today’s global economy. The ability to cope with unexpected events has proven critical for global supply chain success, as demonstrated by Nokia in the past few years, as well as production halts experienced by Toyota and Sony due to the 2011 earthquake and tsunami in Japan.

Preparing for high-impact, low-probability events In a natural disaster, there will inevitably be many who feel that whatever the authorities did was overkill and unnecessary, just as there will be many who feel that the authorities didn’t do enough to (1) prevent the problem, and (2) mitigate the problem after it occurred. It is the nature of emergency management to be unthanked. Transparency, especially during and after a crisis, helps to assure the public that decisions are made on the best available evidence in order to gain public confidence and to manage vested interests. Globalized supply chains, particularly those based on just-in-time methods, are vulnerable. Famous historical examples include Nokia’s response to a March 2000 lightning strike in Albuquerque, NM, leading to a fire in a Royal Philips Electronics fabrication line that supplied RFID chips. Both Nokia and its competitor Ericsson were served by this key supply chain link,5 and Philips estimated that at least a week would be required to return to full production. Ericsson passively waited – but Nokia proactively arranged for alternative supplies, as well as redesigning products to avoid the need for those chips. Nokia gained significant advantage in this market, turning in a profit while Ericsson suffered an operating loss.6 Lee and Preston state that the maximum tolerance for supply chain disruption in a just-in-time global economy is one week. Lee and Preston draw the following recommendations from the Eyjafjallajökull experience: 1. Stress-testing risk mechanisms: This recommendation includes specifics calling for broad coordination with governments and businesses to determine as much as possible that costs and risks of worst-case situations are identified. They also call for scenario-building exercises and sharing of best practices. 2. Crisis communication: Care should be given to develop robust communication, to include websites, especially to keep the public and the media informed of risks. Independent, risk notification hubs supported by governments, businesses and industry associations were called for.

Risk from Natural Disasters

177

3. Enhancing business resilience and shock response: Governments should set up global pooling systems for reinsurance. A reference library of observations from past events was also called for. Businesses were cited as needing preparedness for management continuity, and to conduct cost-benefit analyses for alternative disruption actions. Integrated supply chains have delivered improvements in efficiency and improved the value of products we have available to us as consumers. However, highly optimized supply chain networks are inherently risky, in part because they eliminate most system slack in order to lower costs. The impact of unexpected events (Lee and Preston use SARS in 2003 and the Tōhoku earthquake/ tsunami in 2011 as examples) can be highly disruptive. The white paper does a good job of classifying the risks of various exported products to the European Union. Analysis of the impact of extended disruption was noted. Additional impact of global warming was also evaluated. Be prepared While natural disasters come as surprises, we can nevertheless be prepared. In some cases, such as Hurricane Katrina or Mount Saint Helens, we get warning signs, but we never completely know the extent of what is going to happen.7 Emergency management has been described as the process of coordinating an emergency or its aftermath through communication and organization for deployment and the use of emergency resources.8 Emergency management is a dynamic process conducted under stressful conditions, requiring flexible and rigorous planning, cooperation, and vigilance. During emergencies, a variety of organizations are often involved, and commercial rivalry can lead to normal competition, rivalry, and mutual distrust. At the governmental level, one would expect cooperation in attaining a common goal, but often so many diverse agencies get involved that attention to the overriding shared goal is dimmed by specific agency goals. Cooperation is also hampered by differences in technology.

Risks and emergencies Risks exist in every aspect of our lives. In the food production area, science has made great strides in genetic management. But there are concerns about some of the manipulations involved, with different views prevailing across the globe. In the United States, genetic management is generally viewed as a way to obtain better and more productive sources of food more reliably. However, there are strong objections to bioengineered food in Europe and Asia. Some natural diseases, such as mad cow disease, have appeared that are very difficult to control. The degree of control accomplished is sometimes disputed.

178 Enterprise Risk Management in Finance

Europe has strong controls on bioengineering, but even there a pig breeding scandal involving hazardous feed stock and prohibited medications has arisen.9 Bioengineering risks are important considerations in the food chain.10 Genetic mapping offers tremendous breakthroughs in the world of science, but involve political risks when applied to human resources management.11 Even applying information technology to better manage healthcare delivery risks involves risks.12 Reliance on computer control has been applied to flying aircraft, but hasn’t always worked.13 Risks can be viewed as threats, but businesses exist to cope with risks. Different disciplines have different ways of classifying risks. We propose the following way of classifying risks: field based and property based. ●

Field based classification: Financial risks, which basically include all sorts of risks related to financial sectors and financial aspects in other sectors; these include, but are not restricted to, market risk, credit risk, operational risk, operational risk, liquidity risk.

Nonfinancial risks, which includes risks from sources that are not related to finance. These include, but are not restricted to, political risks, reputational risks, bioengineering risks, and disaster risks. ●

Property based classification: We think risks can have three properties: Probability, dynamics, and dependence. The first two properties have been widely recognized in inter-temporal models from behavior decision and behavior economics.14 The last property is well studied in the finance discipline.

The probability of the occurrence and severity of risks mainly involves the utilization of probability theory and various distributions, to model risks. This can be dated back to the 1700s, when Bernoulli, Poisson, and Gauss used to model normal events, and generalized Pareto distributions and generalized extreme value distribution to model extreme events. The dynamics of risks mainly deals with stochastic process theory in risk management. This can be dated back to the 1930s, when the Markov processes, Brownian motion and Levy processes were developed. The dependence of risks mainly deals with correlation among risk factors; various copula functions are built, and Fourier transformations are also used here.

Technical tools Many tools have been developed to aid emergency management. Reported examples from geoscience include image enhancement through combining multiple images into a clearer composite image (mosaicing).15 Televideos and

Risk from Natural Disasters

179

wireless communications devices have been applied to aid quick response to disasters in the oil and natural gas sector.16 Statistical analysis of growth models have been used to categorize disasters into three categories of growth (damped exponential, normal, fluctuating), so that news stories can be monitored at the onset of a disaster to better predict those events that will dissipate, as opposed to those that will grow into serious disasters.17 Advanced modeling in the form of multi-objective evolutionary algorithms in combination with geographical information systems have been developed to support evacuation planning.18 Even open source software products play a role.19 SAHANA (Sinhalese for relief) is a Sri Lankan open source system built after the 2004 Asian tsunami. SAHANA supports finding missing people, managing volunteers and aid resources, and other disaster-related activities. SAHANA was deployed in the 2008 Burma cyclone and the 2008 Sichuan earthquake. Another open source system is Innovative Support to Emergencies, Diseases and Disasters (InSTEDD), started in 2006 by Larry Brilliant of the Google Foundation. InSTEDD is designed to process data from multiple sources (weather reports, news, field reports, sensor data), to detect and manage disease outbreaks.

Emergency management Local, state and federal agencies in the United States are responsible for responding to natural and man-made disasters. This is coordinated at the federal level through the Federal Emergency Management Agency (FEMA). While FEMA has done much good, it is almost inevitable that more is expected of it than it delivers in some cases, such as hurricane recovery in Florida in various years and the Gulf Coast from Hurricane Katrina in 2005. National security is the responsibility of other agencies, military and civilian (Department of Homeland Security – DHS). They are supported by non-governmental agencies such as the American Red Cross. Again, these systems seem to be effective for the greater part, but are not failsafe, as demonstrated by Pearl Harbor and 9/11/2001. Disasters are abrupt, calamitous events that cause great damage, loss of lives, and destruction. Emergency management is accomplished in every country to some degree. Disasters occur throughout the world, in every form: natural, man-made, and combination. Disasters by definition are unexpected, and tax the ability of governments and other agencies to cope. A number of intelligence cycles have been promulgated, but all are based on the idea of: 1. 2. 3. 4.

Identification of what is not known; Collection – gathering information related to what is not known; Production – answering management questions; Dissemination – getting the answers to the right people (Mueller, 2004).

180

Enterprise Risk Management in Finance

Information technology has been developing at a very rapid pace, creating a dynamic of its own. Many technical systems have been designed to gather, process, distribute, and analyze information in emergencies. These systems include communications and data. Tools to aid emergency planners communicate include telephones, whiteboards, and the internet. Tools to aid in dealing with data include database systems (for efficient data organization, storage, and retrieval), data mining tools (to explore large databases), models to deal with specific problems, and combinations of these resources into decision support systems to assist humans in reaching decisions quickly or expert systems to make decisions rapidly based on human expertise.

Emergency management support systems A number of software products have been marketed to support emergency management. These are often various forms of a decision support system. The Department of Homeland Security in the US developed a National Incident Management System. A similar system used in Europe is the Global Emergency Management Information Network Initiative (Thompson et al., 2006). While many systems are available, there are many challenges due to unreliable inputs at one end of the spectrum, and overwhelmingly massive data content at the other extreme. Decision support systems (DSS) have been in existence since the early 1970s. A general consensus is that DSSs consist of access to tailored data and customized models with real-time access for decision makers. With time, as computer technology has advanced and as the internet has become more available, there has been a great deal of change in what can be accomplished. Database systems have seen tremendous advances since the original concept of DSS. Now weather data from satellites can be stored in data warehouses, as can masses of pointof-sale scanned information for retail organizations, and output from enterprise information systems for internal operations. Many kinds of analytical models can be applied, ranging from spreadsheet models through simulations and optimization models. While the idea of DSS is now over 30 years old, it can still be very useful in support of emergency management. It still can take the form of customized systems accessing specified data from internal and external sources, as well as a variety of models suitable for specific applications needed in emergency management situations. The focus is still on supporting humans making decisions, but if problems can be so structured that computers can operate on their own (Hal in 2001 comes to mind), decision support systems evolve into expert systems. Expert systems can be, and have been, used to support emergency management. An example decision support system directed at aiding emergency response is the Critical Infrastructure Protection Decision Support System (CIPDSS).20

Risk from Natural Disasters

181

CIPDSS was developed by Los Alamos, Sandia, and Argonne National Laboratories, sponsored by the Department of Homeland Security in the US. The system includes a range of applications to organize and present information, as well as system dynamics simulation modeling of critical infrastructure sectors, such as water, public health, emergency services, telecom, energy, and transportation. The system also includes multiattribute utility functions based upon interviews with infrastructure decision-makers. CIPDSS thus serves as an example of what can be done in the way of an emergency management support system. Other systems in place for emergency management include the US National Disaster Medical System (NDMS), providing virtual centers designed as a focal point for information processing, response planning, and inter-agency coordination. Systems have been developed for forecasting earthquake impact21 or the time and size of bioterrorism attacks. This demonstrates the need for DSS support not only during emergencies, but also in the planning stage.

Conclusions Emergencies of two types can arise. One is repetitive – hurricanes have hammered the Gulf Coast of the US throughout history, and will continue to do so (just as tornadoes will hit the Midwest and typhoons the Pacific). A great deal of experience and data can be gathered for those events, and our weather forecasting systems have done a very good job of providing warning systems for actual events over the short term of hours and days. However, humans will still be caught off-guard, as with Hurricane Katrina. The other basic type of emergency are surprises. These can be natural (such as the tsunami of 2004) or human-induced, such as September 11, 2001. We cannot hope to anticipate, nor will we find it economic to massively prepare for, every surprise; we don’t think that, for example, a good asteroid collision prevention system would be a wise investment of our national resources. On the other hand, there is growing support for an effective global warming prevention system. The first type of emergency is an example of risk – we have data to estimate probabilities. The second type is an example of uncertainty – we can’t accurately estimate probabilities for the most part. (People do provide estimates of the probability of asteroid collision, but the odds are so small that they don’t register in our minds. Global warming probabilities are near certainty, but the probability of a compensating cooling event in the near future currently evades calculation.) We have reviewed some of the tools that have been reported for use in supporting disaster or emergency management. This issue includes papers that report on the effectiveness of response systems in the 2008 Sichuan earthquake. It also includes two papers addressing tools, one developed to improve

182

Enterprise Risk Management in Finance

evaluation of damage, the other applying genetic algorithm models to aid rapid decision-making in emergencies. A fourth paper addresses financial tools to deal with the insurance aspects of emergency response. Thus the crux of the problem in supporting emergency management is that tools exist to gather data, and tools such as data mining exist to try to make some sense of it, but the problem is that we usually won’t have the particular data that will be useful to make decisions in real time. It is also reasonably certain that after any event, critics will be able to review what data was available and to point to tell-tale information that could have enlightened decision-makers at the time but didn’t, for example, after World War II, the US was flooded with people who thought that the US Navy should have known the Japanese would bomb Pearl Harbor. CNN and national networks have very predictable scripts for every emergency, with reporters playing to the camera, pointing out the gross malfeasance of those in control in not knowing, preparing for, and countering whatever happened. That’s how they raise their ratings – the audience likes conspiracy theories. But having data is not enough – human minds have to comprehend the core information, and the more information that is provided, the harder that is. The solution is not LESS information, but some filters to focus on the critical core would help provide a solution. The next problem, though, is that we don’t know how to create such filters, especially in new problem domains. Emergency management is thus a no-win game. However, someone has to do it. They need to do the best they can in preplanning: 1. gathering and organizing data likely to be pertinent; 2. developing action plans that can be implemented at the national, regional, and local level; this can include development of and implementation of building codes, environmental awareness, and insurance systems; 3. organizing people into teams to respond nationally, regionally, and locally, trained to identify events, and to respond with all needed systems (rescue, medical, food, transportation, control, etc.); this can include the training and development of planners and managers of response teams.

18

Pricing of Carbon Emission Exchange in the EU ETS

Introduction Carbon emission exchange originated from emission trading proposed by economists in the 1970s. Carbon trading, an important environmental policy in market economy countries, has emerged as the foremost policy instrument for reducing worldwide greenhouse gas emission. The United Nations Intergovernmental Panel on Climate Change (UNIPCC) passed the United Nations Framework Convention on Climate Change (UNFCC) on June 4, 1992. The Kyoto Protocol, passed in December 1997, the first additional convention, uses the market mechanism as a new way to resolve the issue of greenhouse gas reduction, of which carbon emission is the most prominent. Thus carbon emission rights become a tradable commodity, leading to the emergence of a carbon emission exchange mechanism. In accordance with the reduction commitment made in the Kyoto Protocol, some countries from the European Union (EU) signed an expense-sharing convention in June 1998. At the same time, the European Union Commission released a report, entitled Climate Change: Towards an EU Post-Kyoto Strategy, which authorized an exchange system in the European Union before 2005. A draft of the European Union Emission Trading System (EU ETS) was submitted and discussed formally in 2001, and passed in the European Union Parliament in October, 2002. One year later, the revised version was passed in the European Union Parliament and Council in July, and Emission Exchange Directive 2003/87/EC came into effect on October 13, authorizing the EU ETS to start exchanging carbon emissions beginning in January 2005. Thus the European emission exchange system was established. The European Union Emission Trading System (EU ETS) is not only the largest multinational emission trading scheme in the world, but is also a major pillar of the EU climate policy aimed at efficiently reaching the European emission

183

184

Enterprise Risk Management in Finance

commitment targets assigned by the Kyoto Protocol at the minimum cost. The EU ETS comprises three trading periods or phases, the first from January 2005 to December 2007, the second from January 2008 to December 2012, and the third from January 2013 to December 2020. Under the EU ETS, each state member is assigned a national emission quota, precisely stated in the National Allocation Plan (NAP), which has to be approved by the EU Commission. Then national allowances are distributed among their industrial operators with permits, and the actual emission amount in line with NAP is supervised. Allowances are supposed to expire at the end of each year and traded or privately reassigned, on the over-the-counter market or one of the European climate exchanges, such as European Climate Exchange, BlueNext, PowerNext, Nord Pool and others. As a commodity, the supply and demand of allowances in the EU ETS market determines the price for carbon, like other bulk commodities. A greater supply of allowances will lead to a lower carbon price. On the other hand, too much demand for allowances will result in a higher carbon price. During Phase I, all the participating countries accepted most of their allowances freely. However, the share of auctioning allowances was improved greatly during Phase II, which seemed to be more market-driven. In Phase III, a substantial number of permits are being allocated centrally with a large share of auctioning permits, a different method from that used in the National Allocation Plan. Within a trading phase, banking and borrowing is allowed. For example, a 2009 EUA could be used in 2010 (Banking) or in 2008 (Borrowing). However, cross-period borrowing and banking is not allowed. Member states do not have the discretion to bank EUA from Phase I to Phase II. The price mechanism of the carbon emission exchange is one of the crucial factors for its success due to the increasingly intensified world attention to global carbon emission reduction. The advanced price mechanism of the EU ETS was based on five years of experience, and a few scholars conducted initial theoretical and empirical research studying it. Cities in China, such as Beijing, Tianjin, Shanghai, Wuhan, Changsha, Shenzhen and Kunming, have gradually set carbon emission exchanges, but they are only at the preliminary stage and the trading volumes are expected to be low. Therefore, learning or studying the previous EU ETS experience seems useful in establishing the Chinese carbon emission exchange system.

Literature review There is growing interest in carbon emission exchange for scholars, especially with the development of the EU ETS over the past five years. Several scholars have studied the EU ETS price mechanism, which covers three parts:

Pricing of Carbon Emission Exchange in the EU ETS

185

micro-simulation models, empirical research, and determinants of carbon emission exchange. We begin with simulation models. Scholars have studied numerical micro simulation models regarding carbon emission exchange, frequently using the marginal abatement cost curve (MACC) as an analysis tool. Kainuma et al. (1999)1 generated an Asia-Pacific Integrated Model (AIM), which treats carbon rights as a constraint on a production function, and as a result emission targets and trading channels become key factors in influencing the price schemes of carbon emission. The Regional Integrated Model of Climate and the Economy (RIMCE) proposed by Nordhaus and Boyer (1999)2 and by Nordhaus (2001)3 differs from AIM in the way that it includes the participation of the USA and that the emission targets of each country are determinants of CO2 prices. Other researchers4 simulate the allowance price, and their calculations yield between €6 and €35 per ton of CO2, depending on their models and settings. However, these models only reflect the equilibrium price of carbon emission allowances and cannot simulate price fluctuations over periods. Carbon emission trading develops quickly, and some scholars have conducted empirical research on the price mechanism using the abundance of trading data. Daskalakis et al. (2009)5 found that market participants adopt non-arbitrage standard pricing in order to check emission allowance prices and derivatives. Uhrig-Homburg and Wagner (2006)6 studied the optimal design of derivatives based on emission allowances. Paolella and Taschini (2006)7 integrated both the EU ETS market and the US Clean Air Act Amendments, providing an econometric analysis investigating the unconditional tail behavior and the heteroskedastic dynamics in the returns on CO2 and SO2 allowances in order to set up hedging and purchasing strategies. Benz and Trück (2006)8 argued that the emission allowance market is different from the classical stock market as the value of stocks depends on the profit expectations of firms, but the price of carbon emission allowances is determined by the supply and demand of carbon permits in the market. Using the concept of convenience yields, Borak et al. (2006)9 focused on term structure and stochastic properties, and showed that the carbon emission allowances market differs from the existing commodities markets. Seifert et al. (2008)10 set an equilibrium model appropriate to features of the EU ETS to analyze spot price dynamics of CO2. They found that an adequate CO2 process did not necessarily follow any seasonal patterns. Daskalakis et al. (2009) analyzed the effect on pricing of banking and borrowing design among different phases in three exchanges under the EU ETS and provided corresponding interphase and intraphase pricing and arbitraging strategies. But none of the scholars provided both a pricing model and an empirical analysis considering the two phases of the EU ETS. Identifying the determinants of allowances price is yet another research area. Burtraw (1996)11 categorized the influencing factors into three groups: policy

186

Enterprise Risk Management in Finance

issues, market fundamentals and the demand and supply of carbon emission allowances. Burniaux (2000)12 alternatively assumed that the price of fuels, the average carbon content in energy usage, energy substitution possibilities, and the price and availability of clean substitute fuels are the driving factors for CO2 pricing. Considine (2000)13 studied emission movements considering the impact of weather. The hotter and colder seasons have a great impact on energy consumption, and as a result the demand for carbon emission allowances becomes much larger. Therefore, temperature variation is one of the driving factors. On the other hand, the prices of oil and gas have a positive effect on carbon emission allowances prices. Sijm et al. (2005)14 come to a similar conclusion, and find that energy prices, characteristics of the energy sector, demand and supply of allowances, and economic structure play key roles in the formation of allowances price. Ciorba et al. (2001)15 pointed out that the price of energy, the level of emission, the geographical features of a country and its climatic conditions and temperature are the most influential factors on setting allowances prices. Springer (2003)16 indicated that the determinant of prices mainly lies on the level of emission. Springer and Varilek (2004)17 explore other factors that could affect the cost of compliance with the Kyoto Protocol, such as the number of sectors in the economy that do not participate in either the trading of allowances or the inter-period transfer of allowances. From an econometrical perspective, Manasanet-Bataller, Pardo, and Valor (2007)18 attempted estimation of the effects of some determinants of the EU ETS daily forward prices in 2005. Explanatory variables were oil, natural gas and coal prices. That study also included weather variables, which are generally important determinants of allowances prices. In summary, most studies concluded that the level of emission, energy price and weather variables are the main driving factors in the formation of the carbon emission price. In general, very few studies have used data from carbon emission exchanges to test quantitative models for allowances pricing. Furthermore, there is no literature on exploring variations of price dynamics with different phases of the EU ETS. Therefore, in this study we have tried to investigate the EUA price dynamics and select an appropriate model in order to investigate the distinctions and improvements between the two phases.

Price movements In this section, the price movements in the EU ETS over five years are described and explained. Data and news are taken from news and weekly trading reports of Point Carbon and Climate Group, two key websites in carbon emission research. Figure 18.1 shows the price movements of carbon emission in Phase I, a trial phase with large price volatility. It consisted of five stages in price movements. The first stage was recorded from the start of the phase to the middle

Pricing of Carbon Emission Exchange in the EU ETS

187

35 30 25 20 15 10 5

20

05 20 –4– 05 3 20 –6– 05 3 20 –8 05 –3 20 –10 05 –3 – 20 12 06 –3 20 – 2 – 06 3 2 0 –4– 06 3 20 – 6 – 06 3 20 –8 06 –3 20 –10 06 –3 – 2 0 12 0 7 –3 20 –2– 07 3 20 –4– 07 3 20 – 6 – 07 3 20 –8 07 –3 2 0 –10 0 7 –3 –1 2– 3

0

Figure 18.1

EUA price movement in Phase I

of July 2005; the price was stimulated to €28.9 by the European Commission’s announcement further cutting the National Allocation Plan (NAP) for Italy, Poland and the Czech Republic, and the denial of the United Kingdom to extend the amount of certificates. The second stage was from the middle of July to December 2005; early participants in the EU ETS sold their allowances at high prices before the price dropped when new members from Eastern Europe entered the market. The third stage covered January to April 2006; most research reports released an announcement about over-discharging CO2 compared with the EUA supply, and the carbon emissions price went up. The fourth stage was the period from May to December 2006; auditing and supervision by the European Union was minimal, and every country excessively distributed free allowances to their industry operators. Then the 2005 CO2 data indicated that the current allowances in the market were greatly oversupplied and the price dropped dramatically over a few trading days. The fifth stage was between January and December 2007; European countries released their Phase II NAPs and the rule of prohibition on banking and borrowing between Phases I and II was clarified. The EUA price fell to zero. The price movement in Phase II was stable (Figure 18.2) with reduction commitment in accordance with the Kyoto Protocol. The period from January 2008 to 2010 could be divided into three stages. The first between January and June 2008; with weak economic expectations and rising coal prices, the EUA price was reduced to €18.84. However, it rose back up to €28.59 due to the impact of oil and gas prices. The second stage was from July 2008 to February 2009; as the financial crisis burst out all over the world, demand for oil and gas dropped significantly, which led to reduced demand for EUA. At the same time, the ratio for free allotments and tradable allowances rose. The EUA

188

Enterprise Risk Management in Finance

35 30 25 20 15 10 5

Figure 18.2

–2

20

10

–3

–1

–2

2 20

10

1–

2

–1

09

20

20 0

9–

7–

9–

2

–2

9–

20 0

–5

–2 20

09

–3

–2 20

09

–1

–2 11

09 20

8–

20 0

20

08

–9

–2

–2 –7

–2 –5

08 20

–2

08

20

–3 08

20

20

08

–1

–2

0

EUA price movement in Phase II

price declined, with low trading volumes in this stage. The third stage was from March 2009 to 2010; oil, gas and power prices went up with economic recovery and the EUA price increased from €7.9 to €15, over one year fluctuating between €12 and €15. This data shows that the price volatility in Phase I was much larger than in Phase II, primarily influenced by political issues and trading rules. First, dispensed allowances exceeded discharged CO2, which led to the distortion of the EUA price. The price mechanism was especially disturbed by political issues. For example, the European Union announcement regarding the oversupply of allowances led to the EUA price drop to €10 within several trading days. Second, micro statistical data are absent, which is one of the reasons for the oversupply circumstances of EUA. Consequently, a high portion of free allowances distributed to enterprises pushed the EUA price down. Power enterprises received much higher free allowances than other enterprises, and sold their allowances for profits. This made the existing EUA superfluous, adverse to formation of a rational price in Phase I. In Phase II, the EU ETS improved the rules on NAP supervision, distribution of allowances and allotment methods. As a result, the price mechanism returned to fundamentals: energy price, power price, and weather variables. An advanced trading and price mechanism has preliminarily been formed.

Model, data and sample Generally, the right for carbon emission is a public good with externality, and its impact is not directly reflected in market cost and price. However, political conventions, such as the United Nations Framework Convention on Climate Change (UNFCC) and the Kyoto Protocol led to market scarcity and

Pricing of Carbon Emission Exchange in the EU ETS

189

added economic value as a special commodity. Carbon emission exchange makes use of basic market functions in governing climate change, and the price of carbon emission can reflect resource scarcity and governance cost. The rights to carbon emission take on attributes of a commodity through the exchange, which means every participant treats the cost of carbon discharge as an important factor in investment, with price signals providing guidance to internalize external environmental costs. With the expansion of the carbon emission market, improvement of trading transparency, and monetization of carbon, carbon emission rights become a financial asset with high liquidity due to the participation of more financial institutions. The carbon emission market is an important emerging financial market, and it is appropriate to use financial market models to study its price mechanism. We use EGACH19 to model the price volatility of Carbon Emission Exchange in the EU ETS. The EU ETS includes a lot of carbon emission exchanges: PowerNext, Nord Pool and European Climate Exchange (ECX) are the top three. We examine spot close price of carbon emission exchange in the ECX over the period April 2005 to April 2010. This period is divided into two parts, as two independent samples based on the European Commission directive. The first part is from April 2005 to December 2007, the second from January 2008 to April 2010. All data are collected from Bloomberg. Analysis of EUA logreturns We analyze the volatility of EUA logreturns in order to study its price mechanism. Figures 18.3 and 18.4 show logreturns of EUA. The volatility of logreturns in the first phase is apparently much larger than in the second. In the first phase, logreturn fluctuates between 1.0 and −1.5 and some peaks appear in April 2006, in the middle of the time frame, and at the end of 2007. In the 1.0

0.5

0.0

–0.5

–1.0

–1.5 100 Figure 18.3

200

Logreturns of EUA in Phase I

300

400

500

600

190 Enterprise Risk Management in Finance

.15 .10 .05 .00 –.05 –.10 –.15 50 100 150 200 250 300 350 400 450 500 Figure 18.4

Logreturns of EUA in Phase II

Table 18.1 Descriptive statistics of EUA logreturns in Phase I

Series

N

Mean

In-sample 433 −0.002 (2005.4–2006.12) Out-of-sample 237 −0.027 (2007.1–12)

Median Max

Min

Standard Deviation

Skew

Kurt

0

0.418 −0.396

0.046

−0.489 32.792

0

0.916 −1.386

0.145

−2.589 41.592

Table 18.2 Descriptive statistics of EUA logreturns in Phase II

Series

N

Mean

Median

Max

Min

In-sample 458 −0.001 0 0.111 −0.127 (2008.1–2009.12) Out-of-sample 68 0.003 4.26E-05 0.047 −0.033 (2010.1–4)

Standard Deviation

Skew

Kurt

0.029

−0.204 4.770

0.019

0.261 2.552

second phase, volatility is much smaller than in the first, fluctuating between 0.15 and −0.15. To estimate and test the price mechanism of EUA, each phase is divided randomly, one for estimation, the other for forecasting. The estimation stage of Phase I is from April 2005 to December 2006, and the remainder for forecasting. For the second phase, the estimation stage is from January 2008 to December 2009, and the remainder for forecasting. Tables 18.1 and 18.2 give descriptive statistics of the two phases.

Pricing of Carbon Emission Exchange in the EU ETS

191

The skew of estimation and forecasting in Phase I is −0.489 and −2.589, while the kurtosis measures are 32.792 and 41.592 respectively. Similarly, the skew of estimation and forecasting during Phase II are −0.204 and 0.261, with kurtosis of 4.770 and 2.552. These data show large skew and kurtosis. Both stages in Phase I are left-skewed, but during phase II left-skewed in estimation and rightskewed in forecasting. Due to asymmetry, excess kurtosis and heavy tails, the normal distribution does not fit the data very well. Given the large volatility of logreturns, the model used must convey and explain the data characteristics described above. Time series test A root unit test is conducted on logreturns and its first-order difference in order to test time series stationarity. Phase I shows that t statistics values are larger than critical value at the 1%, 5% and 10% levels. They are all significant at these levels, and null hypotheses are rejected. We can conclude that both the logreturns and their first-order difference series are stable. Similarly, the ADF test of logreturns and their first-order difference in Phase II are also significant at the 1%, 5% and 10% levels, and null hypotheses are denied. The logreturns and their first order difference series are also inferred to be stable. Table 18.3 The ADF test of EUA logreturns in Phase I

Augmented Dickey–Fuller test statistic Test critical value 1% level 5% level 10% level

t-Statistic

Probability

−24.631 −3.440 −2.866 −2.569

0.0000

Table 18.4 The ADF test of EUA first-order difference in Phase I

Augmented Dickey–Fuller test statistic Test critical value 1% level 5% level 10% level

t-Statistic

Probability

−17.751 −3.440 −2.866 −2.569

0.0000

Table 18.5 The ADF test of EUA logreturns in Phase II

Augmented Dickey–Fuller test statistic Test critical value 1% level 5% level 10% level

t-Statistic

Probability

−17.023 −3.440 −2.866 −2.569

0.0000

192 Enterprise Risk Management in Finance

Table 18.6 The ADF test of EUA first-order difference in Phase II

Augmented Dickey–Fuller test statistic Test critical value 1% level 5% level 10% level

Autocorrelation

Partial Correlation 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Figure 18.5

PAC

Q-Stat

Prob

0.048 0.048 0.025 0.022 0.024 0.022 0.028 0.025 0.001 –0.003 0.017 0.015 0.024 0.022 0.087 0.084 –0.016 –0.026 –0.027 0.031 0.053 0.053 –0.030 –0.038 0.051 0.055 –0.066 –0.075 –0.044 –0.0443 –0.059 –0.059 0.020 0.031 –0.079 –0.072 –0.011 –0.012 –0.019 –0.007 –0.039 –0.041 0.006 0.029 –0.006 0.007 0.024 0.025 0.051 0.057 0.008 0.011 –0.026 –0.016 0.047 0.040 –0.174 –0.172 –0.008 –0.014 –0.054 –0.050

AC

1.5509 1.9561 2.3391 2.8719 2.8723 3.0644 3.4711 8.6073 8.7861 9.2784 11.228 11.845 13.650 16.667 17.970 20.403 20.665 24.974 25.063 25.311 26.357 26.379 26.407 26.797 28.602 28.648 29.113 30.650 51.976 52.027 54.116

0.213 0.376 0.505 0.579 0.720 0.801 0.838 0.377 0.457 0.506 0.424 0.458 0.399 0.274 0.264 0.203 0.242 0.126 0.158 0.190 0.193 0.236 0.282 0.314 0.281 0.327 0.355 0.333 0.005 0.008 0.006

t-Statistic

Probability

−15.201 −3.440 −2.866 −2.569

0.0000

Autocorrelation

Partial Correlation

AC 1 0.128 2 –0.103 3 0.027 4 –0.010 5 –0.014 6 –0.022 7 0.031 8 0.028 9 –0.005 10 –0.017 11 –0.017 12 –0.008 13 0.051 14 0.016 15 0.019 16 0.013 17 0.046 18 –0.028 19 –0.002 20 –0.013 21 –0.030 22 –0.011 23 –0.050 24 –0.047 25 –0.022 26 –0.055 27 0.041 28 0.044 29 0.044

PAC

Q-Stat

Prob

0.128 –0.121 0.060 –0.036 0.003 –0.029 0.040 0.012 –0.001 –0.016 –0.014 –0.007 0.055 –0.001 0.029 0.002 0.054 –0.043 0.026 –0.034 –0.017 –0.014 –0.049 –0.039 –0.019 –0.060 0.059 0.014 0.055

8.7264 14.301 14.701 14.750 14.847 15.115 15.614 16.030 16.045 16.196 16.345 16.380 17.802 17.936 18.135 18.233 19.383 19.806 19.806 19.898 20.395 20.459 21.847 23.052 23.310 24.967 25.889 26.947 28.052

0.003 0.001 0.002 0.005 0.011 0.019 0.029 0.042 0.066 0.094 0.129 0.174 0.165 0.210 0.256 0.310 0.307 0.344 0.406 0.464 0.496 0.554 0.530 0.517 0.559 0.521 0.525 0.521 0.515

Series correlation of EUA logreturns in Phases I and II

Time series correlations tests are also conducted. Figure 18.5 shows that both the logreturns in Phase I and Phase II are not autocorrelated. Durbin– Watson statistics values are 1.9996 and 1.9852 respectively, which is closer to 2, inferring that there is no autocorrelation. GARCH effect test In the third step, ARCH LM is chosen to test whether there is a GARCH effect or not. Figures 18.3 and 18.4 show a clustering effect of EUA logreturns and it probably has heteroskedasticity. An ARCH LM test is conducted for residual series of EUA logreturns in Phase I and Phase II respectively. GARCH effect was found in both phases (see Table 18.7).

Pricing of Carbon Emission Exchange in the EU ETS

193

Table 18.7 The ARCH LM test for EUA logreturns in Phases I and II ARCH test

Phase I (2005.4 –2007.12) Phase II (2008.1–2010.4)

ARCH-statistic

Critical value

Probability

6.4223e-004 2.3563e-004

3.842 3.842

0.980 0.867

Note: Significance level α = 0.05.

Table 18.8 The Akaike info criterion (AIC) and Schwarz criterion (SC) for the estimated model in Phase I Model Egarch (1, 1)-t Garch (1, 1)-normal Garch (1, 1)-GED

Akaike info criterion (AIC)

Schwarz criterion (SC)

−4.290 −4.121 −4.272

−4.243 −4.092 −4.235

Table 18.9 The Akaike info criterion (AIC) and Schwarz criterion (SC) for the estimated model in Phase II Model Egarch (1, 1)-t Garch (1, 1)-normal Garch (1, 1)-GED

Akaike info criterion (AIC)

Schwarz criterion (SC)

−4.406957 −4.362285 −4.389862

−4.361904 −4.335253 −4.353820

Method selection The above evidence shows that GARCH cluster models fit the price mechanism of carbon emission exchange better. Three typical models, EGARCH (1, 1)-t, GARCH (1, 1)-normal and GARCH (1, 1)-GED, are chosen for comparison to select the best model in the GARCH models cluster. To evaluate the models, the Akaike info criterion (AIC) and Schwarz criterion (SC) are chosen to examine the estimated models. Choosing the most parsimonious model, it is found that the EGARCH (1, 1)-t model is better. In Phase I, while AIC values are −4.290, −4.121, and −4.272 respectively, SC values are −4.243, −4.092 and −4.235. Although the differences between EGARCH (1, 1)-t and GARCH (1, 1)-GED are relatively smaller, EGARCH (1, 1)-t is still the best model considering both criteria. Similarly, in Phase II, AIC and SC values are −4.407, −4.362, −4.390 and −4.362, −4.335 and −4.354 respectively (see Table 18.9). The differences between EGARCH (1, 1)-t and GARCH (1, 1)-GED are very small as well. EGARCH (1, 1)-t is still preferred.

194 Enterprise Risk Management in Finance

In conclusion, EGARCH (1, 1)-t is an adequate approach for modeling EUA logreturns. Estimation and forecasting As mentioned above GARCH models are appropriate, and the EGARCH (1, 1)-t is the best model among the three types for EUA price estimation and forecasting. Therefore, an EGARCH (1, 1)-t model is used to estimate and forecast EUA price in Phase I and Phase II. All estimation and forecasting was obtained using Eviews 6.0 and MATLAB2009a. (1) Estimation and forecasting of logreturns in Phase I In Phase I, the estimation period is from April 2005 to December 2006, and the forecasting period is from January to December 2007. Table 18.10 shows that all the coefficients of variables are significant at the α = 0.05 level. The T-DIST.DOF value of 3.569 means that the model EGARCH(1,1,)-t with a low degree of freedom can describe EUA pricing very well. The AIC and SC values are −4.290 and −424, while the Durbin–Watson statistic of 1.696 shows a low level of autocorrelation. An EUA estimation model in Phase I is obtained. Figure 18.6 shows residual, standard deviation and logreturns series in Phase I. As expected, the trend of logreturns and residual fit closely, because they have the same distribution. The Phase I estimation model is used to forecast returns from January to December, 2007, and the results are tested. Root mean squared error, mean absolute error and mean abs. percent error values are 0.148, 0.061 and 43.038 respectively (see Table 18.11). This estimation model can predict the EUA price trend very well and can be used to forecast EUA price in Phase I.

Table 18.10

Estimation of EUA logreturns in Phase I

Variable

Coefficient

Std. error

z-Statistic

Probability

Variance equation C(1) C(2) C(3) C(4) T-DIST.DOF Akaike info criterion

−1.040 0.530 −0.121 0.903 3.569 −4.290

0.234 0.094 0.057 0.029 0.709 Durbin– Watson stat

−4.436 5.640 −2.141 31.259 5.033 1.696

0.000 0.000 0.032 0.000 0.000

Schwarz criterion

−4.243

Pricing of Carbon Emission Exchange in the EU ETS

195

Innovations Innovation

2 0

Standard Deviation

–2 0

100

200

300

400

500

600

700

500

600

700

500

600

700

Conditional Standard Deviations

1 0.05 0 0

100

200

400

Returns

2 Return

300

0 –2 0

100

200

300

400

Figure 18.6

Residual, standard deviation and logreturns series in Phase I

Table 18.11

EUA Logreturns forecasting in Phase I

Model EGARCH (1,1)-t

Variable no.

Root mean squared error

Mean absolute error

Mean absolute percent error

4

0.148

0.061

43.038

(2) Estimation and forecasting of logreturns in Phase II In Phase II, the estimation period is from January 2008 to December 2009 and the forecasting period is from January to April 2010. Empirical results are illustrated in Table 18.12 using the same method. All the coefficients of variables are significant at the α = 0.05 level. The T-DIST.DOF value of 8.423 means EGARCH (1,1)-t at a low degree of freedom can describe EUA pricing very well. The AIC and SC values are −4.407 and −4.362, while the Durbin–Watson statistic is 1.722, which is closer to 2, exhibiting a low level of autocorrelation. An EUA estimation model in Phase II is also conducted. Figure 18.7 shows residual, standard deviation and logreturns series in Phase II. The trend of logreturns and residual series match those of Phase I well. On the other hand, volatility in Phase II is much smaller than in Phase I (see Figures 18.6 and 18.7).

196 Enterprise Risk Management in Finance

Table 18.12

Estimation of EUA logreturns in Phase II

Variable

Coefficient

Std. error

z-Statistic

Probability

Variance equation C(1) C(2) C(3) C(4) T-DIST.DOF Akaike info criterion

−0.312 0.126 −0.120 0.971 8.423 −4.407

0.110 0.053 0.031 0.012 4.023 Durbin– Watson stat

−2.840 2.363 −3.881 81.385 2.093 1.722

0.005 0.018 0.000 0.000 0.036

Schwarz criterion

−4.362

Innovations Innovation

0.2 0 –0.2 0

100

300

400

500

600

400

500

600

400

500

600

Conditional Standard Deviations

0.1 Standard Deviation

200

0.05 0 0

100

200

Returns

0.2 Return

300

0 –0.2 0

Figure 18.7

100

200

300

Residual, standard deviation and logreturns series in Phase II

The Phase II estimation model is used to forecast returns from January to April 2010 and the results are analyzed. root mean squared error, mean absolute error and mean abs. percent error values are 0.019, 0.015 and 100.000 respectively (see Table 18.13). This estimation model can predict EUA price trend very well, and can be used to forecast EUA price in Phase II.

Pricing of Carbon Emission Exchange in the EU ETS

Table 18.13

197

EUA Logreturn forecasting in Phase II

Model EGARCH(1,1)-t

Variable no.

Root mean squared error

Mean absolute error

Mean absolute percent error

4

0.019

0.0149

100.000

(3) Comparison of results As shown in these results, the EGARCH (1,1)-t estimates and forecasts EUA price very well in both Phase I and Phase II, and is appropriate to predict EUA price. However, the different phases have distinctive price characteristics. First, the price volatility differences in the two phases are seriously large; the EUA price description in Figures 18.1 and 18.2, logreturns of EUA in Figures 18.3 and 18.4, Tables 18.1 and 18.2, and residual, standard deviation and logreturns series in Figures 18.6 and 18.7 all show that large differences exist. The volatility in Phase I is substantially greater than that in Phase II. Second, the pricedriven mechanisms in each phase are different. Estimation and forecasting in Tables 18.8, 18.9, 18.10 and 18.11 show that EGARCH (1,1)-t can be used to predict the EUA price mechanism. But coefficient, standard error and other statistics reveal a diverse price driven mechanism. The data of Phases I and II are modeled as Phase I for estimation and Phase II for forecasting, but the results are not significant. Therefore, all the results of our analysis confirm the divergent EUA price mechanism in each phase.

Conclusions The carbon emission exchange as a market-driven reduction approach is one of the most interesting topics for academic research in the last several years. In this study, the price mechanism of the EU ETS in two phases were analyzed. After a description of the EUA price movements in recent years, three typical GARCH models for EUA pricing were examined: EGARCH (1,1)-t, GARCH (1,1)normal and GARCH (1,1)-GED were selected as the most appropriate to model with AIC and SC. We found that EGARCH (1,1)-t is the best model among the three. Estimation and forecasting in Phase I and Phase II were conducted. The results strongly suggest that both price mechanism and volatility are dramatically different in each phase. With some learning obtained in Phase I operations, an improved price mechanism formed in Phase II. Modeling the price mechanism of EUA will not only be beneficial to traders, brokers and risk managers in the carbon market, but may also enable companies to monitor the costs of CO2 emission in their production processes.

198 Enterprise Risk Management in Finance

The research on carbon emission exchange is at a beginning stage. More organizations and individuals will participate in the market, which will bring more liquidity, trading transparency, and trade volumes. In this study we found the best model for EUA price estimation and forecasting and saw that there are substantial differences between Phase I and Phase II in price volatility and mechanism. Future research should be able to investigate the reasons for the differences between the two phases in order to more completely identify trading rules and mechanisms in the EU ETS.

19

Volatility Forecasting of the Crude Oil Market

Introduction Risk analysis of the crude oil market has always been a core research problem important to both practitioners and academia. Risks arise primarily from changes in oil prices. During the 1970s and 1980s there were a number of steep increases in oil prices; these price fluctuations reached new peaks in 2007 when the price of crude oil doubled during the financial crisis, and double digit fluctuations continued between 2007 and 2008 for short periods. These fluctuations would not be worrisome if oil was not such an important commodity in the world’s economy. But when oil prices become too high and their volatility increases, they have a direct impact on the economy in general, and affect the government decisions regarding market regulation, thus impacting firm and individual consumer incomes.1 Price volatility analysis has been a hot research area for many years. Commodity markets are characterized by extremely high levels of price volatility. Understanding the volatility dynamic process of oil price is a very important and crucial way for producers and countries to hedge various risks and to avoid excess exposures to risks (Hung et al., 2008). To deal with different phases of volatility behavior and the dependence of variability of time series on their cycles, models allowing for autocorrelation as well as heteroskedasticity such as ARCH, GARCH or regime-switching models have been suggested. The former two are very useful in modeling a unique stochastic process with conditional variance; the latter has the advantage of dividing the observed stochastic behavior of a time series into several separate phases with different underlying stochastic processes. Both types of models are widely used in practice. Hung et al. (2008)2 employed three GARCH models (GARCH-N, GARCH-t and GARCH-HT) to investigate the influence of fat-tailed innovation processes

199

200 Enterprise Risk Management in Finance

on the performance of VaR estimates of energy commodities. Narayan et al. (2008)3 used the exponential GARCH models to evaluate the impact of oil price on the nominal exchange rate. Malik and Ewing (2009)4 employed bivariate GARCH models to estimate the relations between five different US sector indexes and oil prices in their validation of cross-market hedging and investor information sharing. Regime-switching has also been used in modeling stochastic processes with different regimes. Alizadeh et al. (2008)5 introduced a Markov regime switching vector error correction model with GARCH error structure, and demonstrated how portfolio risks are reduced using state dependent hedge ratios. Aloui and Jammazi (2009)6 employed a two regime Markov-switching EGARCH model for analysis of oil price changes, and calculated probabilities of transition across regimes. Klaassen (2002)7 developed a regime-switching GARCH model to account for the high persistence of shocks generated by changes in variance processes. Oil shocks were found to better explain the impact of oil on output growth.8 There is no clear evidence regarding which approach outperforms the other. Fan et al. (2008)9 argued that the GED-GARCH-based VaR approach is more realistic and more effective than widely used historical simulation with ARMA forecasts based on their empirical study. The FIAPARCH model is said to outperform the other models in VaR prediction.10 GARCH models also seem to perform better than inversion of the Black equation in estimating implied volatility. GARCH was believed to perform best when assuming GED distributed errors.11 Clear evidence of regime-switching has been discovered in the oil market. Engel (1994)12 concluded that regime-switching models provide a useful framework for predicting the evolution of volatility and forecasting exchange rate volatility. The regime-switching stochastic volatility model performs well in capturing major events affecting the oil market.13

Volatility models Historical volatility We assume εt to be the mean innovation for energy log price changes or price returns. To estimate the volatility at time t over the last N days we have:   1  N −1  VH,t =   ∑ i = 0 ε t2− i    N  

1/2

,

where N is the forecast period. This is actually an N-day simple moving average volatility, where the historical volatility is assumed to be constant over the estimation period and the forecast period. To involve the long-run or

Volatility Forecasting of the Crude Oil Market

201

unconditional volatility using all previous returns available at time t, we have many variations of the simple moving average volatility model.14 ARMA(R,M) Given a time series Xt, the autoregressive moving average (ARMA) model is very useful for predicting future values in time series where both an autoregressive (AR) term and a moving average (MA) term are present. The model is usually then referred to as the ARMA(R,M) model, where R is the order of the first term and M is the order of the second term. The following ARMA(R,M) model contains the AR(R) and MA(M) models: R

M

i =1

j =1

Xt = c + ε t + ∑ ϕ i Xt − i + ∑ θ j ε t − j . where φi and θj are parameters for AR and MA terms respectively. ARMAX(R,M, b) To include the AR(R) and MA(M) models and a linear combination of the last b terms of a known and external time series dt, one can use an ARMAX(R,M, b) model with R autoregressive terms, M moving average terms and b exogenous inputs terms: Xt = c + ε t +

R

∑ϕ X i

i =1

t −i

M

b

j =1

k =1

+ ∑ θ j ε t − j + ∑ η k dt − k ,

where η1, ... ,ηb are the parameters of the exogenous input dt. ARCH(q) Autoregressive Conditional Heteroskedasticity (ARCH) modeling is the predominant statistical technique employed in the analysis of timevarying volatility. In ARCH models, volatility is a deterministic function of historical returns. The original ARCH(q) formulation proposed by Engle15 models conditional variance as a linear function of the first q past squared innovations: q

σ t2 = c + ∑ α i ε t2− i . i =1

This model allows today’s conditional variance to be substantially affected by the (large) squared error term associated with a major market move (in either direction) in any of the previous q periods. It thus captures the conditional

202 Enterprise Risk Management in Finance

heteroskedasticity of financial returns, and offers an explanation of the persistence in volatility. A practical difficulty with the ARCH(q) model is that in many of the applications a long-length q is needed. GARCH(p,q) Bollerslev’s Generalized Autogressive Conditional Heteroskedasticity [GARCH(p,q)] specification16 generalizes the model by allowing the current conditional variance to depend on the first p past conditional variances as well as the q past squared innovations. That is: p

q

i =1

j =1

σ t2 = L + ∑ βi σ t2− i + ∑ α j ε t2− j , where L denotes the long-run volatility. By accounting for the information in the lag(s) of the conditional variance in addition to the lagged t−i terms, the GARCH model reduces the number of parameters required. In most cases, one lag for each variable is sufficient. The GARCH(1,1) model is given by: σ t2 = L + β 1σ t2−1 + α 1 ε t2−1. GARCH can successfully capture thick-tailed returns and volatility clustering. It can also be modified to allow for several other stylized facts of asset returns. EGARCH The Exponential Generalized Autoregressive Conditional Heteroskedasticity (EGARCH) model introduced by Nelson (1991)17 builds in a directional effect of price moves on conditional variance. Large price declines, for instance may have a larger impact on volatility than large price increases. The general EGARCH(p,q) model for the conditional variance of the innovations, with leverage terms and an explicit probability distribution assumption, is: p q  |ε t − j | |ε t − j |  q  εt−j  −E logσ t2 = L + ∑ β i logσ t2− i + ∑ α j    + ∑ Lj  i =1 j =1  σ t − j  σ t − j  j =1  σ t − j 

2 |ε t − j |  for the normal distribution, and where, E {| zt − j |}E  = π  σ t − j   v −1 Γ |ε t − j |  v − 2  2  for the Student’s t distribution with degree E {| zt − j |}E  = π v  σ t − j  Γ  2 of freedom ν > 2.

Volatility Forecasting of the Crude Oil Market

203

GJR(p,q) GJR( p,q) model is an extension of an equivalent GARCH( p,q) model with zero leverage terms. Thus, estimation of initial parameter for GJR models should be identical to those of GARCH models. The difference is the additional assumption with all leverage terms being zero: p

q

q

i =1

j =1

j =1

σ t2 = L + ∑ βi σ t2− i + ∑ α j ε t2− j + ∑ Lj St − j ε t2− j where St–j = 1 if εt–j < 0, St–j = 0 otherwise, with constraints p

q

∑β + ∑α i

i =1

j

+

j =1

1 q ∑ Lj < 1 2 j =1

L ≥ 0, βi ≥ 0, α j ≥ 0, α j + Lj ≥ 0. Regime-switching models Markov regime-switching models have been applied in various fields such as oil and the macroeconomy analysis,18 analysis of business cycles (Hamilton, 198919) and modeling stock market and asset returns (Vo, 2009). We now consider a dynamic volatility model with regime-switching. Suppose a time series yt follows an AR ( p) model with AR coefficients, together with the mean and variance, depending on the regime indicator st: p

yt = µ st + ∑ ϕ j ,st yt − j + ε t , «t ~ i.i.d. Normal (0, sst2 ) j =1

The corresponding density function for yt is: f ( yt | st ,Yt −1 ) =

 ω2  ⋅ exp  − t 2  = f ( yt | st , yt −1 ,..., yt − p , 2πσ s2t  2σ st  1

p

where ω t = yt − ω st − ∑ ϕ j , st yt − j . j =1

The model can be estimated by use of maximum log likelihood estimation. A more practical approach is to allow the density function of yt to depend on not only the current value of the regime indicator st but also on past values of the regime indicator st, which means the density function should take the form of: f(yt | St, St–1, Yt–1) where St−1 = {st−1, st−2, ... } is the set of all the past information on st.

204

Enterprise Risk Management in Finance

Data The data spans a continuous sequence of 866 days from February 2006 to July 2009, showing the closing prices of the NYMEX Crude Oil index during this time period on a day-to-day basis. Weekends and holidays are not included in our data, thus those days are considered as without price movement. Using the logarithm of price changes means that our continuously compounded return is symmetric, preventing us from having nonstationary oil price levels that would affect our return volatility. Table 19.1 presents the descriptive statistics of daily crude oil price changes. In Figure 19.1 we show a plot of crude oil daily price movement. To get a preliminary view of volatility change, Table 19.2 shows descriptive statistics for the logreturn of the Daily Crude Oil Index ranging over the period February 2006 to July 2009. The corresponding plot is given in Figure 19.2. Distribution analysis Figure 19.3 displays a distribution analysis of our data ranging from February 2006 up to July 2009. The data is the log return of the daily crude oil price Table 19.1 Statistics on the daily crude oil index changes, February 2006–July 2009 Statistics

Value

Sample Size Mean Maximum Minimum Standard deviation Skewness Kurtosis

866 77.2329 145.9600 44.4100 20.9270 1.3949 4.3800

160 140 120 100 80 60 40

Figure 19.1

0

100

200

300

400

500

NYMEX crude oil daily price movements

600

700

800

900

Volatility Forecasting of the Crude Oil Market

Table 19.2 Daily crude oil index logreturn statistics, February 2006–July 2009 Statistics

Value 865 1.8293e-005 0.1003 0.0874 0.0218 −0.0962 6.1161

Sample Size Mean Maximum Minimum Standard Deviation Skewness Kurtosis

0.15 0.1 0.05 0 –0.05 –0.1 Figure 19.2

0

100

200

300

400

500

700

800

900

NYMEX crude oil daily logreturn

25

r data Normal distribution T location-scale distribution

20 Density

600

15 10 5 0 –0.08 –0.06 –0.04 –0.02

0

0.02

Data Figure 19.3

Normal distribution vs. t-distribution

0.04

0.06

0.08

0.1

205

206 Enterprise Risk Management in Finance

movements over the time period mentioned above. We can see that the best distribution for our data is a t-Distribution, shown by the blue line in Figure 19.3. The red line represents the normal distribution of our data. So a conditional t-Distribution is preferred to normal distribution in our research. An augmented Dickey–Fuller univariate unit root test yields a p -value of 1.0*e-003, 1.1*e-003 and 1.1*e-003 for lags of 0, 1 and 2 respectively. All p -values are smaller than 0.05, which indicates that the time series has a property of trend-stationary.

Results GARCH modeling We first estimated the parameter of the GARCH(1,1) model using 865 observations in Matlab, and then tried various GARCH models using different probability distributions with the maximum likelihood estimation technique. In many financial time series, the standardized residuals zt = εt / st usually display high levels of kurtosis, which suggests departure from conditional normality. In such cases, the fat-tailed distribution of the innovations driving a dynamic volatility process can be better modeled using the Student’s −t or the Generalized Error Distribution (GED). Taking the square root of the conditional variance and expressing it as an annualized percentage yields a time-varying volatility estimate. A single estimated model can be used to construct forecasts of volatility over any time horizon. Table 19.3 presents the GARCH(1,1) estimation using the t-distribution. The conditional mean process is modeled by use of ARMAX(0,0,0). Substituting these estimated values in the math model, we yield the explicit form as follows: yt = 6.819e − 4 + «t s t2 = 2.216e − 6 + 0.9146 s t2−1 + 0.0815 «t2−1.

Table 19.3 Model

GARCH(1,1) estimation using the t-distribution AIC

BIC

lnL

Parameter

C Mean: K ARMAX −4559.9 −4536.1 2284.97 b1 (0,0,0); a1 Variance: DoF GARCH(1,1)

Value 6.819e−4 2.216e−6 0.915 0.082 34.603

Std error t-Statistic 5.045e−4 1.306e−6 0.017 0.018 8.442e−7

1.352 1.701 52.651 4.554 4.099e+7

Volatility Forecasting of the Crude Oil Market

Innovations

0.1 Innovation

207

0.05 0 –0.05 –0.1

Standard Deviation

0

200

0.15 0.1 0.05 0 –0.05 –0.1

300

400

500

600

700

800

900

600

700

800

900

600

700

800

900

Conditional Standard Deviations

0.05 0.04 0.03 0.02 0.01 0 0

Return

100

100

200

300

400

500

Returns

0

Figure 19.4

100

200

300

400

500

Innovation, standard deviation, return

Figure 19.4 depicts the dynamics of the innovation, standard deviation, and return, using the above estimated GARCH model, that is, the ARMAX(0,0,0) GARCH(1,1) with a log likelihood value of 2284.97. We want to find a higher log likelihood value for other GARCH modeling, so we use the same data with different models in order to increase the robustness of our model. We now try different combinations of ARMAX and GARCH, EGARCH and GJR models. Computational results are presented in Table 19.4. A general rule for model selection is that we should specify the smallest, simplest models that adequately describe data, because simple models are easier to estimate, easier to forecast, and easier to analyze. Model selection criteria such as AIC and BIC penalize models for their complexity when considering best distributions that fit the data. Therefore, we can use log likelihood (LLC), Akaike (AIC) and Bayesian (BIC) information criteria to compare alternative models. Usually, differences in LLC across distributions cannot be compared since distribution functions can have different capabilities for fitting random data, but we can use the minimum AIC and BIC, maximum LLC values as model selection criteria.20 As can be seen from Table 19.4, the log likelihood value of ARMAX(1,1,0) GJR(2,1) yields the highest log likelihood value 2292.32 and the lowest AIC value −4566.6 among all modeling techniques. Thus we select GJR models. The ARMAX(1,1,0) GJR(2,1) model was used in a simulation and a forecast for the standard deviation of a 30-day period using 20,000 realizations.

208 Enterprise Risk Management in Finance

Table 19.4 Various GARCH modeling characteristics Model

AIC

BIC

lnL

Parameter

Value

Std error

t-Statistic

C φ1 Mean: θ1 ARMAX(1,1,0); −4561.0 −4527.7 2287.5 K Variance: β1 GARCH(1,1) α1 DoF

8.995e−4 −0.312 0.236 2.056e−6 0.9175 0.0790 30.107

6.685e−4 0.439 0.447 1.257e−6 0.017 0.017 1.677e−4

1.346 −0.711 0.529 1.636 54.161 4.544 1.795e+5

C φ1 Mean: θ1 ARMAX(1,1,0); K −4557.8 −4524.5 2286.3 Variance: β1 EGARCH(1,1) α1 L1 DoF

6.656e−4 6.237e−4 1.067 −0.307 0.390 −0.787 0.223 0.397 0.561 −0.040 0.030 −1.334 0.995 3.626e−3 274.455 0.146 0.028 5.198 −0.032 0.016 −2.034 37.596 48.455 0.776

C φ1 Mean: θ1 ARMAX(1,1,0); K −4560.9 −4522.8 2288.4 Variance: β1 GJR(1,1) α1 L1 DoF

6.912 e−4 −0.297 0.222 2.151e−6 0.919 0.059 0.034 38.36

6.392e−4 0.450 0.457 1.268e−6 0.017 0.021 0.025 1.197e−4

1.081 −0.660 0.485 1.696 54.718 2.878 1.354 3.205e+4

C φ1 θ1 Mean: K ARMAX(1,1,0); −4566.6 −4523.8 2292.3 β1 Variance: β2 GJR(2,1) α1 L1 DoF

5.647e−4 −0.358 0.284 3.504e−6 0 0.868 0.091 0.068 50.013

6.464e−4 0.403 0.414 1.994e−6 0.026 0.029 0.026 0.035 6.069e−6

0.874 −0.889 0.687 1.757 0.000 29.559 3.571 1.955 8.241e+6

The forecasting horizon was defined to be 30 days (one month). The simulation used 20,000 outcomes over a 30-day period based on our fitted model ARAMX(1,1,0) GJR(2,1) with a horizon of 30 days from ‘Forecasting.’ Figure 19.5 compares forecasts from ‘Forecasting’ with those derived from ‘Simulation.’ The first four panels of Figure 19.5 directly compare each of the forecasted outputs with the corresponding statistical result obtained from simulation. The last two panels of Figure 19.5 illustrate histograms from which we could compute the approximate probability density functions and empirical confidence bounds. When comparing forecasting with its counterpart derived from

209

Volatility Forecasting of the Crude Oil Market

Forecast of STD of cumulative holding period Returns

0.12

0.024 Forecast of STD of residuals

0.11 Standard Deviation

0.09 0.08 0.07 0.06 Forecast results Simulation results

0.05 0.04

Standard Deviation

0.0235

0.1

0.023 0.0225 0.022 Forecast results Simulation results

0.0215 0.021

0.03

0.0205

0.02 0

5

10

15

20

25

30

0

5

Forecast period ×10 15

–4

25

30

Forecast of Returns

Forecast results Simulation results

0.0235

Forecast results Simulation results

0.023

10

0.0225 Return

Standard Deviation

20

15

Forecast period

Standard error of forecast of returns

0.024

10

0.022

5

0.0215 0.021

0

0.021 0.0205

–5 0

5

10

20

15

25

0

30

Forecast period

10

15

20

25

30

Forecast period Simulated Returns at Forecast Horizon

Cumalative Holding Period Returns at Forecast Horizon

3000

5

4500 4000

2500

3500 3000 2500 Count

Count

2000 1500 1000

2000 1500 1000

500

500 0 –0.8

–0.6

–0.4

–0.2

0

0.2

0.4

Return

Figure 19.5

0.6

0 –0.2 –0.15 –0.1 –0.05

0

0.05

0.1

0.15

Return

Simulation and forecasting

the Monte Carlo simulation, we show computation for four parameters in the first four panels of Figure 19.5: the conditional standard deviations of future innovations, the MMSE forecasts of the conditional mean of the Nasdaq return series, cumulative holding period returns and the root mean square errors (RMSE) of the forecasted returns. The fifth panel of Figure 19.5 uses a histogram to illustrate the distribution of the cumulative holding period return obtained

210

Enterprise Risk Management in Finance

if an asset was held for the full 30-day forecast horizon. In other words, we plot the logreturn obtained by investing in NYMEX Crude Oil Index today, and sold after 30 days. The last panel of Figure 19.5 uses a histogram to illustrate the distribution of the single-period return at the forecast horizon, that is, the return of the same mutual fund, the 30th day from now.

Markov regime-switching modeling We now use a Markov regime-switching model. The purpose is twofold: first, to see if Markov switching regressions can beat GARCH models in time series modeling; second, to find turmoil regimes in historical time series. Table 19.5 illustrates the results. The model in Table 19.5 assumes Normal distribution and allows all parameters to switch. We used S = [1 1 1] to control switching dynamics, where the first elements of S control the switching dynamic of the mean equation, while the last terms control the switching dynamic of the residual vector, including distribution parameters mean and variance. A value of 1 in S indicates that switching is allowed in the model while a value of 0 in S indicates that parameter is not allowed to change states. Then the model for the mean equation is: State 1 (St = 1)

State 2(St = 2 )

yt = −0.0015 − 0.0667yt −1 + ε t

yt = 0.0012 − 0.0934yt −1 + ε t

ε t ∼ N (0,0.03062 )

ε t ∼ N (0,0.01152 ),

where εt is a residual vector which follows a particular distribution. The tran⎡0.99 0.01⎤ sition matrix, P = ⎢ ⎥ , controls the probability of a regime switch ⎣ 0.01 0.99⎦ from state 1(2) (column 1(2)) to state 2(1) (row 2(1)). The sum of each column in P is equal to 1, since they represent full probabilities of the process for each state. In order to obtain the best fitted Markov regime-switching models, we tried various parameter settings for the traditional Hamilton model and complicated settings using t-distribution and Generalized Error Distribution. Computational results are given in Tables 19.6, 19.7 and 19.8. A comparison of log Likelihood values indicate that complicated setting using t-distribution and Generalized Error Distribution usually are preferred. The best fitted Markov regime-switching models should assume GED and allow all parameters to change states (see Table 19.8). We now focus on analysis using the best fitted Markov regime-switching model, that is, ‘MS model, S = [1 1 1 1 1] (GED)’ in Table 19.8. Figure 19.6

Volatility Forecasting of the Crude Oil Market

Table 19.5

Markov regime-switching computation example

Model log (Distribution likeassumption) lihood MS Model, S = [1 1 1] (Normal)

211

Switching parameters state 1 state 2

Transition probabilities matrix

Model’s STD 0.0306 0.0115 Indep column 1 −0.0015 0.0012 Indep column 2 −0.0667 −0.0934

0.99 0.01 0.01 0.99

Nonswitching parameters

2257.36 N/A

Table 19.6 Markov regime-switching using Hamilton’s (1989) model

log Model (Distribution likeassumption) lihood The Hamilton (1989) model, 2212.38 S = [1 0 1] (t)

The Hamilton (1989) model, 2257.34 S = [1 1 1] (t)

Nonswitching parameters

0.0135

N/A

Switching parameters state 1 state 2 Degrees of Freedom (t dist) Indep column 1

100.00

0.0008

Model’s STD 0.0264 7.8238 Degrees of Freedom (t dist) −0.0012 Indep column

Transition probabilities matrix

1.5463

−0.0002

1.00 0.00 0.00 1.00

0.0113 112.3094

0.0010

0.99 0.01 0.01 0.99

presents transitional probabilities in Markov regime-switching with GED: fitted state probabilities and smoothed state probabilities. Based on such a transitional probability figure, we can classify historical data into two types according to their historical states. Figure 19.7 depicts the logreturn of two regimes in historical time series. Figure 19.8 depicts the price of two regimes in historical time series. As can be seen from Figures 19.7 and 19.8, the total historical time series are divided into two regimes: a normal one with small change (state 2) and a turmoil regime with high risk (state 1). For each state, the regime-switching model identifies three periods of data. The normal regime includes two periods: February 10, 2006~December 11, 2006, and January 30, 2007~October 14, 2007. The turmoil regime also includes two periods: December 12, 2006~January 29, 2007, and October 15, 2007~July 7, 2009. The first turmoil lasted only one and a half months, but the second one covered almost the total financial crisis.

212

Enterprise Risk Management in Finance

Table 19.7

Markov regime-switching using t-distribution

Model (Distribution log like- Non-switching assumption) lihood parameters

Transition probabilities matrix

Switching parameters state 1 state 2

Indep 0.0021 column 1 Indep −0.3925 2172.41 Degrees of 2.9506 column 2 Freedom (t dist)

−0.0010

MS Model, S=[1 1 0 0] (t)

0.0117

MS Model, S=[1 1 1 1] (t)

0.0130 Model’s STD Degrees of 3.2408 freedom (t dist) 0.0013 Indep column 1 −0.2015 Indep column 2

STD

2174.86

MS Model, S=[1 1 1 1 1] (t)

Table 19.8

2260.95

0.0128

N/A

N/A

0.2553

0.45 0.57 0.55 0.43

2.3637

−0.0034

0.80 0.98 0.20 0.02

0.9080

Model’s 0.0262 0.0113 STD Degrees of 7.4904 100.000 freedom (t dist) −0.0012 0.0011 Indep column 1 −0.0736 −0.0915 Indep column 2 −0.0121 0.0422 Indep column 3

0.99 0.01 0.01 0.99

Markov regime-switching using GED

Model (Distribution assumption) MS Model, S=[1 1 1 1] (GED)

MS Model, S=[1 1 1 1 1] (GED)

log likelihood

Transition probabilities matrix

Switching parameters state 1 state 2

2172.16

Model‘s STD Value of k (GED dist) Indep column 1 Indep column 2

0.0029 1.4987 0.0020 0.8905

0.0094 0.8011 0.0013 0.2207

0.0203

0.0120

2263.06

Model’s standard deviation Value of k (GED dist) Indep column 1 Indep column 2 Indep column 3

0.7122 0.0014 0.0706 0.0287

0.4675 0.0010 0.0848 0.0384

0.06 0.26 0.94 0.74

0.99 0.01 0.01 0.99

213

Volatility Forecasting of the Crude Oil Market

(a)

(b) 1.4

1 State 1 State 2

0.8

Smoothed states probabilities

Filtered states probabilities

0.9

0.7 0.6 0.5 0.4 0.3 0.2 0.1

State 1 State 2

1.2 1 0.8 0.6 0.4 0.2 0

0 0

0

100 200 300 400 500 600 700 800 900

100 200 300 400 500 600 700 800 900 Time

Time

Transitional probabilities in Markov regime-switching with GED

Figure 19.6

20

2

6 00

-12

0

1 7-0

-12

State 1

-29

0 20

0 20

7-1

0-1

0 20

2

-2-

7-0

7

5

State 2

6 00

9-0

6-1

1 2-1 2

10 2

7 00

-01

-30

Figure 19.7 Returns of two regimes in historical time series

7 00

-10

-14

214

Enterprise Risk Management in Finance

State 1

2

2

6 00

-12

7 00

-12

-01

-29 2

0 20

7-1

0-1

2

2

Figure 19.8

-2-

-07

-07

5

State 2

6 00

9 00

6 00

-12

-11 2

10 0 20

7-0

1-3

7 00

-10

-14

0

Price of two regimes in historical time series

Conclusions We examined crude oil price volatility dynamics using daily data for the period February 13, 2006 up to July 21, 2009. We employed the GARCH, EGARCH and GJR models and various Markov regime-switching models using the maximum likelihood estimation technique to model volatility. Codes were written in Matlab language. We compared several parameter settings in all models. In GARCH models, the ARMAX (1,1,0)/GJR(2,1) yielded the best fitted result, with maximum log likelihood value of 2292.32 when assuming that our data followed a t-distribution. Markov regime-switching models generated a similar fitted result, but with a bit lower log likelihood value. Markov regimeswitching modeling gave interesting results in classifying historical data into two states: a normal period and a turmoil period. This can account for some market performance during the financial crisis.

20

Confucius Three-stage Learning of Risk Management

Introduction During my class of risk management, I teach a Confucius three-step approach of learning risk management to my students. Students seem to be well-motivated to learn this interdisciplinary course using my approach. Confucius, a philosopher born in 551 BCE during the Chou dynasty, may well be one of the most influential philosophers in history.1 He wandered all around China, trying to serve as an adviser to various rulers in the Spring–Autumn period. Confucius, his students and many other followers developed ‘Four Books’ in Confucianism. One of the most significant was ‘The Great Learning’, which addresses classical themes of Chinese philosophy and political issues, and has been extremely popular and very influential in both traditional and modern Chinese thinking. ‘Eight steps’ were developed showing how to ‘investigate things,’ which becomes one of the first stages to understanding ‘The Great Learning’.2 Three famous steps are expressed: Wishing to order well their States, they first regulated their families. Wishing to regulate their families, they first cultivated their persons. Their persons being cultivated, their families were regulated. Their families being regulated, their States were rightly governed. I borrow three words to describe these three stages: self-cultivation, family regulation, and state harmonization. I first taught this at the RiskLab executive course on risk management in 2008, later on courses offered to various Chinese universities such as the Management School at the University of the Chinese Academy of Sciences, the School of Economics and Management at Beihang University, and the School of Business at Central South University.

215

216

Enterprise Risk Management in Finance

Self-cultivation Let us start with an example of NETLIB Problem PILOT4 to illustrate use of self-cultivation in learning sophisticated financial risk management tools.3 Netlib has been a repository of linear programming (LP) problems available to test new codes and compare performance. This example is a LP with 1000 variables and 410 constraints, where the 372nd constraint is as follows. aT x ≡ −15.79081x826 − 8.598819x827 − 1.88789x828 − 1.362417x829 − 1.526049x830 − 0.031883x849 − 28.725555x850 − 10.792065x851 − 0.19004 x852 − 2.757176 x853 − 12.290832 x854 + 717.562256 x855 − 0.057865x856 − 3.785417x857 − 78.30661x858 − 122.163055x859 − 6.46609x860 − 0.48371x861 − 0.615264 x862 − 1.353783x863 − 84.644257x864 − 122.459045x865 − 43.15593x866 − 1.712592 x870 − 0.401597x871 + x880 − 0.946049x898 − 0.946049x916 ≥

b ≡ 23.387405

Assuming that one investor employed a naïve investment strategy to maximize return of portfolio, this financial engineering problem structure turns out to be very similar to a Netlib problem. We observe that most coefficients in the above constraint look ugly, with an accuracy of 5–6 digits. It is natural to believe that coefficients of this type characterize certain market uncertainties and risks. Beyond operations research, probability theory is fundamental in financial risk management development. Probability was not established until 1600 (although fashionable in 1700), and according to a lecturer by Professor Robert Shiller was probably first used by William Shakespeare, who wrote a story about a young lady, describing a man she likes, saying: ‘I like him very much. I find him very probable’. Another interesting concept related to Shakespeare is the Infinite Monkey Theorem, which states that if enough monkeys hit keys at random on a typewriter keyboard for an infinite amount of time, one will almost surely type a complete works of William Shakespeare. The infinite monkey theorem is obviously one popular application of the law of very large numbers from probability theory.4 Here I would like to summarize some fundamental math theories, especially probability theory, useful in risk management. To do this, I need a basic classification of risk properties. Risk, measured by volatility in a security market, is quoted by unit measures in terms of basis points. That explains to some extent the relevance of having properties for risks as an object. I use four levels of classifications: Uncertainty of risks, Dynamics of risks, Clustering-dependenceinterconnection of risks and Complexity of risks. Basic probability theory, to include various distribution and density functions, is used to characterize

Confucius Three-stage Learning of Risk Management

217

the first property of risks. This was developed in the 1600s and popularized in 1700s with representative scholars such as Bernoulli, de Moivre, Laplace, Poisson, Gauss, and Pareto. The second type of risk property uses various stochastic modeling tools developed intensively since the 1930s by well-known scholars such as Lévy, Khintchine, Kolmogorov, and Doob. The third property of risks is new research problems in finance that have been developed since the 1950s by well-known scholars such as Fréchet and Sklar, who develop fundamental theories that are useful for derivative pricing. Complexity of risks can be described by complexity science theory developed since the 1960s.5

Family regulation I am going to tell a story extracted from a cartoon picture in the book The Cartoon Introduction to Economics.6 I give the story a name: a story of family risk management. This is a family of three people: the child, the mommy and the daddy. On Monday morning, the family members all have questions about what they are going to do. The child is going to school and raises a question: is it going to rain today? The mom is planning to buy a second-hand BMW car and raises a question: is this used BMW a lemon or a peach? The daddy is reading news on the stock market and raises the question: am I going to buy stock in Facebook or mutual funds Comfort? The child’s question reflects his attitude of risk-aversion to today’s weather. This is the core theme in expected utility theory originated from classical economics in Adam Smith’s time. The daddy’s question indicates his diversification strategy for investment in risky assets: a portfolio optimization strategy might be preferred. There are milestone works in financial risk theory: Harry Markowitz presented mean-variance framework in a 1952 paper and a 1959 book7 on how to find the best possible diversification strategy, and William Sharpe’s capital asset pricing model (CAPM) in 19648 provided tools to determine a theoretically appropriate required rate of return of an asset, if that asset is to be added to an already well-diversified portfolio, given that asset’s non-diversifiable risk. James Tobin expanded on Markowitz’s work by adding a risk-free asset to the analysis in 1958,9 leading to the development of a super-efficient portfolio and the capital market line. All these works won a Nobel Prize, although they include strong assumptions that are opposed, such as perfect capital markets, or log-normally distributed market data. The New York Times ran a story when James Tobin died in 2002: After he won the Nobel Prize, reporters asked him to explain the portfolio theory. When he tried to do so, one journalist interrupted, ‘Oh, no, please explain it in lay language.’ So he described the theory of diversification by

218 Enterprise Risk Management in Finance

saying: ‘You know, don’t put all your eggs in one basket.’ Headline writers around the world the next day created some version of ‘Economist Wins Nobel for Saying, “Don’t Put Eggs in One Basket.”’ The mommy’s question gives a key risk management problem in economics: adverse selection. The mommy is checking a used car with a relatively high price but maybe reasonably good quality. Because of information asymmetry, she is not sure of the quality, so decides to buy another used car with lower price. This continues until a dealer with a low-priced used car can sell her a particular car. The result is that owners of good cars will not place their cars on the used car market but low-priced used cars with low quality prevail. So adverse selection works against social welfare and refers to a market process where undesired results occur when market players have asymmetric information. This is another very important piece that helped the economist George Akerlof to win the Nobel Memorial Prize in Economic Sciences in 2001.10 The story of family risk management demonstrates important risk problems and theories from various disciplines. From a simple family activity, we see that risk theory is embedded in many other theories including expected utility theory in decision sciences and microeconomics, portfolio optimization theory and CAPM theory in modern finance, adverse selection, asymmetric information, and contract theory in economics, especially in industry organization theory.

State harmonization After the self-cultivation and family-regulation steps of learning risk management tools, you probably are armed with fundamental knowledge that can assist you to treat real-world risk management problems at a state-harmonization level. One typical example is the systemic risks that to some extent result in the recent economic crisis that originated from the secondary mortgage market. Systemic risk refers to the potential collapse of the entire financial system, as a result of risk associated with one individual entity, group or component, leading to a cascading failure of the entire system. These are risks imposed by inter-linkages and interdependencies in a system or market. These interdependencies and the potential ‘clustering’ of institution failure are important stateharmonization issues which policy-makers have to consider in order to protect a system or system of systems against systemic risk. Government agencies such as the US Fed and SEC (Securities and Exchange Commission) or central banks usually have rules for safeguarding the trading interests of the market as a whole, claiming that the investors in markets are exposure to dependencies of risks arising from their inter-linkage.11

Confucius Three-stage Learning of Risk Management

219

In Table 20.1 we map previous chapters to different levels of learning risk management.

Table 20.1 Risk management links Chapter 1 – Overview

Enterprise Risk Management

Family regulation

Chapter 2 – Practice

Enron

State harmonization

Chapter 3 – Overview

Financial Risk Management

Family regulation

Chapter 4 – Practice

The Real Estate Crash of 2008

State harmonization

Chapter 5 – Theory

Financial Risk Forecast Using Machine Learning and Sentiment Analysis

Self-cultivation

Chapter 6 – Theory

On-line Stock Forum Sentiment Analysis

Self-cultivation

Chapter 7 – Theory

DEA Risk Scoring Model of Internet Stocks

Self-cultivation

Chapter 8 – Theory

Bank Credit Scoring

Self-cultivation

Chapter 9 – Theory

Credit Scoring using Multiobjective Data Mining

Self-cultivation

Chapter 10 – Theory

Performance Evaluation and Risk Analysis of Online Banking

Self-cultivation

Chapter 11 – Overview

Economic Perspective

Family regulation

Chapter 12 – Practice

British Petroleum Deepwater Horizon

State harmonization

Chapter 13– Theory

Bank Efficiency Analysis

Self-cultivation

Chapter 14 – Theory

Catastrophe Bond and Risk Modeling

Self-cultivation

Chapter 15 – Theory

Bilevel Programming Merger Analysis in Banking

Self-cultivation

Chapter 16 – Overview

Sustainability and Risk in Globalization

Family regulation

Chapter 17 – Practice

Risk from Natural Disaster

State harmonization

Chapter 18 – Theory

Pricing of Carbon Emission Exchange in the EU ETS

Self-cultivation

Chapter 19 – Theory

Volatility Forecasting of the Crude Oil Market

Self-cultivation

Conclusions This chapter presents a Confucian three-step process of learning risk management: self-cultivation, family regulation, and state harmonization.

220

Enterprise Risk Management in Finance

The self-cultivation perspective is the root and the first stage of learning risk management. Fundamental math theories that are useful in risk management are summarized. A basic classification of risk properties is given by Uncertainty of risks, Dynamics of risks, Clustering/Dependence/interconnection of risks and Complexity of risks. For each property, I gave a related math theory that may be useful in treating risk problems. A story of family risk management was presented to show the second stage of learning risk management. This perspective demonstrates that risk theory is embedded in many other theories, including expected utility theory in decision sciences and microeconomics, portfolio optimization theory and CAPM theory in modern finance, adverse selection, asymmetric information, and contract theory in economics, especially in industry organization theory. Armed with the self-cultivation stage tool and family-regulation stage knowledge, one might be ready for handling real-world risk management problems at a state-harmonization level. I gave systemic risk as a typical problem example for treating economic crisis originated from the secondary mortgage market.

Notes 1

Enterprise Risk Management

1. D.W. Hubbard (2009) The Failure of Risk Management: Why It ’s Broken and How to Fix It. John Wiley & Sons. 2. H.N. Higgins (2012) ‘Learning internal controls from a fraud case at Bank of China,’ Issues in Accounting Education, 27(4): 1171–1192. 3. B. Ballou, D.L. Heitger (2005) ‘A building-block approach for implementing COSO’s enterprise risk management–integrated framework,’ Management Accounting Quarterly, 6(2): 1–10. 4. D. Williamson (2007) ‘The COSO ERM framework: A critique from systems theory of management control,’ International Journal of Risk Assessment and Management, 7(8): 1089–1119. 5. F. Caron, J. Vanthienen, B. Baesens (2013) ‘A comprehensive investigation of the applicability of process mining techniques for enterprise risk management,’ Computers in Industry, 64, 464–475. 6. L. Rittenberg, F. Martens (2012) Enterprise Risk Management: Understanding and Communicating Risk Appetite. COSO. 7. The Association of Risk Managers (2010) A Structured Approach to Enterprise Risk Management (ERM) and the Requirements of ISO 31000. COSO. 8. B.M. Bowling, L. Rieger (2005) ‘Success factors for implementing enterprise risk management,’ Bank Accounting and Finance, 18(3): 21–26.

2

Enron

1. G. Ailon (2012) ‘The discursive management of financial risk scandals: the case of Wall Street Journal commentaries on LTCM and Enron,’ Qualitative Sociology, 35: 251–270. 2. J.E. Stiglitz (2003) The Roaring Nineties: A New History of the World’s Most Prosperous Decade. W.W. Norton & Co. 3. L. Fox (2003) Enron: The Rise and Fall. Wiley. 4. C. Hurt (2014) ‘The duty to manage risk,’ The Journal of Corporate Law, 39(2): 153–267. 5. Ailon (2012), op cit. 6. C. Hollingsworth (2012) ‘Risk management in the post-SOX era,’ International Journal of Auditing, 16: 35–53. 7. H.N. Butler, L.E. Ribstein (2006) The Sarbanes–Oxley Debacle: What We’ve Learned; How to Fix It, AEI. 8. A. Dey (2010) ‘The chilling effect of Sarbanes–Oxley: a discussion of Sarbanes–Oxley and corporate risk-taking,’ Journal of Accounting and Economics, 49(1–2), 53–57. 9. J.D. Piotroski, S. Srinivasan (2008) ‘Regulation and bonding: the Sarbanes–Oxley Act and the flow of international listings,’ Journal of Accounting Research, 46(2): 383–425. 10. L. Bargeron, K. Lehn, C. Zutter (2009) ‘Sarbanes–Oxley and corporate risk-taking,’ Journal of Accounting and Economics, 49(1–2): 34–52.

221

222

3

Notes

Financial Risk Management

1. R. Lowenstein (2000) When Genius Failed: The Rise and Fall of Long-Term Capital Management, Random House. 2. H.M. Markowitz (1952) ‘Portfolio selection,’ The Journal of Finance, 17(1): 77–91. 3. W.F. Sharpe (1964) ‘Capital asset prices: a theory of market equilibrium under conditions of risk,’ The Journal of Finance, 19(3): 425–442. 4. F. Black, M. Scholes (1972) ‘The valuation of option contracts and a test of market efficiency,’ The Journal of Finance, 27(2): 399–417. 5. G.J. Alexander, A.M. Baptista (2004) ‘A comparison of VaR and CVaR constraints on portfolio selection with the mean-variance model,’ Management Science, 50(9): 1261– 1273; V. Chavez-Demoulin, P. Embreechts, J. Nešlehová (2006) ‘Quantitative models for operational risk: extremes, dependence and aggregation,’ Journal of Banking & Finance, 30: 399–417. 6. J. Von Neumann, O. Morgenstern (1944) Theory of Games and Economic Behaviour, 2nd ed. Princeton University Press. 7. M. Friedman, L.J. Savage (1948) ‘The utility analysis of choices involving risk,’ The Journal of Political Economy, 56(4): 279–304. 8. S.G. Mandis (2013) What Happened to Goldman Sachs: An Insider ’s Story of Organizational Drift and Its Unintended Consequences. Harvard Business Review Press. 9. N.N. Taleb (2012) Antifragile: Things That Gain from Disorder. Random House. 10. D.X. Li (2000) ‘On default correlation: a copula approach,’ Journal of Fixed Income, 9(4): 43–54. 11. F. Salmon (2009) ‘Recipe for disaster: the formula that killed Wall Street,’ Wired, 17(3).

4

The Real Estate Crash of 2008

1. C.M. Reinhart, K.S. Rogoff (2008) ‘Is the 2007 Subprime Crisis so different? An international historical comparison,’ American Economic Review, 98(2), 339–344. 2. G. Cooper (2008) The Origin of Financial Crises: Central Banks, Credit Bubbles and the Efficient Market Fallacy. Vintage Books. 3. N. Dunbar (1999) Investing Money: The Story of Long-Term Capital Management and the Legends Behind It. Wiley; R. Lowenstein (2000) When Genius Failed: The Rise and Fall of Long-Term Capital Management. Random House. 4. B. Cohen (1997) The Edge of Chaos: Financial Booms, Bubbles, Crashes and Chaos. John Wiley & Sons, Ltd. 5. A.S. Blinder (2013) After the Music Stopped: The Financial Crisis, the Response, and the Work Ahead. The Penguin Press. 6. L. Laeven, F. Valencia (2010) ‘Resolution of banking crises: the good, the bad, and the ugly,’ IMF Working Paper WP/10/146. 7. J. Taylor (2009) Getting Off Track: How Government Actions and Interventions Caused, Prolonged, and Worsened the Financial Crisis. Hoover Press; B. Keys, T. Mukherjee, A. Seru, V. Vig (2010) ‘Did securitization lead to lax screening? Evidence from subprime loans,’ Quarterly Journal of Economics, 125, 307–362. 8. G. Gorton (2008) ‘The panic of 2007,’ NBER Working Paper No. 14358; M. Brunnermeier (2009) ‘Deciphering the liquidity and credit crunch 2007–2008,’ Journal of Economic Perspectives, 23, 77–100; G. Dell’Arriccia, L. Laeven, D. Igan (2008) ‘Credit booms and lending standards: evidence from the subprime mortgage market,’ IMF Working Paper 08/106.

Notes 223

9. F.B. Wiseman (2013) Some Financial History Worth Reading: A Look at Credit, Real Estate, Investment Bubbles & Scams, and Global Economic Superpowers. Abcor Publishers. 10. R. Boyd (2011) Fatal Risk: A Cautionary Tale of AIG’s Corporate Suicide. Wiley. 11. S. Patterson (2010) The Quants: How a New Breed of Math Whizzes Conquered Wall Street and Nearly Destroyed It. Crown Business. 12. V. Sampath (2009) ‘The need for greater focus on nontraditional risks: the case of Northern Rock,’ Journal of Risk Management in Financial Institutions, 2(3): 301–305. 13. H.S. Shin (2009) ‘Reflections on Northern Rock: the bank run that heralded the global financial crisis,’ Journal of Economic Perspectives, 23(1): 101–119. 14. P. Goldsmith-Pinkham, T. Yorulmazer (2010) ‘Liquidity, bank runs, and bailouts: Spillover effects during the Northern Rock episode,’ Journal of Financial Service Research, 37(2/3): 83–98. 15. R. Shelp, A. Ehrbar (2009) Fallen Giant: The Amazing Story of Hank Greenberg and the History of AIG. Wiley. 16. Ibid. 17. Ibid. 18. J.F. Egginton, J.I. Hilliard, A.P. Liebenberg, I.A. Liebenberg (2010) ‘What effect did AIG’s bailout, and the preceding events, have on its competitors?’ Risk Management and Insurance Review, 13(2): 225–249. 19. J. Hobbs (2011) ‘Financial derivatives, the mismanagement of risk and the case of AIG,’ CPCU eJournal, 64(7): 1–8. 20. P.M. Linsley, R.E. Slack (2013) ‘Crisis management and an ethic of care: The case of Northern Rock Bank,’ Journal of Business Ethics, 113(2): 285–295.

5 Financial Risk Forecast Using Machine Learning and Sentiment Analysis 1. D. Dong, Q. Dong (2003) ‘HowNet – a hybrid language and knowledge resource,’ Proceedings of 2003 International Conference on Natural Language Processing and Knowledge Engineering, 820–824, October 26–29. 2. T. Bollerslev (1986) ‘Generalized autoregressive conditional heteroskedasticity,’ Journal of Econometrics, 31(3): 307–327. 3. B. Freisleben, K. Ripper (1997) ‘Volatility estimation with a neural network,’ Proceedings of the IEEE/IAFE on Computational Intelligence for Financial Engineering, 177–181, March 24–25; C. Burges (1998) ‘A tutorial on support vector machines for pattern recognition,’ Data Mining and Knowledge Discovery, 2(2): 121–167. 4. B. Freisleben and K. Ripper (1997) ‘Volatility estimation with a neural network,’ Proceedings of the IEEE/IAFE on Computational Intelligence for Financial Engineering, 177–181, March 24–25. 5. J.A.K. Suykens, T.V. Gestel, J.D. Brabanter et al. (2002) ‘Least squares support vector machines,’ Singapore :World Scientific Press; H.F. Wang, D.J. Hu (2005) ‘Comparison of SVM and LS-SVM for Regression,’ International Conference on Neural Networks and Brain 2005, 1, 279–283, October 13–15, 2005.

6

Online Stock Forum Sentiment Analysis

1. B. Watkins (2003) ‘Riding the wave of sentiment: an analysis of return consistency as a predictor of future returns,’ Journal of Behavioral Finance, 4(4): 191–200.

224 Notes

2. W. Antweiler, M. Frank (2004) ‘Is all that talk just noise? The information content of internet stock message boards,’ Journal of Finance, 59(3): 1259–1295. 3. R.F. Engle (1982) ‘Autoregressive conditional heteroscedasticity with estimates of variance of United Kingdom inflation,’ Econometrica, 50: 987–1008. 4. C. Burges (1998) ‘A tutorial on support vector machines for pattern recognition,’ Data Mining and Knowledge Discovery, 2(2): 121–167. 5. W. Chan (2003) ‘Stock price reaction to news and no-news: drift and reversal after headlines,’ Journal of Financial Economics, 70: 223–260. 6. M. Baker, J. Wurgler (2006) ‘Investor sentiment and the cross-section of stock returns,’ Journal of Finance, 61(4): 1645–1680. 7. J. Seigel (2002) Stocks for the Long Run. 3rd ed. McGraw-Hill. 8. D. Bathia, D. Bredin (2012) ‘An examination of investor sentiment effect on G7 stock market returns,’ European Journal of Finance, DOI:10.1080/1351847X.2011.636834.

7

DEA Risk Scoring Model of Internet Stocks

1. S.A. Ross, R.M. Westerfield, B.D. Jordan (2007) Corporate Finance Essentials. McGrawHill/Irwin. 2. G.J. Fielding, T.T. Babitsky, M.E. Brenner (1985) ‘Performance evaluation for bus transit,’ Transportation Research, 19A(1): 73–82; L.V. Utikin (2007) ‘Risk analysis under partial prior infermation and nonmonotone utility functions,’ International Journal of Information Technology and Decision Making, 6: 625–647. 3. C.-T. Ho and D.S. Zhu (2004) ‘Performance measurement of Taiwan’s commercial banks,’ International Journal of Productivity and Performance Management, 53(5): 425–434; D. Wu (2009) ‘Performance evaluation: an integrated method using data envelopment analysis and fuzzy preference relations,’ European Journal of Operational Research, 194(1): 227–235. 4. K. Cengiz, C. Ufuk and U. Ziya (2003) ‘Multi-criteria supplier selection using fuzzy AHP,’ Logistics Information Management, 16(6): 382–394; T.L. Saaty (2008) ‘Decision making with the analytic hierarchy process,’ International Journal of Services Sciences, 1(1): 83–98. 5. T.S. Felix and H.J. Chan (2003) ‘An innovative performance measurement method for supply chain management,’ Supply Chain Management: An International Journal, 8(3): 209–223. 6. C.-T. Ho (2006) ‘Measuring bank operations performance: an approach based on grey relation analysis,’ Journal of the Operational Research Society, 57, 227–349. 7. R.S. Kaplan and D.P. Norton (2006) Alignment: Using the Balanced Scorecard to Create Corporate Synergies. Harvard Business School Press Books. 8. P. Espahbodi (1991) ‘Identification of problem banks and binary choice models,’ Journal of Banking and Finance, 15, 53–71. 9. D. Wu (2006) ‘A note on DEA efficiency assessment using ideal point: an improvement of Wang and Luo’s model,’ Applied Mathematics and Computation, 2: 819–830. 10. M.J. Farrell (1957) ‘The measurement of productive efficiency,’ Journal of the Royal Statistical Society, 120, 253–281. 11. W. Charnes, A. Cooper, E. Rhodes (1978) ‘Measuring the efficiency of decision making units,’ European Journal of Operational Research, 2: 429–444. 12. R.D. Banker, A. Charnes, W. Cooper (1984) ‘Some models for estimating technical and scale inefficiencies in data envelopment analysis,’ Management Science, 30, 1078–1092.

Notes 225

13. P.S. Sudarsanam and R.J. Taffler (1995) ‘Financial ratio proportionality and intertemporal stability: an empirical analysis,’ Journal of Banking & Finance, 19(1): 45–60. 14. C.-T. Ho, D.S. Zhu (2004) ‘Performance measurement of Taiwan’s commercial banks,’ International Journal of Productivity and Performance Management, 53(5): 425–434. 15. S.N. Huang, T.L. Kao (2006) ‘Measuring managerial efficiency in non-life insurance companies: an application of two-stage data envelopment analysis,’ International Journal of Management, 23(3): 699–720. 16. L.M. Seiford, J. Zhu (1999) ‘Profitability and marketability of the top 55 U.S. commercial banks,’ Management Science, 45(9): 1270–1288; M. Gulser, M. Ilhan (2001) ‘Risk and return in the world’s major stock markets,’ Journal of Investing, (Spring): 62–67; X. Luo (2003) ‘Evaluating the profitability and marketability efficiency of large banks: an application of data envelopment analysis,’ Journal of Business Research, 56: 627–635; A. Barua, P.L. Brockett, W.W. Cooper, H. Deng, B.R. Parker, T.W. Ruefli, A. Whinston (2004) ‘Multi-factor performance measure model with an application to fortune 500 companies,’ Socio-Economic Planning Sciences, 38: 233–253; C.S. Carlos, F.C. Yolanda, M.M. Cecilio (2005) ‘Measuring DEA efficiency in internet companies,’ Decision Support Systems, 38: 557–573; D. Wu, Z. Yang, L. Liang (2006) ‘Using DEA-neural network approach to evaluate branch efficiency of a large Canadian bank,’ Expert Systems with Applications, 31(1): 108–115; S.F. Lo, W.M. Lu (2006) ‘Does size matter? finding the profitability and marketability benchmark of financial holding companies,’ Journal of Operational Research, 23(2): 229–246; H.C. Tsai, C.M. Chen, G.H. Tzeng (2006) ‘The comparative productivity efficiency for global telecoms,’ International Journal of Production Economics, 103: 509–526. 17. D. Wu (2006) ‘A note on DEA efficiency assessment using ideal point: an improvement of Wang and Luo’s model,’ Applied Mathematics and Computation, 2: 819–830; N. Ahmad, D. Berg, G.R. Simons (2006) ‘The integration of analytic hierarchy process and data envelopment analysis in a multi-criteria decision-making problem,’ International Journal of Information Technology and Decision Making, 5: 263–276. 18. J. Yao, Z. Li, K.W. Ng (2006) ‘Model risk in VaR estimation: an empirical study,’ International Journal of Information Technology and Decision Making, 5: 503–512. 19. Y. Shi, Y. Peng, G. Kou, Z. Chen (2006) ‘Classifying credit card accounts for business intelligence and decision making: a multiple-criteria quadratic programming approach,’ International Journal of Information Technology and Decision Making, 4: 1–19; S. Deng, Z. Xia (2006) ‘A real options approach for pricing electricity tolling agreements,’ International Journal of Information Technology and Decision Making, 5, 421–436.

8

Bank Credit Scoring

1. G. Dickinson (2001) ‘Enterprise risk management: its origins and conceptual foundation,’ The Geneva Papers on Risk and Insurance, 26(3): 360–366. 2. E.G. Baranoff (2004) ‘Risk management: a focus on a more holistic approach three years after September 11,’ Journal of Insurance Regulation, 22(4): 71–81. 3. M. Crouhy, D. Galai, R. Mark (1998) ‘Model risk,’ Journal of Financial Engineering, 7(3/4): 267–288, reprinted in Model Risk: Concepts, Calibration and Pricing (ed. R. Gibson), Risk Book, 2000, 17–31; M. Crouhy, D. Galai, R. Mark (2000) ‘A comparative analysis of current credit risk models,’ Journal of Banking & Finance, 24: 59–117.

226

4.

5.

6. 7.

8.

9.

10. 11.

12. 13.

14. 15.

9

Notes

G.J. Alexander, A.M. Baptista (2004) ‘A comparison of VaR and CVaR constraints on portfolio selection with the mean-variance model,’ Management Science, 50(9): 1261–1273; V. Chavez-Demoulin, P. Embrechts, J. Nešlehová (2006) ‘Quantitative models for operational risk: extremes, dependence and aggregation,’ Journal of Banking & Finance, 30: 2635–2658; R. Garcia, É. Renault, G. Tsafack (2007) ‘Proper conditioning for coherent VaR in portfolio management,’ Management Science, 53(3): 483–494; N. Taylor (2007) ‘A note on the importance of overnight information in risk management models,’ Journal of Banking & Finance, 31: 161–180. T. Jacobson, J. Lindé, K. Roszbach (2006) ‘Internal ratings systems, implied credit risk and the consistency of banks’ risk classification policies,’ Journal of Banking & Finance, 30, 1899–1926. H. Elsinger, A. Lehar, M. Summer (2006) ‘Risk assessment for banking systems,’ Management Science, 52(9): 1301–1314. R.S. Kaplan, D.P. Norton (1992) ‘The balanced scorecard – measures that drive performance,’ Harvard Business Review, 70(1): 71–79; and R.S. Kaplan, D.P. Norton (2006) Alignment: Using the Balanced Scorecard to Create Corporate Synergies. Harvard Business School Press Books. D. Bigio, R.L. Edgeman, T. Ferleman (2004) ‘Six sigma availability management of information technology in the office of the chief technology officer of Washington, DC.,’ Total Quality Management, 15(5–6): 679–687; S. Scandizzo (2005) ‘Risk mapping and key risk indicators in operational risk management,’ Economic Notes by Banca Monte dei Paschi di Siena SpA, 34(2): 231–256. A. Papalexandris, G. Ioannou, G. Prastacos, K.E. Soderquist (2005) ‘An integrated methodology for putting the balanced scorecard into action,’ European Management Journal, 23(2): 214–227. J. Calandro, Jr., S. Lane, S. (2006) ‘An introduction to the enterprise risk scorecard,’ Measuring Business Excellence, 10(3): 31–40. U. Anders, M. Sandstedt (2003) ‘An operational risk scorecard approach,’ Risk, 16(1): 47–50; H. Wagner (2004) ‘The use of credit scoring in the mortgage industry,’ Journal of Financial Services Marketing, 9(2): 179–183. S. Caudle (2005) ‘Homeland security,’ Public Performance & Management Review, 28(3): 352–375. H.S.B. Herath, W.G. Bremser (2005) ‘Real-option valuation of research and development investments: implications for performance measurement,’ Managerial Auditing Journal, 20(1): 55–72. F. Lhabitant (2000) ‘Coping with model risk,’ in The Professional Handbook of Financial Risk Management, M. Lore, L. Borodovsky (eds), Butterworth-Heinemann. J. Sobehart, S. Keenan (2001) ‘Measuring default accurately,’ Credit Risk Special Report, Risk, 14: 31–33.

Credit Scoring using Multiobjective Data Mining

1. C.L. Hwang, K. Yoon (1981) Multiple Attribute Decision Making: Methods and Applications. Springer-Verlag. 2. Y. Peng (2000) Management Decision Analysis. Science Publication; T.-C. Chu (2002) ‘Facility location selection using fuzzy TOPSIS under group decisions,’ International Journal of Uncertainty, Fuzziness & Knowledge-Based Systems, 10(6): 687–701. 3. D.L. Olson, D. Wu (2005) ‘Decision making with uncertainty and data mining,’ Advanced Data Mining and Applications: First International Conference, ADMA,

Notes 227

4. 5.

6. 7. 8. 9.

10.

11. 12. 13.

14.

X. Li, S. Wang, Z.Y. Dong eds, Lecture Notes in Artificial Intelligence. Keynote paper. Springer, 1–9. D.L. Olson (2005) ‘Comparison of weights in TOPSIS models,’ Mathematical and Computer Modelling, 40: 721–727. M. Freimer, P.L. Yu (1976) ‘Some new results on compromise solutions for group decision problems,’ Management Science, 22(6): 688–693; T.E. Dielman (2005) ‘Least absolute value regression: recent contributions,’ Journal of Statistical Computation & Simulation, 75(4): 263–286. S.C. Caples, M.E. Hanna (1997) ‘Least squares versus least absolute value in real estate appraisals,’ Appraisal Journal, 65(1): 18–24. G.W. Bassett, Jr. (1997) ‘Robust sports rating based on least absolute errors,’ American Statistician, 51(2): 99–105. Olson (2005), ibid. S.M. Lee, D.L. Olson (2004) ‘Goal programming formulations for a comparative analysis of scalar norms and ordinal vs. ratio data,’ Information Systems and Operational Research, 42(3): 163–174. A. Barnes (1987) ‘The analysis and use of financial ratios: a review article’, Journal of Business and Finance Accounting, 14: 449–461; H. Deng, C.-H. Yeh, R.J. Willis (2000) ‘Inter-company comparison using modified TOPSIS with objective weight’, Computers & Operations Research, 27: 963–973. D.L. Olson (2004) ‘Data set balancing’, Lecture Notes in Computer Science: Data Mining and Knowledge Management, Y. Shi, W. Xu, Z. Chen, eds. Springer, 71–80. J. Laurikkala (2002) ‘Instance-based data reduction for improved identification of difficult small classes’, Intelligent Data Analysis, 6(4): 311–322. B. Bull (2005) ‘Exemplar sampling: Nonrandom methods of selecting a sample which characterizes a finite multivariate population,’ American Statistician, 59(2): 166–172. D.L. Olson, D. Wu (2006) ‘Simulation of fuzzy multiattribute models for grey relationships,’ European Journal of Operational Research, 175(1): 111–120.

10 Online Banking Efficiency and Risk Evaluation with Principal Component Analysis 1. K. Furst, W.W. Lang, D. Nolle (2000) ‘Internet banking: developments and prospects,’ Economic and Policy Analysis, Working Paper 2000–9. 2. Dominion (2001) ‘Internet banking struggles for profits,’ available at www.stuff. co.nz/inl/index/0,1008,779016a28,FF.html. 3. B. Stafford (2001) ‘Risk management and internet banking: what every banker needs to know,’ Community Banker, 10(2): 48–49. 4. C-T.B. Ho, D. Wu (2009) ‘Online banking performance evaluation using data envelopment analysis and principal component analysis,’ Computers & Operations Research, 36(6): 1835–1842. 5. Jupiter Research.(2004) ‘FIND research, Institute for Information Industry,’ available at http://www.find.org.tw. 6. C. Serrano-Cinca, Y. Fuertes-Calle’n, C. Mar-Molinero (2005) ‘Measuring DEA efficiency in Internet companies,’ Decision Support Systems, 38: 557–573. 7. H.D. Sherman, F. Gold (1985) ‘Bank branch operating efficiency: evaluation with data envelopment analysis,’ Journal of Banking and Finance, 9(2): 297–316.

228

Notes

8. A. Soteriou, S.A. Zenios, (1999) ‘Operations, quality, and profitability in the provision of banking services,’ Management Science, 45(9): 1221–1238. 9. H. Tulkens (1993) ‘On FDH efficiency analysis: some methodological issues and applications to retail banking, courts and urban transit,’ Journal of Productivity Analysis, 4(1–2): 183–210. 10. A.N. Berger, D.B. Humphrey (1992) ‘Measurement and efficiency issues in commercial banking,’ in Z. Griliches, ed., Output Measurement in the Service Sectors, NBER Studies in Income and Wealth, 245–300. The University of Chicago Press. 11. A.N. Berger, D.B. Humphrey (1997) ‘Efficiency of financial institutions: international survey and direction for future research,’ European Journal of Operational Research, 98:175–212. 12. A.N. Berger, D. Hancock, D.B. Humphrey (1993) ‘Bank efficiency derived from the profit function,’ Journal of Banking and Finance, 17(2–3): 317–348. 13. D.D. Wu (2009) ‘Performance evaluation: an integrated method using data envelopment analysis and fuzzy preference relations,’ European Journal of Operational Research, 194(1): 227–235. 14. K. Eriksson, K. Kerem, D. Nilsson (2008) ‘The adoption of commercial innovations in the former Central and Eastern European markets: the case of internet banking in Estonia,’ International Journal of Bank Marketing, 26(3): 154–169. 15. Bank of America (2007) Annual Report 2007, available at www.rbs.com/microsites/ gra2007/downloads/RBS_GRA_2007.pdf. 16. Citibank (2007) Annual Report 2007, available at www.citi.com/citi/fin/data/k07c. pdf. 17. HSBC (2007) Annual Report 2007, available at www.investis.com/reports/hsbc_ ar_2007_En/report.php?type=1. 18. Barclays (2007) Annual Report 2007, available at www.barclaysannualreport.com/ index.html. 19. Chase (2007) Annual Report 2007, available at: http://investor.shareholder.com/ common/. 20. Wells Fargo (2007) Annual Report 2007, available at www.wellsfargo.com/downloads/ pdf/invest_relations/wf2007annualreport.pdf. 21. Lloyds (2007) Annual Report 2007, available at www.investorrelations.lloydstsb.com/ media/pdf_irmc/ir/2007/2007_LTSB_Group_R&A.pdf. 22. Royal Bank of Scotland (2007) Annual Report 2007, available at www.rbs.com/microsites/gra2007/downloads/RBS_GRA_2007.pdf. 23. SunTrust (2007) Annual Report 2007, available at www.suntrustenespanol.com/ suntrust. 24. Wachovia (2007) Annual Report 2007, available at www.wachovia.com/file/2007_ Wachovia_Annual_Report.pdf. 25. Basel (2005) ‘Amendment to the Capital Accord to the Incorporate Market Risks,’ Basel Committee on Banking Supervision, Basel.

11

Economic Perspective

1. F.H. Knight (1921) Risk, Uncertainty and Profit. Hart, Schaffner & Marx. 2. J.G. Courcelle-Seneuil (1852) ‘Profit,’ in Coquelin and Guillaumin, eds, Dictionnaire de l’ėconomie politique, 2nd ed. 3. J.H. Von Thünen (1826) The Isolated State.

Notes 229

4. F.B. Hawley (1907) Enterprise and the Productive Process. 5. A. Ganegoda, J. Evans (2012) ‘A framework to manage the measurable, immeasurable and the unidentifiable financial risk,’ Australian Journal of Management, 39(5): 5–34. 6. D. Li (2000) ‘On default correlation: a copula approach,’ Journal of Fixed Income, 9(4): 43–54. 7. H.M. Markowitz (1952) ‘Portfolio selection,’ The Journal of Finance, 17(1): 77–91. 8. E.F. Fama (1965) ‘Random walks in stock market prices,’ Financial Analysts Journal, 51(1): 404–419. 9. K. Buehler, A. Freeman, R. Hulme (2008) ‘The new arsenal of risk management,’ Harvard Business Review, 86(9): 93–100. 10. W.F. Sharpe (1970) Portfolio Theory and Capital Markets. McGraw-Hill Book Company. 11. D. Kahneman, A. Tversky (1972) ‘Subjective probability: a judgment of representativeness,’ Cognitive Psychology, 3: 430–454. 12. C. Mackay (1841) Extraordinary Popular Delusions and the Madness of Crowds. 13. S. Patterson (2010) The Quants: How a New Breed of Math Whizzes Conquered Wall Street and Nearly Destroyed It. Crown Business. 14. J.N. Stanard, M.G. Wacek (1991) ‘The spiral in the catastrophe retrocessional market,’ Casualty Actuarial Society Discussion Paper, May, Arlington, VA. 15. R. Lowenstein (2000) When Genius Failed: The Rise and Fall of Long-Term Capital Management. Random House. 16. N.N. Taleb (2007) The Black Swan: The Impact of the Highly Improbable. Penguin Books. 17. D. Kahneman, A. Tversky (2000) Choices, Values, and Frames. The University of Cambridge. 18. G.A. Akerlof, R.J. Shiller (2009) Animal Spirits: How Human Psychology Drives the Economy, and Why It Matters for Global Capitalism. Princeton University Press. 19. N. Doherty, J. Lamm-Tennant, L.T. Starks (2009) ‘Lessons from the financial crisis on risk and capital management: the case of insurance companies,’ Journal of Applied Corporate Finance, 21(4): 52–59. 20. M.R. Powers, T.Y. Powers, S. Gao (2012) ‘Risk finance for catastrophe losses with Pareto-calibrated Lévy-stable severities,’ Risk Analysis, 32(11): 1967–1977.

12

British Petroleum Deepwater Horizon

1. R. Zolkos, M. Bradford (2011) ‘Risk management faulted in probe of BP disaster’, Business Insurance, 45(36): 4–25. 2. R.T. Silves, L.K. Comfort (2012) ‘The Exxon Valdez and BP Deepwater Horizon oil spills: Reducing risk in socio-technical systems’, American Behavioral Scientist, 56(1): 76–103. 3. C. Perrow (1984) Normal Accidents: Living with High-risk Technologies, Basic Books. 4. A. Borison, G. Hamm (2011) ‘Black swan or black sheep?’ Risk Management, 58(3): 48–53. 5. P. Eckle, P. Burgherr (2013) ‘Bayesian data analysis of severe fatal accident risk in the oil chain’, Risk Analysis, 33(1): 146–160. 6. H. Abbasinejad, Y. Gudarzi Farahani, E. Ashari Ghara (2012) ‘Energy consumption in Iran with Bayesian approach’, OPEC Energy Review, 36(4): 444–455.

230 Notes

13

Bank Efficiency Analysis

1. H.D. Sherman, F. Gold (1985) ‘Bank branch operating efficiency: evaluation with data envelopment analysis,’ Journal of Banking and Finance, 9(2): 297–316; H. Tulkens (1993) ‘On FDH efficiency analysis: some methodological issues and applications to retail banking, courts and Urban transit,’ Journal of Productivity Analysis, 4(1/2): 183–210; A.N. Berger, D.B. Humphrey (1997) ‘Efficiency of financial institutions: international survey and directions for future research,’ European Journal of Operational Research, 98, 175–212; J.A. Clark (1996) ‘Economic cost, scale efficiency, and competitive viability in banking,’ Journal of Money, Banking and Credit, 28(3): 342–364; R. Deyoung (1997) ‘A diagnostic test for the distribution-free efficiency estimator: an example using US commercial bank data,’ European Journal of Operational Research, 98(2): 243–249. 2. S.H. Wang (2003) ‘Adaptive non-parametric efficiency frontier analysis: a neuralnetwork-based model,’ Computers & Operations Research, 30: 279–295. 3. A.D. Athanassopoulos, S.P. Curram (1996) ‘A comparison of data envelopment analysis and artificial neural networks as tool for assessing the efficiency of decision making units,’ Journal of the Operational Research Society, 47(8): 1000–1016. 4. A. Costa, R.N. Markellos (1997) ‘Evaluating public transport efficiency with neural network models,’ Transportation Research C, 5(5): 301–312. 5. A.R. Fleissig, T. Kastens, D. Terrell (2000) ‘Evaluating the semi-nonparametric fourier, aim, and neural networks cost functions,’ Economics Letters, 68(3): 235–244. 6. D. Santin, F.J. Delgado, A. Valino (2004) ‘The measurement of technical efficiency: a neural network approach,’ Applied Economics, 36(6): 627–635. 7. P.C. Pendharkar, J.A. Rodger (2003) ‘Technical efficiency-based selection of learning cases to improve forecasting accuracy of neural networks under monotonicity assumption,’ Decision Support Systems, 36(1): 117–136. 8. Athanassopoulos and Curram (1996), op cit. 9. R.D. Banker, A. Charnes, W.W. Cooper (1984) ‘Some models for estimating technical and scale inefficiencies in data envelopment analysis,’ Management Science, 30: 1078–1092. 10. A. Charnes, W.W. Cooper, E. Rhodes (1978) ‘Measuring the efficiency of decision making units,’ European Journal of Operational Research, 6(2): 429–444. 11. R. Hecht Nielsen (1990) ‘Neural computing,’ Addison Wesley, 124–133; J.W. Shavlik, R.J. Mooney, G. G.Towell (1991) ‘Symbolic and neural learning algorithms: an experimental comparison,’ Machine Learning, 6: 111–143. 12. Pendharkar and Rodger (2003), op. cit. 13. M.D. Troutt, A. Rai, A. Zhang (1995) ‘The potential use of DEA for credit applicant acceptance systems,’ Computers and Operations Research, 4: 405–408.

14 Catastrophe Bond and Risk Modeling 1. A. Dassios, J. Jang (2003) ‘Pricing of catastrophe reinsurance and derivatives using the cox process with shot noise intensity,’ Finance and Stochastics, 7: 73–95. 2. H. Geman, M. Yor (1997) ‘Stochastic time changes in catastrophe option pricing,’ Insurance: Mathematics and Economics, 21: 185–193. 3. R.T. Silves, L.K. Comfort (2012) ‘The Exxon Valdez and BP Deepwater Horizon oil spills: reducing risk in socio-technical systems,’ American Behavioral Scientist, 56(1): 76–103.

Notes 231

4. Dassios and Jang (2003), op cit. 5. S. Jaimungal, T. Wang (2005) ‘Catastrophe options with stochastic interest rates and compound Poisson losses,’ Insurance: Mathematics and Economics, 38: 469–483. 6. S. Cox, H. Pedersen (2000) ‘Catastrophe risk bonds,’ North American Actuarial Journal, 48: 56–82. 7. J. Lee, M. Yu (2007) ‘Variation of catastrophe reinsurance with catastrophe bonds,’ Insurance: Mathematics and Economics, 41: 264–278. 8. R.C. Merton (1974) ‘On the pricing of corporate debt: the risk structure of interest rates,’ The Journal of Finance, 29(2): 449–470. 9. C. Perrow (1984) Normal Accidents: Living with High-risk Technologies, Basic Books. 10. V.E. Vaugirard (2003) ‘Pricing catastrophe bonds by an arbitrage approach,’ The Quarterly Review of Economics and Finance, 43: 119–132. 11. L. Zhu (2008) ‘Double exponential jump diffusion model for catastrophe bonds pricing,’ Journal of Fujian University of Technology, 6: 336–338 (in Chinese). 12 . L.F. Chang, M.W. Hung (2009) ‘Analytical valuation of catastrophe equity options with negative exponential jumps,’ Insurance: Mathematics and Economics, 44: 59–69. 13. Geman and Yor (1997), op cit.; Dassios and Jung (2003), op cit. 14. Andrieu, N. de Freitas, A. Doucet, M.I. Jordan (2003) ‘An introduction to MCMC for machine learning,’ Machine Learning, 50: 5–43. 15. D. Wu, D.L. Olson (2010) ‘Enterprise risk management: coping with model risk in a large bank,’ Journal of the Operational Research Society, 61(2): 179–190.

15

Bilevel Programming Merger Analysis in Banking

1. K. Aquino, A. Reed II (1998) ‘A social dilemma perspective on cooperative behavior in organizations: the effects of scarcity, communication, and unequal access on the use of a shared resource,’ Group & Organization Management, 23: 390–413. 2. P. Bonacich (1987) ‘Communication networks and collective action,’ Social Networks, 9: 389–396; J.A. Sniezek, D.R. May, J.E. Sawyer (1990) ‘Social uncertainty and interdependence: a study of resource allocation decision in groups,’ Organizational Behavior and Human Decision Processes, 46: 155–180. 3. E.A. Mannix (1993) ‘Organizations as resource dilemmas: the effects of power balance on coalition formation in small groups,’ Organizational Behavior and Human Decision Processes, 55: 1–22. 4. MEI Computer Technology Group Inc. (2011) ‘2011 Trade Promotion Management Trends.’ 5. Jemison, B. David, S.B. Sitkin (1986) ‘Corporate acquisitions: a process perspective,’ Academy of Management Review, 11(1): 145–163. 6. J. Paradi, S. Vela, H. Zhu (2010) ‘Adjusting for cultural differences, a new DEA model applied to a merged bank,’ Journal of Productivity Analysis, 33: 109–123. 7. S. Kreipl, M. Pinedo (2004) ‘Planning and scheduling in supply Chain: an overview of issues in practice,’ Production and Operations Management, 13(1): 29–77. 8. D.D. Wu, J.R. Birge (2012) ‘Serial chain merger evaluation model and application to mortgage banking,’ Decision Sciences, 43(1): 5–36. 9. Mannix (1993), op cit. 10. J. Bard (1998) Practical Bilevel Optimization: Algorithms and Applications. Kluwer Academic Publishers.

232 Notes

11. P. Hansen, B. Jaumard, G. Savard (1992) ‘New branch and bound rules for linear bilevel programming,’ SIAM Journal on Scientific and Statistical Computing, 13(5): 1194–1217. 12. Bard (1998), op cit. 13. W.W.Cooper, L.M. Seiford, K. Tone (2000) Data Envelopment Analysis. Kluwer. 14. S.C. Ray (2004) Data Envelopment Analysis: Theory and Techniques for Economics and Operations Research. Cambridge University Press, 189–208. 15. P. Bogetoft, D. Wang (2005) ‘Estimating the potential gains from mergers,’ Journal of Productivity Analysis, 23: 145–171. 16. Wu and Birge (2012), op cit. 17. R. Maddigan, J. Zaima (1985) ‘The profitability of vertical integration,’ Managerial and Decision Economics, 6(3): 178–179. 18. E.H. MacDonald (2001) ‘GIS in banking: evaluation of Canadian Bank mergers,’ Canadian Journal of Regional Science, 24(3): 419–442. 19. S. Finkelstein, H. Jerayr (2002) ‘Understanding acquisition performance: the role of transfer effects,’ Organization Science, 13(1): 36–47. 20. Cooper et al. (2000), op cit. 21. C.H. Wang, R. Gopal, S. Zionts (1997) ‘Use of data envelopment analysis in assessing information technology impact on firm performance,’ Annals of Operations Research, 73: 191 –213. 22. Ibid. 23. J.D. Cummins, X. Xie (2008) ‘Mergers and acquisitions in the US property-liability insurance industry: productivity and efficiency effects,’ Journal of Banking & Finance, 2(1): 30–55.

16

Sustainability and Risk in Globalization

1. E.G. Baranoff (2004) ‘Risk management: a focus on a more holistic approach three years after September 11,’ Journal of Insurance Regulation, 22(4): 71–81. 2. D.B. McDonald (2011) ‘When risk management collides with enterprise sustainability,’ Journal of Leadership, Accountability and Ethics, 8(3): 56–66. 3. I.I. Mitroff, M.C. Alpaslan (2003) ‘Preparing for evil,’ Harvard Business Review, 81(4): 109–115. 4. C. Perrow (1999) Normal Accidents: Living with High-Risk Technologies. Princeton University Press. 5. M. Drew (2007) ‘Information risk management and compliance – expect the unexpected,’ BT Technology Journal, 25(1), 19–29. 6. T. Lambooy (2011) ‘Corporate social responsibility: sustainable water use,’ Journal of Cleaner Production, 19(8): 852–866. 7. D. Ng, P.D. Goldsmith (2010) ‘Bio energy entry timing from a resource based view and organizational ecology perspective,’ International Food & Agribusiness Management Review, 13(2): 69–100. 8. D. Meyler, J.P. Stimpson, M.P. Cutghin (2007) ‘Landscapes of risk,’ Organization & Environment, 20(2): 204–212. 9. M. Santiago (2011) ‘The Huasteca rain forest,’ Latin American Research Review, 46: 32–54. 10. T.K. Zhelev (2005) ‘On the integrated management of industrial resources incorporating finances,’ Journal of Cleaner Production, 13(5): 469–474.

Notes 233

11. T.M. Mata, R.L. Smith, D.M. Young, C.A.V. Costa (2005) ‘Environmental analysis of gasoline blending components through their life cycle,’ Journal of Cleaner Production, 13(5): 517–523. 12. H. Von Blottnitz, M.A. Curran (2007) ‘A review of assessments conducted on bioethanol as a transportation fuel from a net energy, greenhouse gas, and environmental life cycle perspective,’ Journal of Cleaner Production, 15(7): 607–619. 13. A. Akcil (2006) ‘Managing cyanide: health, safety and risk management practices at Turkey’s Ovacik gold–silver mine,’ Journal of Cleaner Production, 14(8): 727–735. 14. N. Gȕlpinar, E. Canakoglu, D. Pachamanova (2014) ‘Robust investment decisions under supply disruption in petroleum markets,’ Computers & Operations Research, 44: 75–91. 15. F. Cucchiella, M. Gastaldi (2006) ‘Risk management in supply chains: a real option approach,’ Journal of Manufacturing Technology Management, 17(6): 700–720. 16. B. Ritchie, C. Brindley (2007) ‘An emergent framework for supply chain risk management and performance measurement,’ Journal of the Operational Research Society, 58: 1398–1411. 17. F.B. Hawley (1907) Enterprise and the Productive Process. 18. F.H. Knight (1921) Risk, Uncertainty, and Profit. Hart, Schaffner & Marx. 19. D. Kahneman, A. Tversky (2000) Choices, Values, and Frames. The University of Cambridge.

17

Risk from Natural Disasters

1. C. McDonald (2009) ‘New PRIMA president sees public RMs as masters of disaster,’ National Underwriter/Property & Casualty Risk & Benefits Management, 113(21): 17–31. 2. W.-J. Tan, P. Enderwick (2006) ‘Managing threats in the global era: the impact and response to SARS,’ Thunderbird International Business Review, 48(4): 515–536. 3. B. Lee, F. Preston (2012) Preparing for High-impact, Low-probability Events: Lessons from Eyjafjallajökull. Chatham House Report. 4. N.N. Taleb, D.G. Goldstein, M.W. Spitznagel (2009) ‘The six mistakes executives make in risk management,’ Harvard Business Review, 87(10): 78–81. 5. K. Hopkins (2003) ‘Value opportunity three: improving the ability to fulfill demand,’ Business Week, January 13. 6. A.S. Mukherjee (2008) The Spider’s Strategy: Creating Networks to Avert Crisis, Create Change, and Really Get Ahead. FT Press. 7. N. Kapucu, M. Van Wart (2008) ‘Making matters worse: an anatomy of leadership failures in managing catastrophic events,’ Administration & Society, 40(7): 711–740. 8. D. Alexander (2003) ‘Towards the development of standards in emergency management training and education,’ Disaster Prevention and Management, 12: 113–123. 9. G. Suder, D.W. Gillingham (2007) ‘Paradigms and paradoxes of agricultural risk governance,’ International Journal of Risk Assessment and Management, 7(3): 444–457. 10. L.A. Reilly, O. Courtenay (2007) ‘Husbandry practices, badger sett density and habitat composition as risk factors for transient and persistent bovine tuberculosis on UK cattle farms,’ Preventive Veterinary Medicine, 80(2–3): 129–142.

234 Notes

11. K.S. Markel, L.A. Barclay (2007) ‘The intersection of risk management and human resources: an illustration using genetic mapping,’ International Journal of Risk Assessment and Management, 7(3): 326–340. 12. D.H. Smaltz, R. Carpenter, J. Saltz (2007) ‘Effective IT governance in healthcare organizations: a tale of two organizations,’ International Journal of Healthcare Technology and Management, 8(1/2): 20–41. 13. D. Dalcher (2007) ‘Why the pilot cannot be blamed: a cautionary note about excessive reliance on technology,’ International Journal of Risk Assessment and Management, 7(3): 350–366. 14. M. Baucells, F.H. Heukamp (2009) ‘Probability and time tradeoff,’ Working Paper, http://ssrn.com/abstract=970570. 15. J. Pan, M. Wang, D. Li, J. Le4 (2009) ‘Automatic generation of seamline network using area Voronoi diagrams with overlap,’ IEEE Transactions on Geoscience and Remote Sensing, 47(6): 1737–1744. 16. D. Engel (2009) ‘Hi-tech solutions for crisis management,’ African Business, 352: 50. 17. J. Wei, D. Zhao, L. Liang (2009) ‘Estimating the growth models of news stories on disasters,’ Journal of the American Society for Information Science and Technology, 60(9): 1741–1755. 18. M. Saadatseresht, A. Mansourian, M. Taleai (2009) ‘Evacuation planning using multiobjective evolutionary optimization approach,’ European Journal of Operational Research, 198(1): 305–314. 19. R. Morelli, A. Tucker, N. Danner, T.R. de Lanerolle, H.J.C. Ellis, O. Izmirli, D. Krizanc, G. Parker (2009) ‘Revitalizing computing education through free and open source software for humanity,’ Communications of the ACM, 52(8): 67–75. 20. N. Santella, L.J. Steinberg, K. Parks (2009) ‘Decision making for extreme events: Modeling critical infrastructure interdependencies to aid mitigation and response planning,’ Review of Policy Research, 26(4): 409–422. 21. F. Aleskerov, A.L. Say, A. Toker, H.OL. Akin, G. Altay (2005) ‘A cluster-based decision support system for estimating earthquake damage and casualties,’ Disasters, (3): 255–276.

18 Pricing of Carbon Emission Exchange in the EU ETS 1. M. Kainuma, Y. Matsuoka, T. Morita (1999) ‘Development of AIM (Asian-Pacific Integrated Model) for coping with global warming,’ Proceedings of the IEEE International Conference on System Man and Cybernetics, 6: 569–574. 2. W.D. Nordhaus, J.G. Boyer (1999) ‘Requiem for Kyoto: an economic analysis of the Kyoto protocol’, The Energy Journal, 20: 93–130. 3. W.D. Nordhaus (2001) ‘Climate change: global warming economics’, Science, 294(5545): 1283–1284. 4. P. Capros, L. Mantzos (2000) ‘The economic effects of industry-level emission trading to reduce greenhouse gases’, Report to DG environment, E3M-Laboratory 21 at ICCS/NTUA; P. Criqui, A. Kitous (2003) ‘Impacts of linking JI and CDM credits to the European emission allowance trading scheme,’ KPI technical report; G. Klepper, S. Peterson (2004) ‘The EU emissions trading scheme: allowance prices, trade flows, competitiveness effects’, European Environment, 14(4): 201–218; G. Klepper, S. Peterson (2006) ‘Emissions trading, CDM, JI and more – the climate strategy of the EU’, Energy Journal, 27(2): 1–26.

Notes 235

5. G. Daskalakis, D. Psychoyios, R.N. Markellos (2009) ‘Modeling CO2 emission allowance prices and derivatives: evidence from the European trading’, Journal of Banking & Finance, 33(7): 1230–1241. 6. M. Uhrig-Homburg, M. Wagner (2006) ‘Success chances and optimal design of derivatives on CO2 emission certificates,’ Working Paper, University of Karlsruhe. 7. M.S. Paolella, L. Taschini (2006) ‘An econometric analysis of emission trading allowances,’ Research Paper Series 06–26, FINRISK: National Center of Competence in Research Financial Valuation and Risk Management. 8. E. Benz, S. Trück, (2006) ‘CO2 emission allowances trading in Europe – specifying a new class of assets’, Problems and Perspectives in Management, 4(3): 30–40. 9. S. Borak, W. Härdle, S. Trück, R. Weron (2006) ‘Convenience yields for CO2 emission allowance future contracts,’ SFB 649 discussion paper 2006–076, SFB Economic Risk Berlin. 10. J. Seifert, M. Uhrig-Hombur, M. Wagner (2008) ‘Dynamic behavior of CO2 spot prices,’ Journal of Environmental Economics and Management, 56(2): 180–194. 11. D. Burtraw (1996) ‘Cost savings sans allowance trades? Evaluating the SO2 emission trading program to date,’ Discussion Paper 95–30-REV. 12. J.M. Burniaux, J.O. Martins (2000) ‘Carbon emission leakages: a general equilibrium view,’ OECD Economics Department Working Papers No. 242. 13. T.J. Considine (2000) ‘The impacts of weather variations on energy demand and carbon emissions,’ Resource and Energy Economics, 22: 295–314. 14. J.Sijm, S. Bakker, Y. Chen, H. Harmesen, W. Lise (2005) ‘CO2 price dynamics: the implications of EU emissions trading on the price of electricity,’ Report ECNC05–81, Energy Research Center of the Netherlands (ECN). 15. U. Ciorba, A. Lanza, F. Pauli (2001) ‘Kyoto protocol and emission trading: does the US make a difference?’ FEEM working paper 90.2001, Milan. 16. U. Springer (2003) ‘The market for tradable GHG permits under the Kyoto Protocol: a survey of model studies’, Energy Economics, 25: 527–551. 17. U. Springer, M. Varilek (2004) ‘Estimating the price of tradable permits for greenhouse gas emissions in 2008–2012’, Energy Policy, 32: 611–621. 18. M. Manasanet-Bataller, A. Pardo, E. Valor (2007) ‘CO2 prices, energy and weather’. The Energy Journal, 28(3): 73–92. 19. D.B. Nelson (1991) ‘Conditional heteroskedasticity in asset returns: a new approach’, Econometrica, 59, 347–370.

19 Volatility Forecasting of the Crude Oil Market 1. R. Bacon, M. Kojima (2008) ‘Coping with Oil Price Volatility,’ Energy sector management assistance program, Energy Security Special Report 005/08. 2. J.C. Hung, M.C. Lee, H.C. Liu (2008) ‘Estimation of value-at-risk for energy commodities via fat-tailed GARCH models,’ Energy Economics, 30(3):1173–1191. 3. P.K. Narayan, S. Narayan, A. Prasad (2008) ‘Understanding the oil price-exchange rate nexus for the Fiji islands,’ Energy Economics, 30(5): 2686–2696. 4. F. Malik, B.T. Ewing (2009) ‘Volatility transmission between oil prices and equity sector returns,’ International Review of Financial Analysis, 18(3): 95–100. 5. A.H. Alizadeh, N.K. Nomikos, P.K. Pouliasis (2008) ‘A Markov regime switching approach for hedging energy commodities,’ Journal of Banking & Finance, 32(9):1970–1983.

236

Notes

6. C. Aloui, R. Jammazi (2009) ‘The effects of crude oil shocks on stock market shifts behaviour: a regime switching approach,’ Energy Economics, 31(5): 789–799. 7. F. Klaassen (2002) ‘Improving GARCH volatility forecasts with regime-switching GARCH,’ Empirical Economics, 27: 363–394. 8. A. Cologni, M. Manera (2009) ‘The asymmetric effects of oil shocks on output growth: a Markov–Switching analysis for the G-7 countries,’ Economic Modelling, 26(1): 1–29. 9. Y. Fan, Y.J. Zhang, H.T. Tsaic, Y.M. Wei (2008) ‘Estimating ‘Value at Risk’ of crude oil price and its spillover effect using the GED-GARCH approach,’ Technological Change and the Environment, 30(6): 3156–3171. 10. C. Aloui, S. Mabrouk (2009) ‘Value-at-risk estimations of energy commodities via long-memory, asymmetry and fat-tailed GARCH models,’ Energy Policy, 38(5): 2326–2339. 11. P. Agnolucci (2009) ‘Volatility in crude oil futures: a comparison of the predictive ability of GARCH and implied volatility models,’ Energy Economics, 31(2): 316–321. 12. C. Engel (1994) ‘Can the Markov switching model forecast exchange rates? ‘ Journal of International Economics, 36(1): 151–165. 13. M.T. Vo (2009) ‘Regime-switching stochastic volatility: evidence from the crude oil market,’ Energy Economics, 31(5): 779–788. 14. E. Fama (1970) ‘Efficient capital markets: a review of theory and empirical work,’ Journal of Finance, 25: 383–417. 15. R.F. Engle (1982) ‘Autoregressive conditional heteroscedasticity with estimates of variance of United Kingdom inflation,’ Econometrica, 50: 987–1008. 16. T. Bollerslev (1986) ‘Generalized autoregressive conditional heteroskedasticity,’ Journal of Econometrics, 31: 307–327. 17. D.B. Nelson (1991) ‘Conditional heteroskedasticity in asset returns: a new approach,’ Econometrica, 59: 347–370. 18. J.E. Raymond, R.W. Rich (1997) ‘Oil and the macroeconomy: a Markov stateswitching approach,’ Journal of Money, Credit and Banking, 29(2): 193–213. 19. J.D. Hamilton (1989) ‘A new approach to the economic analysis of nonstationary time series and the business cycle,’ Econometrica, 57(2): 357–384. 20. D. Cousineau, S. Brown, A. Heathcote (2004) ‘Fitting distributions using maximum likelihood: methods and packages,’ Behavior Research Methods, Instruments, & Computers, 36: 742–756.

20 Confucius Three-stage Learning of Risk Management 1. D. Gardner (2007) The Four Books. The Teachings of the Later Confucian Tradition. Hackett Publishing. 2. X. Yao, H. Yao (2000) An Introduction to Confucianism. Cambridge University Press. 3. A. Ben-Tal, A. Nemirovski (2000) ‘Robust solutions of linear programming problems contaminated with uncertain data’, Mathematical Programming, 88, 411–424. 4. J.C. Smith (2009) Pseudoscience and Extraordinary Claims of the Paranormal: A Critical Thinker. Wiley-Blackwell. ISBN 978–1405181228. 5. D. Wu, D.L. Olson (2009) ‘Introduction to the special section on optimizing risk management. Methods and tools’, Human and Ecological Risk Assessment, 15(2): 220–226.

Notes 237

6. Y. Bauman, G. Klein (2010) The Cartoon Introduction to Economics: Volume One: Microeconomics. Hill and Wang. 7. H.M. Markowitz (1959) Portfolio Selection: Efficient Diversification of Investments. John Wiley & Sons (reprinted by Yale University Press, 1970). 8. W.F. Sharpe (1964) ‘Capital asset prices: a theory of market equilibrium under conditions of risk’, Journal of Finance, 19(3): 425–442. 9. J. Tobin (1958) ‘Liquidity preference as behavior towards risk’, The Review of Economic Studies, 25: 65–86. 10. G.A. Akerlof (1970) ‘The market for “lemons”: quality uncertainty and the market mechanism,’ Quarterly Journal of Economics, 84(3): 488–500. 11. Counterparty Risk Management Policy Group III (2008) Containing Systemic Risk, August 6.

References H. Abbasinejad, Y. Gudarzi Farahani, E. Ashari Ghara (2012) ‘Energy consumption in Iran with Bayesian approach’, OPEC Energy Review, 36(4): 444–455. P. Agnolucci (2009) ‘Volatility in crude oil futures: a comparison of the predictive ability of GARCH and implied volatility models’, Energy Economics, 31(2): 316–321. N. Ahmad, D. Berg, G.R. Simons (2006) ‘The integration of analytic hierarchy process and data envelopment analysis in a multi-criteria decision-making problem’, International Journal of Information Technology and Decision Making, 5: 263–276. G. Ailon (2012) ‘The discursive management of financial risk scandals: the case of Wall Street Journal commentaries on LTCM and Enron’, Qualitative Sociology, 35: 251–270. A. Akcil (2006) ‘Managing cyanide: health, safety and risk management practices at Turkey’s Ovacik gold-silver mine’, Journal of Cleaner Production, 14(8): 727–735. G.A. Akerlof (1970) ‘The Market for “Lemons”: Quality Uncertainty and the Market Mechanism,’ Quarterly Journal of Economics, 84(3): 488–500. G.A. Akerlof, R.J. Shiller (2009) Animal Spirits: How Human Psychology Drives the Economy, and Why It Matters for Global Capitalism. Princeton University Press. F. Aleskerov, A.L. Say, A. Toker, H.O.L. Akin, G. Altay (2005) ‘A cluster-based decision support system for estimating earthquake damage and casualties’, Disasters, 3: 255–276. D. Alexander (2003) ‘Towards the development of standards in emergency management training and education’, Disaster Prevention and Management, 12, 113–123. G.J. Alexander, A.M. Baptista (2004) ‘A comparison of VaR and CVaR constraints on portfolio selection with the mean-variance model’, Management Science, 50(9): 1261–1273. A.H. Alizadeh, N.K. Nomikos, P.K. Pouliasis (2008) ‘A Markov regime switching approach for hedging energy commodities’, Journal of Banking & Finance, 32(9): 1970–1983. C. Aloui, R. Jammazi (2009) ‘The effects of crude oil shocks on stock market shifts behaviour: a regime switching approach’, Energy Economics, 31(5): 789–799. C. Aloui, S. Mabrouk (2009) ‘Value-at-risk estimations of energy commodities via longmemory, asymmetry and fat-tailed GARCH models’, Energy Policy, 38(5): 2326–2339. U. Anders, M. Sandstedt (2003) ‘An operational risk scorecard approach’, Risk, 16(1): 47–50. W. Antweiler, M. Frank (2004) ‘Is all that talk just noise? The information content of internet stock message boards,’ Journal of Finance, 59(3): 1259–1295. K. Aquino, A. Reed II (1998) ‘A social dilemma perspective on cooperative behavior in organizations: the effects of scarcity, communication, and unequal access on the use of a shared resource,’ Group & Organization Management, 23: 390–413. The Association of Risk Managers (2010) A Structured Approach to Enterprise Risk Management (ERM) and the Requirements of ISO 31000. COSO. A.D. Athanassopoulos, S.P. Curram (1996) ‘A comparison of data envelopment analysis and artificial neural networks as tool for assessing the efficiency of decision making units’, Journal of the Operational Research Society, 47(8): 1000–1016. R. Bacon, M. Kojima (2008) ‘Coping with Oil Price Volatility’, Energy Sector Management Assistance Program, Energy Security Special Report 005/08. M. Baker, J. Wurgler (2006) ‘Investor sentiment and the cross-section of stock returns,’ Journal of Finance, 61(4): 1645–1680. 238

References

239

B. Ballou, D.L. Heitger (2005) ‘A building-block approach for implementing COSO’s enterprise risk management-integrated framework’, Management Accounting Quarterly, 6(2): 1–10. R.D. Banker, A. Charnes, W.W. Cooper (1984) ‘Some models for estimating technical and scale inefficiencies in data envelopment analysis’, Management Science, 30: 1078–1092. E.G. Baranoff (2004) ‘Risk management: a focus on a more holistic approach three years after September 11’, Journal of Insurance Regulation, 22(4): 71–81. J. Bard (1998) Practical Bilevel Optimization: Algorithms and Applications. Kluwer Academic Publishers. L. Bargeron, K. Lehn, C. Zutter (2009) ‘Sarbanes-Oxley and corporate risk-taking’, Journal of Accounting and Economics, 49(1–2): 34–52. A. Barnes (1987) ‘The analysis and use of financial ratios: a review article’, Journal of Business and Finance Accounting, 14: 449–461. A. Barua, P.L. Brockett, W.W. Cooper, H. Deng, B.R. Parker, T.W. Ruefli, A. Whinston (2004) ‘Multi-factor performance measure model with an application to fortune 500 companies’, Socio-Economic Planning Sciences, 38: 233–253. Basel Committee on Banking Supervision (2005) Amendment to the Capital Accord to the Incorporate Market Risks. Basel. G.W. Bassett, Jr. (1997) ‘Robust sports rating based on least absolute errors’, American Statistician, 51(2): 99–105. D. Bathia, D. Bredin (2012) ‘An examination of investor sentiment effect on G7 stock market returns,’ European Journal of Finance, DOI: 10.1080/1351847X.2011.636834. Y. Bauman, G. Klein (2010) The Cartoon Introduction to Economics: Volume One: Microeconomics. Hill and Wang. A. Ben-Tal, A. Nemirovski (2000) ‘Robust solutions of linear programming problems contaminated with uncertain data’, Mathematical Programming, 88: 411–424. E. Benz, S. Trück, (2006) ‘CO2 emission allowances trading in Europe – specifying a new class of assets,’ Problems and Perspectives in Management, 4(3): 30–40. A.N. Berger, D. Hancock, D.B. Humphrey (1993) ‘Bank efficiency derived from the profit function,’ Journal of Banking and Finance, 17(2–3): 317–348. A.N. Berger, D.B. Humphrey (1992) ‘Measurement and efficiency issues in commercial banking,’ in Z. Griliches, ed., Output Measurement in the Service Sectors, NBER Studies in Income and Wealth. The University of Chicago Press, pp. 245–300. A.N. Berger, D.B. Humphrey (1997) ‘Efficiency of financial institutions: international survey and direction for future research,’ European Journal of Operational Research, 98: 175–212. D. Bigio, R.L. Edgeman, T. Ferleman (2004) ‘Six sigma availability management of information technology in the office of the chief technology officer of Washington, DC,’ Total Quality Management, 15(5–6): 679–687. F. Black, M. Scholes (1972) ‘The valuation of option contracts and a test of market efficiency’, The Journal of Finance, 27(2): 399–417. A.S. Blinder (2013) After the Music Stopped: The Financial Crisis, the Response, and the Work Ahead. The Penguin Press. P. Bogetoft, D. Wang (2005) ‘Estimating the potential gains from mergers’, Journal of Productivity Analysis, 23: 145–171. T. Bollerslev (1986) ‘Generalized autoregressive conditional heteroskedasticity’, Journal of Econometrics, 31(3): 307–327. P. Bonacich (1987) ‘Communication networks and collective action’, Social Networks, 9: 389–396.

240 References

S. Borak, W. Härdle, S. Trück, R. Weron (2006) ‘Convenience yields for CO2 emission allowance future contracts,’ SFB 649 discussion paper 2006–076, SFB Economic Risk Berlin. A. Borison, G. Hamm (2011) ‘Black swan or black sheep?’ Risk Management, 58(3): 48–53. B.M. Bowling, L. Rieger (2005) ‘Success factors for implementing enterprise risk management,’ Bank Accounting and Finance, 18(3): 21–26. R. Boyd (2011) Fatal Risk: A Cautionary Tale of AIG’s Corporate Suicide. Wiley. M. Brunnermeier (2009) ‘Deciphering the liquidity and credit crunch 2007–2008,’ Journal of Economic Perspectives, 23, 77–100. K. Buehler, A. Freeman, R. Hulme (2008) ‘The new arsenal of risk management,’ Harvard Business Review, 86(9): 93–100. B. Bull (2005) ‘Exemplar sampling: nonrandom methods of selecting a sample which characterizes a finite multivariate population,’ American Statistician, 59(2): 166–172. C. Burges (1998) ‘A tutorial on support vector machines for pattern recognition,’ Data Mining and Knowledge Discovery, 2(2): 121–167. D. Burtraw (1996) ‘Cost savings sans allowance trades? Evaluating the SO2 emission trading program to date,’ Discussion Paper 95–30-REV. J.M. Burniaux, J.O. Martins (2000) ‘Carbon emission leakages: a general equilibrium view,’ OECD Economics Department Working Papers No. 242. H.N. Butler, L.E. Ribstein (2006) The Sarbanes-Oxley Debacle: What We’ve Learned; How to Fix It, AEI. J. Calandro, Jr., S. Lane (2006) ‘An introduction to the enterprise risk scorecard,’ Measuring Business Excellence, 10(3), 31–40. S.C. Caples, M.E. Hanna (1997) ‘Least squares versus least absolute value in real estate appraisals,’ Appraisal Journal, 65(1): 18–24. P. Capros, L. Mantzos (2000) ‘The economic effects of industry-level emission trading to reduce greenhouse gases,’ Report to DG environment, E3M-Laboratory 21 at ICCS/ NTUA. C.S. Carlos, F.C. Yolanda, M.M. Cecilio (2005) ‘Measuring DEA efficiency in internet companies,’ Decision Support Systems, 38: 557–573. F. Caron, J. Vanthienen, B. Baesens (2013) ‘A comprehensive investigation of the applicability of process mining techniques for enterprise risk management,’ Computers in Industry, 64: 464–475. S. Caudle (2005) ‘Homeland security,’ Public Performance & Management Review, 28(3): 352–375. K. Cengiz, C. Ufuk, U. Ziya (2003) ‘Multi-criteria supplier selection using Fuzzy AHP, Logistics Information Management, 16(6): 382–394. W. Chan (2003) ‘Stock price reaction to news and no-news: drift and reversal after headlines,’ Journal of Financial Economics, 70: 223–260. L.F. Chang, M.W. Hung, (2009) ‘Analytical valuation of catastrophe equity options with negative exponential jumps,’ Insurance: Mathematics and Economics, 44: 59–69. A. Charnes, W.W. Cooper, E. Rhodes (1978) ‘Measuring the efficiency of decision making units,’ European,’ Journal of Operational Research, 6(2): 429–444. V. Chavez-Demoulin, P. Embreechts, J. Nešlehová (2006) ‘Quantitative models for operational risk: extremes, dependence and aggregation,’ Journal of Banking & Finance, 30: 399–417. X. Chen, Z. Wang, D.D. Wu (2013) ‘Modeling the price mechanism of carbon emission exchange in the European Union Emission Trading System,’ Human and Ecological Risk Assessment, 19(5): 1309–1323.

References

241

Chien-Ta Ho, D.D. Wu, D.L. Olson (2009) ‘A risk scoring model and application to measuring internet stock performance,’ International Journal of Information Technology and Decision Making, 8(1): 133–149. T.-C. Chu (2002) ‘Facility location selection using fuzzy TOPSIS under group decisions,’ International Journal of Uncertainty, Fuzziness & Knowledge-Based Systems, 10(6): 687–701. U. Ciorba, A. Lanza, F. Pauli (2001) ‘Kyoto protocol and emission trading: does the US make a difference?’ FEEM working paper 90.2001, Milan. J.A. Clark (1996) ‘Economic cost, scale efficiency, and competitive viability in banking,’ Journal of Money, Banking and Credit, 28(3): 342–364. B. Cohen (1997) The Edge of Chaos: Financial Booms, Bubbles, Crashes and Chaos. John Wiley & Sons Ltd. A. Cologni, M. Manera (2009) ‘The asymmetric effects of oil shocks on output growth: a Markov–Switching analysis for the G-7 countries,’ Economic Modelling, 26(1): 1–29. T.J. Considine (2000) ‘The impacts of weather variations on energy demand and carbon emissions,’ Resource and Energy Economics, 22: 295–314. G. Cooper (2008) The Origin of Financial Crises: Central Banks, Credit Bubbles and the Efficient Market Fallacy. Vintage Books. W.W. Cooper, L.M. Seiford, K. Tone (2000) Data Envelopment Analysis. Kluwer. COSO (2004) Enterprise Risk Management – Integrated Framework: Executive Summary. September. A. Costa, R. N Markellos (1997) ‘Evaluating public transport efficiency with neural network models,’ Transportation Research C, 5(5): 301–312. Counterparty Risk Management Policy Group III (2008) Containing Systemic Risk, August 6. J.G. Courcelle-Seneuil (1852) ‘Profit,’ in Coquelin and Guilllaumin, eds, Dictionnaire de l’ėconomie politique, 2nd ed. D. Cousineau, S. Brown, A. Heathcote (2004) ‘Fitting distributions using maximum likelihood: methods and packages,’ Behavior Research Methods, Instruments, & Computers, 36: 742–756. S. Cox, H. Pedersen (2000) ‘Catastrophe risk bonds,’ North American Actuarial Journal, 48: 56–82. P. Criqui, A. Kitous (2003) ‘Impacts of linking JI and CDM credits to the European emission allowance trading scheme,’ KPI technical report. M. Crouhy, D. Galai, R. Mark (1998) ‘Model Risk,’ Journal of Financial Engineering, 7(3/4), 267–288; reprinted in Model Risk: Concepts, Calibration and Pricing, (ed. R. Gibson), Risk Book, 2000, 17–31. M. Crouhy, D. Galai, R. Mark (2000) ‘A comparative analysis of current credit risk models,’ Journal of Banking & Finance, 24, 59–117. F. Cucchiella, M. Gastaldi (2006) ‘Risk management in supply chains: a real option approach,’ Journal of Manufacturing Technology Management, 17(6): 700–720. J.D. Cummins, X. Xie (2008), ‘Mergers and acquisitions in the US property-liability insurance industry: productivity and efficiency effects,’ Journal of Banking & Finance, 2(1): 30–55. D. Dalcher (2007) ‘Why the pilot cannot be blamed: a cautionary note about excessive reliance on technology,’ International Journal of Risk Assessment and Management, 7(3): 350–366. G. Daskalakis, D. Psychoyios, R.N. Markellos (2009) ‘Modeling CO2 emission allowance prices and derivatives: evidence from the European trading,’ Journal of Banking & Finance, 33(7): 1230–1241.

242 References

A. Dassios, J. Jang (2003) ‘Pricing of catastrophe reinsurance and derivatives using the cox process with shot noise intensity,’ Finance and Stochastics, 7: 73–95. G. Dell’Arriccia, L. Laeven, D. Igan (2008) ‘Credit booms and lending standards: evidence from the subprime mortgage market,’ IMF Working Paper 08/106. H. Deng, C.-H., Yeh, R.J. Willis (2000) ‘Inter-company comparison using modified TOPSIS with objective weight,’ Computers & Operations Research, 27: 963–973. S. Deng, Z. Xia (2006) ‘A real options approach for pricing electricity tolling agreements,’ International Journal of Information Technology and Decision Making, 5: 421–436. A. Dey (2010) ‘The chilling effect of Sarbanes-Oxley: a discussion of Sarbanes-Oxley and corporate risk-taking,’ Journal of Accounting and Economics, 49(1–2): 53–57. R. Deyoung (1997) ‘A diagnostic test for the distribution-free efficiency estimator: an example using US commercial bank data,’ European Journal of Operational Research, 98(2): 243–249. G. Dickinson (2001) ‘Enterprise risk management: its origins and conceptual foundation,’ The Geneva Papers on Risk and Insurance, 26(3): 360–366. T.E. Dielman (2005) ‘Least absolute value regression: recent contributions,’ Journal of Statistical Computation & Simulation, 75(4): 263–286. N. Doherty, J. Lamm-Tennant, L.T. Starks (2009) ‘Lessons from the financial crisis on risk and capital management: the case of insurance companies,’ Journal of Applied Corporate Finance, 21(4): 52–59. D. Dong, Q. Dong (2003) ‘HowNet – a hybrid language and knowledge resource,’ Proceedings of 2003 International Conference on Natural Language Processing and Knowledge Engineering, 820 – 824, October 26–29. M. Drew (2007) ‘Information risk management and compliance – expect the unexpected,’ BT Technology Journal, 25(1): 19–29. N. Dunbar (1999) Investing Money: The Story of Long-Term Capital Management and the Legends Behind It. Wiley. P. Eckle, P. Burgherr (2013) ‘Bayesian data analysis of severe fatal accident risk in the oil chain,’ Risk Analysis, 33(1): 146–160. J.F. Egginton, J.I. Hilliard, A.P. Liebenberg, I.A. Liebenberg (2010) ‘What effect did AIG’s bailout, and the preceding events, have on its competitors?’ Risk Management and Insurance Review, 13(2): 225–249. H. Elsinger, A. Lehar, M. Summer (2006) ‘Risk assessment for banking systems,’ Management Science, 52(9), 1301–1314. C. Engel (1994) ‘Can the Markov switching model forecast exchange rates?’ Journal of International Economics, 36(1): 151–165. D. Engel (2009) ‘Hi-tech solutions for crisis management,’ African Business, 352, 50. R.F. Engle (1982) ‘Autoregressive conditional heteroscedasticity with estimates of variance of United Kingdom inflation,’ Econometrica, 50, 987–1008. K. Eriksson, K. Kerem, D. Nilsson (2008) ‘The adoption of commercial innovations in the former Central and Eastern European markets: the case of internet banking in Estonia,’ International Journal of Bank Marketing, 26(3): 154–169. P. Espahbodi (1991) ‘Identification of problem banks and binary choice models,’ Journal of Banking and Finance, 15, 53–71. E.F. Fama (1965) ‘Random walks in stock market prices,’ Financial Analysts Journal, 51(1): 404–419. E. Fama (1970) ‘Efficient capital markets: a review of theory and empirical work,’ Journal of Finance, 25: 383–417. Y. Fan, Y.J. Zhang, H.T. Tsaic, Y.M. Wei (2008) ‘Estimating ‘value at risk’ of crude oil price and its spillover effect using the GED-GARCH approach,’ Technological Change and the Environment, 30(6): 3156–3171.

References

243

M.J. Farrell (1957) ‘The measurement of productive efficiency,’ Journal of the Royal Statistical Society 120: 253–281. T.S. Felix, H.J. Chan (2003) ‘An innovative performance measurement method for supply chain management,’ Supply Chain Management: An International Journal, 8(3): 209–223. G.J. Fielding, T.T. Babitsky, M.E. Brenner (1985) ‘Performance evaluation for bus transit,’ Transportation Research, 19A(1): 73–82. S. Finkelstein, H. Jerayr (2002) ‘Understanding acquisition performance: the role of transfer effects,’ Organization Science, 13(1): 36–47. A.R. Fleissig, T. Kastens, D. Terrell (2000) ‘Evaluating the semi-nonparametric fourier, aim, and neural networks cost functions,’ Economics Letters, 68(3): 235–244. L. Fox (2003) Enron: The Rise and Fall. Wiley. M. Freimer, P.L. Yu (1976) ‘Some new results on compromise solutions for group decision problems,’ Management Science, 22(6): 688–693. B. Freisleben, K. Ripper (1997) ‘Volatility estimation with a neural network,’ Proceedings of the IEEE/IAFE on Computational Intelligence for Financial Engineering, 177–181, March 24–25. M. Friedman, L.J. Savage (1948) ‘The utility analysis of choices involving risk,’ The Journal of Political Economy, 56(4): 279–304. K. Furst, W.W. Lang, D. Nolle (2000) ‘Internet banking: developments and prospects,’ Economic and Policy Analysis, Working Paper 2000–9. A. Ganegoda, J. Evans (2012) ‘A framework to manage the measurable, immeasurable and the unidentifiable financial risk,’ Australian Journal of Management, 39(5): 5–34. R. Garcia, É. Renault, G. Tsafack (2007) ‘Proper conditioning for coherent VaR in portfolio management,’ Management Science, 53(3), 483–494. D. Gardner (2007) The Four Books. The Teachings of the Later Confucian Tradition. Hackett Publishing. H. Geman, M. Yor (1997) ‘Stochastic time changes in catastrophe option pricing,’ Insurance: Mathematics and Economics, 21: 185–193. P. Goldsmith-Pinkham, T. Yorulmazer (2010) ‘Liquidity, bank runs, and bailouts: spillover effects during the Northern Rock episode,’ Journal of Financial Service Research, 37(2/3): 83–98. G. Gorton (2008) ‘The panic of 2007,’ NBER Working Paper No. 14358. N. Gülpinar, E. Canakoglu, D. Pachamanova (2014) ‘Robust investment decisions under supply disruption in petroleum markets,’ Computers & Operations Research, 44: 75–91. M. Gulser, M. Ilhan (2001) ‘Risk and return in the world’s major stock markets,’ Journal of Investing, (Spring): 62–67. J.D. Hamilton (1989) ‘A new approach to the economic analysis of nonstationary time series and the business cycle,’ Econometrica, 57(2): 357–384. P. Hansen, B. Jaumard, G. Savard (1992) ‘New branch and bound rules for linear bilevel programming,’ SIAM Journal on Scientific and Statistical Computing, 13(5): 1194–1217. F.B. Hawley (1907) Enterprise and the Productive Process. R. Hecht Nielsen (1990) ‘Neural Computing,’ Addison Wesley: 124–133. H.S.B. Herath, W.G. Bremser (2005) ‘Real-option valuation of research and development investments: implications for performance measurement,’ Managerial Auditing Journal, 20(1): 55–72. H.N. Higgins (2012) ‘Learning internal controls from a fraud case at Bank of China,’ Issues in Accounting Education, 27(4): 1171–1192. C-T. Ho (2006) ‘Measuring bank operations performance: an approach based on grey relation analysis,’ Journal of the Operational Research Society, 57: 227–349.

244 References

C-T. Ho, D.S. Zhu (2004) ‘Performance measurement of Taiwan’s commercial banks,’ International Journal of Productivity and Performance Management, 53(5): 425–434. C-T. Ho, D. Wu (2009) ‘Online banking performance evaluation using data envelopment analysis and principal component analysis,’ Computers & Operations Research, 36(6): 1835–1842. J. Hobbs (2011) ‘Financial derivatives, the mismanagement of risk and the case of AIG,’ CPCU eJournal, 64(7): 1–8. C. Hollingsworth (2012) ‘Risk management in the post-SOX era,’ International Journal of Auditing, 16: 35–53. K. Hopkins (2003) ‘Value opportunity three: improving the ability to fulfill demand,’ Business Week, January 13. S.N. Huang, T.L. Kao (2006) ‘Measuring managerial efficiency in non-life insurance companies: an application of two-stage data envelopment analysis,’ International Journal of Management, 23(3): 699–720. D.W. Hubbard (2009) The Failure of Risk Management: Why It’s Broken and How to Fix It. John Wiley & Sons. J.C. Hung, M.C. Lee, H.C. Liu (2008) ‘Estimation of value-at-risk for energy commodities via fat-tailed GARCH models,’ Energy Economics, 30(3): 1173–1191. C. Hurt (2014) ‘The duty to manage risk,’ The Journal of Corporate Law, 39(2): 153–267. C.L. Hwang, K. Yoon (1981) Multiple Attribute Decision Making: Methods and Applications. Springer-Verlag. T. Jacobson, J. Lindé, K. Roszbach (2006) ‘Internal ratings systems, implied credit risk and the consistency of banks’ risk classification policies,’ Journal of Banking & Finance, 30: 1899–1926. S. Jaimungal, T. Wang (2005) ‘Catastrophe options with stochastic interest rates and compound Poisson losses,’ Insurance: Mathematics and Economics, 38: 469–483. Jemison, B. David, S.B. Sitkin (1986) ‘Corporate acquisitions: a process perspective,’ Academy of Management Review, 11(1): 145–163. D. Kahneman, A. Tversky (1972) ‘Subjective probability: a judgment of representativeness,’ Cognitive Psychology, 3, 430–454. D. Kahneman, A. Tversky (2000) Choices, Values, and Frames. The University of Cambridge. M. Kainuma, Y. Matsuoka, T. Morita (1999) ‘Development of AIM (Asian-Pacific Integrated Model) for coping with global warming,’ Proceedings of the IEEE International Conference on System Man and Cybernetics, 6: 569–574. R.S. Kaplan, D.P. Norton (1992) ‘The balanced scorecard – measures that drive performance,’ Harvard Business Review, 70(1): 71–79. R.S. Kaplan, D.P. Norton (2006) Alignment: Using the Balanced Scorecard to Create Corporate Synergies. Harvard Business School Press Books. N. Kapucu, M. Van Wart (2008) ‘Making matters worse: an anatomy of leadership failures in managing catastrophic events,’ Administration & Society, 40(7): 711–740. B. Keys, T. Mukherjee, A. Seru, V. Vig (2010) ‘Did securitization lead to lax screening? Evidence from subprime loans,’ Quarterly Journal of Economics, 125: 307–362. F. Klaassen (2002) ‘Improving GARCH volatility forecasts with regime-switching GARCH,’ Empirical Economics, 27: 363–394. G. Klepper, S. Peterson (2004) ‘The EU emissions trading scheme: allowance prices, trade flows, competitiveness effects,’ European Environment, 14(4): 201–218. G. Klepper, S. Peterson (2006) ‘Emissions trading, CDM, JI and more – the climate strategy of the EU,’ Energy Journal, 27(2): 1–26. F.H. Knight (1921) Risk, Uncertainty and Profit. Hart, Schaffner & Marx.

References

245

S. Kreipl, M. Pinedo (2004) ‘Planning and scheduling in supply chain: an overview of issues in practice,’ Production and Operations Management, 13(1): 29–77. L. Laeven, F. Valencia (2008) ‘Systemic banking crises: a new database,’ International Monetary Fund Working Paper WP/08/224. L. Laeven, F. Valencia (2010) ‘Resolution of banking crises: the good, the bad, and the ugly,’ IMF Working Paper WP/10/146. T. Lambooy (2011) ‘Corporate social responsibility: sustainable water use,’ Journal of Cleaner Production, 19(8): 852–866. J. Laurikkala (2002) ‘Instance-based data reduction for improved identification of difficult small classes,’ Intelligent Data Analysis, 6(4): 311–322. B. Lee, F. Preston (2012) Preparing for High-impact, Low-probability Events: Lessons from Eyjafjallajőkull. Chatham House Report. J. Lee, M. Yu (2007) ‘Variation of catastrophe reinsurance with catastrophe bonds,’ Insurance: Mathematics and Economics, 41: 264–278. S.M. Lee, D.L. Olson (2004) ‘Goal programming formulations for a comparative analysis of scalar norms and ordinal vs. ratio data,’ Information Systems and Operational Research, 42(3): 163–174. F. Lhabitant (2000) ‘Coping with Model Risk,’ in The Professional Handbook of Financial Risk Management, M. Lore, L. Borodovsky (eds). Butterworth-Heinemann. D.X. Li (2000) ‘On default correlation: a copula approach,’ Journal of Fixed Income, 9(4): 43–54. N. Li, X. Liang, X. Li, C. Wang, Desheng D. Wu (2009) ‘Network environment and financial risk using machine learning and sentiment analysis,’ Human and Ecological Risk Assessment, 15(2): 227–252. P.M. Linsley, R.E. Slack (2013) ‘Crisis management and an ethic of care: the case of Northern Rock Bank,’ Journal of Business Ethics, 113(2): 285–295. S.F. Lo, W.M. Lu (2006) ‘Does size matter? Finding the profitability and marketability benchmark of financial holding companies,’ Journal of Operational Research, 23(2), 229–246. R. Lowenstein (2000) When Genius Failed: The Rise and Fall of Long-Term Capital Management. Random House. C. Luo, L.A. Seco, H. Wang, D.D. Wu (2010) ‘Risk modeling in crude oil market: a comparison of Markov switching and GARCH models,’ Kybernetics, 39(5): 750–769. X. Luo (2003) ‘Evaluating the profitability and marketability efficiency of large banks – an application of data envelopment analysis,’ Journal of Business Research, 56: 627–635. E.H. MacDonald (2001) ‘GIS in banking: evaluation of Canadian bank mergers,’ Canadian Journal of Regional Science, 24(3): 419–442. C. Mackay (1841) Extraordinary Popular Delusions and the Madness of Crowds. Richard Bentley. R. Maddigan, J. Zaima (1985) ‘The Profitability of Vertical Integration,’ Managerial and Decision Economics, 6(3): 178–179. F. Malik, B.T. Ewing (2009) ‘Volatility transmission between oil prices and equity sector returns,’ International Review of Financial Analysis, 18(3): 95–100. M. Manasanet-Bataller, A. Pardo, E. Valor (2007) ‘CO2 prices, energy and weather,’ The Energy Journal, 28(3): 73–92. S.G. Mandis (2013) What Happened to Goldman Sachs: An Insider’s Story of Organizational Drift and Its Unintended Consequences. Harvard Business Review Press. E.A. Mannix (1993) ‘Organizations as resource dilemmas: the effects of power balance on coalition formation in small groups,’ Organizational Behavior and Human Decision Processes, 55: 1–22.

246

References

K.S. Markel, L.A. Barclay (2007) ‘The intersection of risk management and human resources: an illustration using genetic mapping,’ International Journal of Risk Assessment and Management, 7(3): 326–340. H.M. Markowitz (1952) ‘Portfolio selection,’ The Journal of Finance, 17(1): 77–91. H.M. Markowitz (1959) Portfolio Selection: Efficient Diversification of Investments. John Wiley & Sons (reprinted by Yale University Press, 1970). T.M. Mata, R.L. Smith, D.M. Young, C.A.V. Costa (2005) ‘Environmental analysis of gasoline blending components through their life cycle,’ Journal of Cleaner Production, 13(5): 517–523. C. McDonald (2009) New PRIMA president sees public RMs as masters of disaster. National Underwriter/Property & Casualty Risk & Benefits Management, 113(21): 17–31. D.B. McDonald (2011) ‘When risk management collides with enterprise sustainability,’ Journal of Leadership, Accountability and Ethics, 8(3): 56–66. MEI Computer Technology Group Inc. (2011) ‘2011 Trade Promotion Management Trends.’ R.C. Merton (1974) ‘On the pricing of corporate debt: the risk structure of interest rates,’ The Journal of Finance, 29(2): 449–470. D. Meyler, J.P. Stimpson, M.P. Cutghin (2007) ‘Landscapes of risk,’ Organization & Environment, 20(2): 204–212. I.I. Mitroff, M.C. Alpaslan (2003) ‘Preparing for evil,’ Harvard Business Review, 81(4): 109–115. R. Morelli, A. Tucker, N. Danner, T.R. de Lanerolle, H.J.C. Ellis, O. Izmirli, D. Krizanc, G. Parker (2009) ‘Revitalizing computing education through free and open source software for humanity,’ Communications of the ACM, 52(8): 67–75. A.S. Mukherjee (2008) The Spider’s Strategy: Creating Networks to Avert Crisis, Create Change, and Really Get Ahead. FT Press. P.K. Narayan, S. Narayan, A. Prasad (2008) ‘Understanding the oil price-exchange rate nexus for the Fiji islands,’ Energy Economics, 30(5): 2686–2696. D.B. Nelson (1991) ‘Conditional heteroskedasticity in asset returns: a new approach,’ Econometrica, 59: 347–370. D. Ng, P.D. Goldsmith (2010) ‘Bio energy entry timing from a resource based view and organizational ecology perspective,’ International Food & Agribusiness Management Review, 13(2): 69–100. W.D. Nordhaus (2001) ‘Climate change: global warming economics,’ Science, 294(5545): 1283–1284. W.D. Nordhaus, J.G. Boyer (1999) ‘Requiem for Kyoto: an economic analysis of the Kyoto protocol,’ The Energy Journal, 20: 93–130. D.L. Olson (2004) ‘Data set balancing,’ Lecture Notes in Computer Science: Data Mining and Knowledge Management, Y. Shi, W. Xu, & Z. Chen, eds. Springer, 71–80. D.L. Olson (2005) ‘Comparison of weights in TOPSIS models,’ Mathematical and Computer Modelling, 40: 721–727. D.L. Olson, D. Wu (2005) ‘Decision making with uncertainty and data mining,’ Advanced Data Mining and Applications: First International Conference, ADMA, X. Li, S. Wang, Z.Y. Dong eds, Lecture Notes in Artificial Intelligence. Keynote paper. Springer, 1–9. D.L. Olson, D. Wu (2006) ‘Simulation of fuzzy multiattribute models for grey relationships,’ European Journal of Operational Research, 175(1): 111–120. D.L. Olson, D. Wu (2008) Enterprise Risk Management. World Scientific. J. Pan, M. Wang, D. Li, J. Le (2009) ‘Automatic generation of seamline network using area Voronoi diagrams with overlap,’ IEEE Transactions on Geoscience and Remote Sensing, 47(6): 1737–1744.

References

247

M.S. Paolella, L. Taschini (2006) ‘An econometric analysis of emission trading allowances,’ Research Paper Series 06–26, FINRISK: National Center of Competence in Research Financial Valuation and Risk Management. A. Papalexandris, G. Ioannou, G. Prastacos, K.E. Soderquist (2005) ‘An integrated methodology for putting the balanced scorecard into action,’ European Management Journal, 23(2): 214–227. J. Paradi, S. Vela, H. Zhu (2010) ‘Adjusting for cultural differences, a new DEA model applied to a merged bank,’ Journal of Productivity Analysis, 33: 109–123. S. Patterson (2010) The Quants: How a New Breed of Math Whizzes Conquered Wall Street and Nearly Destroyed It. Crown Business. P.C. Pendharkar, J.A. Rodger (2003) ‘Technical efficiency-based selection of learning cases to improve forecasting accuracy of neural networks under monotonicity assumption,’ Decision Support Systems, 36(1): 117–136. Y. Peng (2000) Management Decision Analysis. Science Publication. C. Perrow (1984) Normal Accidents: Living with High-risk Technologies. Basic Books. C. Perrow (1999) Normal Accidents: Living with High-Risk Technologies. Princeton University Press. J.D. Piotroski, S. Srinivasan (2008) ‘Regulation and bonding: the Sarbanes-Oxley Act and the flow of international listings,’ Journal of Accounting Research, 46(2): 383–425. M.R. Powers, T.Y. Powers, S. Gao (2012) ‘Risk finance for catastrophe losses with Paretocalibrated Lévy-stable severities,’ Risk Analysis, 32(11): 1967–1977. S.C. Ray (2004) Data Envelopment Analysis: Theory and Techniques for Economics and Operations Research. Cambridge University Press, 189–208. J.E. Raymond, R.W. Rich (1997) ‘Oil and the macroeconomy: a Markov state-switching approach,’ Journal of Money, Credit and Banking, 29(2): 193–213. L.A. Reilly, O. Courtenay (2007) ‘Husbandry practices, badger sett density and habitat composition as risk factors for transient and persistent bovine tuberculosis on UK cattle farms,’ Preventive Veterinary Medicine, 80(2–3): 129–142. C.M. Reinhart, K.S. Rogoff (2008) ‘Is the 2007 subprime crisis so different? An international historical comparison,’ American Economic Review, 98(2): 339–344. B. Ritchie, C. Brindley (2007) ‘An emergent framework for supply chain risk management and performance measurement,’ Journal of the Operational Research Society, 58: 1398–1411. L. Rittenberg, F. Martens (2012) Enterprise Risk Management: Understanding and Communicating Risk Appetite. COSO. S.A. Ross, R.M. Westerfield, B.D. Jordan (2007) Corporate Finance Essentials. McGrawHill/Irwin. M. Saadatseresht, A. Mansourian, M. Taleai (2009) ‘Evacuation planning using multiobjective evolutionary optimization approach,’ European Journal of Operational Research, 198(1): 305–314. T.L. Saaty (2008) ‘Decision making with the analytic hierarchy process,’ International Journal of Services Sciences, 1(1): 83–98. F. Salmon (2009) ‘Recipe for disaster: the formula that killed Wall Street,’ Wired, 17(3). V. Sampath (2009) ‘The need for greater focus on nontraditional risks: the case of Northern Rock,’ Journal of Risk Management in Financial Institutions, 2(3): 301–305. N. Santella, L.J. Steinberg, K. Parks (2009) ‘Decision making for extreme events: modeling critical infrastructure interdependencies to aid mitigation and response planning,’ Review of Policy Research, 26(4): 409–422. M. Santiago (2011) ‘The Huasteca rain forest,’ Latin American Research Review, 46: 32–54.

248

References

D. Santin, F.J. Delgado, A. Valino (2004) ‘The measurement of technical efficiency: a neural network approach,’ Applied Economics, 36(6): 627–635. S. Scandizzo (2005) ‘Risk mapping and key risk indicators in operational risk management,’ Economic Notes by Banca Monte dei Paschi di Siena SpA, 34(2): 231–256. J. Seifert, M. Uhrig-Hombur, M. Wagner (2008) ‘Dynamic behavior of CO2 spot prices,’ Journal of Environmental Economics and Management, 56(2): 180–194. L.M. Seiford, J. Zhu (1999) ‘Profitability and marketability of the top 55 U.S commercial banks,’ Management Science, 45(9): 1270–1288. J. Seigel (2002) Stocks for the Long Run, 3rd ed. McGraw-Hill. C. Serrano-Cinca, Y. Fuertes-Calle’n, C. Mar-Molinero (2005) ‘Measuring DEA efficiency in Internet companies,’ Decision Support Systems, 38: 557–573. W.F. Sharpe (1964) ‘Capital asset prices: a theory of market equilibrium under conditions of risk,’ The Journal of Finance, 19(3): 425–442. W.F. Sharpe (1970) Portfolio Theory and Capital Markets. McGraw-Hill Book Company. J.W. Shavlik, R.J. Mooney, G.G. Towell (1991) ‘Symbolic and neural learning algorithms: an experimental comparison,’ Machine Learning, 6: 111–143. R. Shelp, A. Ehrbar (2009) Fallen Giant: The Amazing Story of Hank Greenberg and the History of AIG. Wiley. H.D. Sherman, F. Gold (1985) ‘Bank branch operating efficiency: evaluation with data envelopment analysis,’ Journal of Banking and Finance, 9(2): 297–316. Y. Shi, Y. Peng, G. Kou, Z. Chen (2006) ‘Classifying credit card accounts for business intelligence and decision making: a multiple-criteria quadratic programming approach,’ International Journal of Information Technology and Decision Making, 4: 1–19. H.S. Shin (2009) ‘Reflections on Northern Rock: the bank run that heralded the global financial crisis,’ Journal of Economic Perspectives, 23(1): 101–119. J. Sijm, S. Bakker, Y. Chen, H. Harmesen, W. Lise (2005) ‘CO2 price dynamics: the implications of EU emissions trading on the price of electricity,’ Report ECNC-05–81, Energy Research Center of the Netherlands (ECN). R.T. Silves, L.K. Comfort (2012) ‘The Exxon Valdez and BP Deepwater Horizon oil spills: reducing risk in socio-technical systems,’ American Behavioral Scientist, 56(1): 76–103. D.H. Smaltz, R. Carpenter, J. Saltz (2007) ‘Effective IT governance in healthcare organizations: a tale of two organizations,’ International Journal of Healthcare Technology and Management, 8(1/2): 20–41. J.C. Smith (2009) Pseudoscience and Extraordinary Claims of the Paranormal: A Critical Thinker. Wiley-Blackwell. ISBN 978–1405181228. J.A. Sniezek, D.R. May, J.E. Sawyer (1990), ‘Social uncertainty and interdependence: a study of resource allocation decision in groups,’ Organizational Behavior and Human Decision Processes, 46, 155–180. J. Sobehart, S. Keenan (2001) ‘Measuring Default Accurately,’ Credit Risk Special Report, Risk, 14: 31–33. A. Soteriou, S.A. Zenios (1999) ‘Operations, quality, and profitability in the provision of banking services,’ Management Science, 45(9): 1221–1238. U. Springer (2003) ‘The market for tradable GHG permits under the Kyoto Protocol: a survey of model studies,’ Energy Economics, 25: 527–551. U. Springer, M. Varilek (2004) ‘Estimating the price of tradable permits for greenhouse gas emissions in 2008–2012,’ Energy Policy, 32: 611–21. B. Stafford (2001) ‘Risk management and internet banking: what every banker needs to know,’ Community Banker, 10(2): 48–49.

References

249

J.N. Stanard, M.G. Wacek (1991) ‘The spiral in the catastrophe retrocessional market,’ Casualty Actuarial Society Discussion Paper, May, Arlington, VA. J.E. Stiglitz (2003) The Roaring Nineties: A New History of the World’s Most Prosperous Decade. W.W. Norton & Co. P.S. Sudarsanam, R.J. Taffler (1995) ‘Financial ratio proportionality and inter-temporal stability: an empirical analysis,’ Journal of Banking & Finance, 19(1): 45–60. G. Suder, D.W. Gillingham (2007) ‘Paradigms and paradoxes of agricultural risk governance,’ International Journal of Risk Assessment and Management, 7(3): 444–457. J.A.K. Suykens, T.V. Gestel, J.D. Brabanter. (2002) Least Squares Support Vector Machines. World Scientific Press. N. Taleb (2012) Antifragile: Things That Gain from Disorder. Random House. N.N. Taleb (2007) The Black Swan: The Impact of the Highly Improbable. Penguin Books. N.N. Taleb, D.G. Goldstein, M.W. Spitznagel (2009) ‘The six mistakes executives make in risk management,’ Harvard Business Review, 87(10): 78–81. W.-J. Tan, P. Enderwick (2006) ‘Managing threats in the global era: the impact and response to SARS,’ Thunderbird International Business Review, 48(4): 515–536. J. Taylor (2009) Getting Off Track: How Government Actions and Interventions Caused, Prolonged, and Worsened the Financial Crisis. Hoover Press. N. Taylor (2007) ‘A note on the importance of overnight information in risk management models,’ Journal of Banking & Finance, 31: 161–180. J. Tobin (1958) ‘Liquidity preference as behavior towards risk,’ The Review of Economic Studies, 25: 65–86. M.D. Troutt, A. Rai, A. Zhang (1995) ‘The potential use of DEA for credit applicant acceptance systems,’ Computers and Operations Research, 4: 405–408. H.C. Tsai, C.M. Chen, G.H. Tzeng (2006) The comparative productivity efficiency for global telecoms,’ International Journal of Production Economics, 103: 509–526. H. Tulkens (1993) ‘On FDH efficiency analysis: some methodological issues and applications to retail banking, courts and urban transit,’ Journal of Productivity Analysis, 4(1–2): 183–210. M. Uhrig-Homburg, M. Wagner (2006) ‘Success chances and optimal design of derivatives on CO2 emission certificates,’ Working Paper, University of Karlsruhe. L.V. Utikin (2007) ‘Risk analysis under partial prior information and nonmonotone utility functions,’ International Journal of Information Technology and Decision Making, 6, 625–647. V.E. Vaugirard (2003) ’Pricing catastrophe bonds by an arbitrage approach,’ The Quarterly Review of Economics and Finance, 43: 119–132. M.T. Vo (2009) ‘Regime-switching stochastic volatility: evidence from the crude oil market,’ Energy Economics, 31(5): 779–788. H. Von Blottnitz, M.A. Curran (2007) ‘A review of assessments conducted on bio-ethanol as a transportation fuel from a net energy, greenhouse gas, and environmental life cycle perspective,’ Journal of Cleaner Production, 15(7): 607–619. J. Von Neumann, O. Morgenstern (1944) Theory of Games and Economic Behaviour, 2nd ed. Princeton University Press. J.H. Von Thünen (1826) The Isolated State. H. Wagner (2004) ‘The use of credit scoring in the mortgage industry,’ Journal of Financial Services Marketing, 9(2): 179–183. C.H. Wang, R. Gopal, S. Zionts (1997) ‘Use of data envelopment analysis in assessing information technology impact on firm performance,’ Annals of Operations Research, 73: 191–213.

250 References

H.F. Wang, D.J. Hu (2005) ‘Comparison of SVM and LS-SVM for regression,’ International Conference on Neural Networks and Brain 2005, 1: 279–283, October 13–15, 2005. S.H. Wang (2003) ‘Adaptive non-parametric efficiency frontier analysis: a neural-network-based model,’ Computers & Operations Research, 30: 279–295. B. Watkins (2003) ‘Riding the wave of sentiment: an analysis of return consistency as a predictor of future returns,’ Journal of Behavioral Finance, 4(4): 191–200. J. Wei, D. Zhao, L. Liang (2009) ‘Estimating the growth models of news stories on disasters,’ Journal of the American Society for Information Science and Technology, 60(9): 1741–1755. D. Williamson (2007) ‘The COSO ERM framework: a critique from systems theory of management control,’ International Journal of Risk Assessment and Management, 7(8): 1089–1119. F.B. Wiseman (2013) Some Financial History Worth Reading: A Look at Credit, Real Estate, Investment Bubbles & Scams, and Global Economic Superpowers. Abcor Publishers. D. Wu (2006) ‘A note on DEA efficiency assessment using ideal point: an improvement of Wang and Luo’s model,’ Applied Mathematics and Computation, 2: 819–830. D.D. Wu (2009) ‘Performance evaluation: an integrated method using data envelopment analysis and fuzzy preference relations,’ European Journal of Operational Research, 194(1): 227–235. D.D. Wu (2014) ‘An approach for learning risk management: confucianism system and risk theory,’ International Journal of Financial Services Management. Accepted and in press. D.D. Wu, J.R. Birge (2012) ‘Serial chain merger evaluation model and application to mortgage banking,’ Decision Sciences, 43(1): 5–36. D. Wu, C. Luo, H. Wang, J.R. Birge (2014) ‘Bilevel programming merger evaluation and application to banking operations,’ Production and Operations Management. DOI: 10.1111/poms.12205. Accepted and in press. D. Wu, D.L. Olson (2006) ‘A TOPSIS data mining demonstration and application to credit scoring,’ International Journal of Data Warehousing & Mining, 2(3): 1–10. D. Wu, D.L. Olson (2009) ‘Introduction to the special section on optimizing risk management. Methods and Tools,’ Human and Ecological Risk Assessment, 15(2): 220–226. D.D. Wu, D.L. Olson (2010) ‘Enterprise risk management: coping with model risk in a large bank,’ Journal of the Operational Research Society, 61(2): 774–787. D. Wu, D.L. Olson (2010) ‘Enterprise risk management: coping with model risk in a large bank,’ Journal of the Operational Research Society, 61(2): 179–190. D. Wu, D.D. Wu (2010) ‘Performance evaluation and risk analysis of online banking service,’ Kybernetics, 39(5): 723–734. D. Wu, Z. Yang, L. Liang (2006) ‘Using DEA-neural network approach to evaluate branch efficiency of a large Canadian bank,’ Expert Systems with Applications, 31(1): 108–115. D. Wu, L. Zheng, D.L. Olson (2014) A Decision Support Approach for Online Stock Forum Sentiment Analysis. IEEE Transactions on Systems Man and Cybernetics. Accepted and in press. DOI: 10.1109/TSMC.2013.2295353. D. Wu, Y. Zhou (2010) ‘Catastrophe bond and risk modeling: a review and calibration using Chinese earthquake loss data,’ Human and Ecological Risk Assessment, 16(3): 510–523. J. Yao, Z. Li, K.W. Ng (2006) ‘Model risk in VaR estimation: an empirical study,’ International Journal of Information Technology and Decision Making, 5: 503–512. X. Yao, H. Yao (2000) An Introduction to Confucianism. Cambridge University Press. T.K. Zhelev (2005) ‘On the integrated management of industrial resources incorporating finances,’ Journal of Cleaner Production, 13(5): 469–474.

References

251

L. Zhu (2008) ‘Double exponential jump diffusion model for catastrophe bonds pricing,’ Journal of Fujian University of Technology, 6: 336–338 (in Chinese). R. Zolkos, M. Bradford (2011) ‘Risk management faulted in probe of BP disaster,’ Business Insurance, 45(36): 4–25.

Company websites Bank of America (2007) Annual Report 2007, available at www.rbs.com/microsites/ gra2007/downloads/RBS_GRA_2007.pdf. Barclays (2007) Annual Report 2007, available at www.barclaysannualreport.com/index. html. Chase (2007) Annual Report 2007, available at http://investor.shareholder.com/ common/. Citibank (2007) Annual Report 2007, available at www.citi.com/citi/fin/data/k07c.pdf. Dominion (2001) ‘Internet banking struggles for profits,’ available at www.stuff.co.nz/ inl/index/0,1008,779016a28,FF.html. HSBC (2007) Annual Report 2007, available at www.investis.com/reports/hsbc_ar_2007_ En/report.php?type=1. Jupiter Research.(2004) ‘FIND research, Institute for Information Industry,’ available at http://www.find.org.tw. Lloyds (2007) Annual Report 2007, available at www.investorrelations.lloydstsb.com/ media/pdf_irmc/ir/2007/2007_LTSB_Group_R&A.pdf. Royal Bank of Scotland (2007) Annual Report 2007, available at www.rbs.com/microsites/ gra2007/downloads/RBS_GRA_2007.pdf. SunTrust (2007) Annual Report 2007, available at www.suntrustenespanol.com/suntrust. Wachovia (2007) Annual Report 2007, available at www.wachovia.com/file/2007_ Wachovia_Annual_Report.pdf. Wells Fargo (2007) Annual Report 2007, available at www.wellsfargo.com/downloads/pdf/ invest_relations/wf2007annualreport.pdf.

Index accounting perspective, 3–7 Adelphia, 13 AIG, 2, 15, 29–30, 73 Air Canada, 147 allocative efficiency (AE), 61 Ameriquest, 26 Analytic Hierarchy Process (AHP), 58, 59 anchoring, 111 Arbitrage pricing theory, 110 Arthur Andersen, 13, 164 artificial neural networks (ANN), 32, 43, 48, 125, 127–134 asset price volatility, 32 autoregressive conditional heteroscedasticity (ARCH) model, 192, 199, 201 autoregressive moving average (ARMA) model, 201 autoregressive process, 121, 139 availability, 111 backpropagation neural networks, 127–128 bags-of-words, 33 balanced scorecard, 58, 60, 73–74 Bank Credit Scoring, 72–86 Bank Efficiency Analysis, 99–107, 124–135 Bank of America, 2, 100, 104 Banker, Charnes & Cooper (BCC) DEA model, 62, 67–70, 126, 156 Barclay’s, 100 Basel Accords, 16, 73, 106 Bayesian analysis, 121–122 Beta (book to market value), 66 bilevel programming, 145–162 binomial option pricing model, 110 Black, Scholes, and Merton, 24, 112 British Petroleum, 118–123 bubbles, 15, 24, 111–112

California electricity, 11, 12–13 Canadian Imperial Bank of Commerce (CIBS), 147 Capital Asset Pricing Model (CAPM), 16, 109, 110 Capital market instruments, 73 carbon emission pricing, 183–198 catastrophe bonds, 73, 136–144 catastrophe equity puts (cat-e-puts), 73 catastrophe risk instruments, 136–138 Charnes, Cooper & Rhodes DEA model, 62, 127, 156 Chase, 101 Chinese earthquakes, 1, 136, 137, 175 Citigroup, 100 closeness coefficient, 89 coincidence matrix, 94–95 collateralized debt obligations (CDOs), 16, 73, 108 collateralized mortgage obligations (CMOs), 21 Committee on Sponsoring Organizations (COSO), 3–4, 6, 7 Compound Poisson loss model, 138 Conditional-Value-at-Risk (CVaR), 18 Confucius three-stage learning, 215–220 contingent surplus notes, 73 copulas, 15, 20–21 COSO ERM Cube, 4 COSO framework, 3–4 COSO internal control process, 3 Countrywide, 26 credit default swaps (CDSs), 16, 21, 29, 30 credit rating, 74 credit scorecard, 75–84 credit scoring, 90–97 Critical Infrastructure Protection Decision Support System (CIPDSS), 180–181 crude oil, 199–214

253

254 Index

daily volatility model, 38–39 Data Envelopment Analysis (DEA), 57–71, 100, 124, 126–127, 149–162 data mining, 87–90 decision making unit (DMU), 61, 66, 70, 126, 131, 149, 150, 151 decision support system (DSS), 180–181 decision tree, 93–96 Deep Water Horizon, 118–123 derivatives, 24, 73 Distribution Free Approach, 100, 124 double marginalization, 151 DuPont model, 57 economic perspective, 108–117 Efficient Market Hypothesis, 23 efficient market theory, 109 emergency management, 179–180 emergency management support systems (EMSS), 180–181 energy risk, 167 Enron, 11–14, 15, 72, 163, 164 enterprise resource planning (ERP) systems, 14 Ericsson, 176 ERM process, 7–8 European Climate Exchange (ECX), 184, 189 Expected Utility Theory, 16 Exponential Generalized Autoregressive Conditional Heteroscedasticity (EGARCH) model, 193–194, 197, 202 Eyjafjallajokull, 168, 175 family regulation, 217–218 FEMA, 179 financial risk forecasting, 32–48 financial risk management, 15–22 financial statement analysis, 58, 60–61 Florida hurricanes, 1 food risk, 166 Fourier transformations, 17 framing, 111 free-disposal hull, 100, 124 Fuzzy set theory, 58, 59–60 Gaussian copula, 20, 108 Generalized Autoregressive Conditional Heteroscedasticity (GARCH) model,

33, 34–42, 48, 49, 55–56, 192–197, 199–200, 202, 206–210 globalization, 168–169, 171–173 Golden West, 26 Green Tree, 26 grey relation analysis, 58, 60 H1N1 virus, 2 hedging, 15, 27 herding, 111 HSBC, 100 Icelandic volcano, 1 ICTCLAS System, 51 implementation issues, 7–8 incentive incompatibility, 151 indifference theory, 110 information sentiment, 33–34 Innovative Support To Emergencies, Diseases And Disasters (InSTEDD), 179 International Organization for Standardization (ISO), 14 investment collars, 17 Kolmogorov-Smirnov statistic, 75, 77, 142 Kyoto protocol, 183, 184, 187 Lehman Brothers, 15, 30 Levy process, 17 Lexicon approach, 50 Li David X., 20, 108 Lloyd’s of London, 1, 72, 101, 104, 174 London Market Exchange (LMX), 112 Long Term Capital Management (LTCM), 15, 24, 112, 164 Lorenz curve, 75, 78 machine learning, 32–48, 50 Macondo, 118, 119, 120 malicious activities, 164 marginal abatement cost curve, 185 Markov chain Monte Carlo, 144 Markov process, 17 Markov regime switching model, 210–214 Markowitz, 16, 109 mean variance, 110 merger evaluation, 150–152

Index 255

Merrill Lynch, 2, 15 Minerals Management Service (MMS), 120, 121 Monte Carlo simulation, 19, 96, 110, 121, 140, 144, 209 Moody’s, 72 moral hazard, 115 mortgage system, 26 multiattribute utility, 181 Multivariate Statistical Analysis, 58, 106 mutualization, 115 National Disaster Medical System, 181 natural disasters, 164, 175–182 neural networks, see artificial neural networks Nokia, 176 Nord Pool, 184 Northern Rock, 25, 26–29 online banking, 99–107 options-pricing model, 110 overall efficiency (OE), 61 overconfidence, 111 Pacific Gas & Electric, 12 Pareto distribution, 17 part of speech (POS) tagging, 50 Peregrine Systems, 13 performance validation, 74–75 perturbation, 96 Philips Electronics, 176 polarity tagging, 49 PowerNext, 184 principal component analysis (PCA), 99–107 real estate crash of 2008, 23–31 real estate cycle, 25 regime switching models, 203 Regional Integrated Model of Climate and The Economy (RIMCE), 185 Reinhart and Rogoff, 23 return on assets (ROA), 63 return on equity (ROE), 63 risk analysis, 7, 165 risk appetite, 6 risk exchange swaps, 73 risk identification, 7

risk management definition, 2 risk management framework, 110 risk management modeling, 9 risk management responsibilities, 8 risk management theories, 111 risk mitigation, 114 risk scoring, 64 risk tolerance, 114 Royal Bank of Scotland, 101, 104 SAHANA, 179 San Diego Gas & Electric Company, 12 Sarbanes-Oxley Act, 3, 13–14 self cultivation, 216–217 semantic techniques, 32 sentiment analysis, 39–41, 44–48, 49–56 Sharpe, William, 109 squared correlation coefficient, 45 Stackelberg game, 148 Standard & Poor’s (S&P), 72 state harmonization, 218–219 state preference theory, 110 Stochastic frontier analysis, 100, 124, 156 Stock Forum, 49–51, 52, 56 stock price volatility, 54 subprime banking crisis, 2 SunTrust, 101 supply chains, 147–149, 169, 170–171 support vector machines (SVM), 32, 39, 41–42, 43, 48, 49, 55–56 sustainability, 163–174 sustainable risk, 166–168 Swine Flue epidemic, 136 systemic failures, 164 Taleb, N.N., 20, 113 technical efficiency (TE), 61 terrorism, 1 thick frontier approach, 100, 124 TOPSIS, 87–98 trading volume volatility, 48 tranches, 15, 21–22, 26 Treadway Commission, 3 tsunamis, 1, 164, 176 Tyco International, 13 underinvestment problem, 110 United Nations intergovernmental panel on climate change (UNIPCC), 183

256 Index

Value-at-risk (VaR), 15, 17–20, 31, 73, 121, 200 variable selection, 64–66 volatility forecasting model, 33–39, 200–214 volatility trend forecast accuracy, 45 volcanoes, 1

Wachovia, 101 weather derivatives, 73 Wells Fargo, 101 Wenchuan earthquake, 136, 137, 175 word segmentation, 50, 51 WorldCom, 13, 15, 72, 163, 164