Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Abstraction and Intermediation: The Implications of Behavioral Decision Making
for Three Applications in Finance
A dissertation submitted to the
Department of Social and Decision Sciences in partial fulfillment of the requirements
for the degree of
Doctor of Philosophy in Social and Decision Sciences
by
David C. Rode
Dissertation Committee:
Paul Fischbeck (Chair) Baruch Fischhoff
Jay Apt Eric Gold
Carnegie Mellon University Pittsburgh, Pennsylvania
October 2017
i
Table of Contents
TABLE OF CONTENTS ................................................................................................................................. i
TABLE OF EXHIBITS .................................................................................................................................. iii
ACKNOWLEDGEMENTS ........................................................................................................................... vi
PRÉCIS .......................................................................................................................................................... viii
INTRODUCTION ........................................................................................................................................ 1-1
LITERATURE REVIEW ............................................................................................................................ 2-1
2.1 ABSTRACTION ................................................................................................................................... 2-1 2.2 INTERMEDIATION .............................................................................................................................. 2-4 2.3 RESEARCH QUESTIONS & PREVIOUS WORK ...................................................................................... 2-6
2.3.1 Descriptive Models of Choice and Securities Design ............................................................. 2-6 2.3.2 Expert Judgment and Regulatory Decision-Making ............................................................... 2-8 2.3.3 Computational Modeling of Risky Intertemporal Choice ....................................................... 2-9 2.3.4 Presentation Compression and Complexity Reduction ......................................................... 2-12
CATASTROPHIC RISK AND SECURITIES DESIGN ........................................................................... 3-1
3.1 INTRODUCTION .................................................................................................................................. 3-1 3.2 THE PC INDUSTRY AND THE NATURE OF CATASTROPHIC RISKS ....................................................... 3-7
3.2.1 The Insurance Market ............................................................................................................. 3-8 3.2.2 Regulation ............................................................................................................................... 3-9 3.2.3 Intra-Industry Competition ..................................................................................................... 3-9 3.2.4 Definitions ............................................................................................................................. 3-11 3.2.5 Estimating the Risks of Catastrophic Events ......................................................................... 3-12
3.3 STRUCTURE OF CATASTROPHE BONDS ............................................................................................ 3-15 3.4 HUMAN BEHAVIOR AND CATASTROPHE BONDS .............................................................................. 3-19
3.4.1 Cognitive Complexity ............................................................................................................ 3-21 3.4.2 Exaggerated Comprehensiveness .......................................................................................... 3-23 3.4.3 Illusion of Control ................................................................................................................. 3-23 3.4.4 Reliance on Availability ........................................................................................................ 3-24 3.4.5 Overweighting Small Probabilities ....................................................................................... 3-24 3.4.6 Violations of Extensionality .................................................................................................. 3-25 3.4.7 Dimensions of Risk ................................................................................................................ 3-27 3.4.8 Asymmetric Information ........................................................................................................ 3-28
3.5 MARKET EQUILIBRATION ................................................................................................................ 3-29 3.5.1 Immature Market Structure ................................................................................................... 3-29 3.5.2 A Dual Equilibrium ............................................................................................................... 3-30
3.6 CONCLUSIONS AND IMPLICATIONS FOR SECURITIES DESIGN ........................................................... 3-32 3.7 EPILOGUE ........................................................................................................................................ 3-33
REGULATED EQUITY RETURNS: A PUZZLE .................................................................................... 4-1
4.1 INTRODUCTION .................................................................................................................................. 4-1 4.2 REGULATED EQUITY RETURNS AND THE CAPM ............................................................................... 4-3 4.3 REGULATED ELECTRIC UTILITY RETURNS ON EQUITY SINCE 1980 ................................................... 4-6
ii
4.3.1 Historical Authorized Return on Equity Data ......................................................................... 4-6 4.3.2 The Regulated Equity Premium............................................................................................... 4-9
4.4 POTENTIAL THEORETICAL EXPLANATIONS FOR THE PREMIUM ....................................................... 4-12 4.4.1 Capital Structure Effects ....................................................................................................... 4-12 4.4.2 Asset-Specific Risk ................................................................................................................ 4-14 4.4.3 The Market Risk Premium ..................................................................................................... 4-15 4.4.4 Testing a Theoretical Model of the Risk Premium ................................................................ 4-17
4.5 POSSIBLE IMPLICATIONS ................................................................................................................. 4-21 4.5.1 Potential Alternative Finance Explanations ......................................................................... 4-23 4.5.2 Potential Public Policy Explanations ................................................................................... 4-25 4.5.3 Potential Behavioral Economics Explanations ..................................................................... 4-27 4.5.4 Potential Public Choice Explanations .................................................................................. 4-29
4.6 CONCLUSION ................................................................................................................................... 4-31
A COMPUTATIONAL FRAMEWORK FOR RISKY INTERTEMPORAL CHOICE ....................... 5-1
5.1 INTRODUCTION .................................................................................................................................. 5-1 5.2 LITERATURE REVIEW ........................................................................................................................ 5-2
5.2.1 Choice Under Risk and Uncertainty ....................................................................................... 5-3 5.2.2 Choice Over Time ................................................................................................................... 5-4 5.2.3 Risky Intertemporal Choice ..................................................................................................... 5-5
5.3 SIMULATING THE POPULATION OF DECISION MAKERS ...................................................................... 5-7 5.3.1 Parameterization and Decision Demographics ...................................................................... 5-7 5.3.2 Decision Processes and Sequencing ..................................................................................... 5-17
5.4 ANALYSIS AND FINDINGS ................................................................................................................ 5-18 5.4.1 Incorporating Probabilistic Information into Decision Problems ........................................ 5-20 5.4.2 Evaluating Novel and Complex Problems ............................................................................. 5-23
5.5 APPLICATIONS AND IMPLICATIONS .................................................................................................. 5-34
PRESENTATION COMPRESSION: INVESTMENT METRICS & HEURISTICS ............................ 6-1
6.1 INTRODUCTION .................................................................................................................................. 6-1 6.2 THE LEVELIZED COST OF ENERGY IN THEORY AND PRACTICE .......................................................... 6-3
6.2.1 The Levelized Cost of Energy .................................................................................................. 6-3 6.2.2 Levelized Cost vs. Levelized Cash Flow .................................................................................. 6-6
6.3 THE DECISION PROBLEM ................................................................................................................... 6-7 6.3.1 Decision Problem Structure .................................................................................................. 6-11
6.4 ANALYSIS AND FINDINGS ................................................................................................................ 6-14 6.4.1 Power Investments as Intertemporal Lotteries ...................................................................... 6-14 6.4.2 Basic Prospect Structures and Design Elements................................................................... 6-15 6.4.3 Compression and Different Prospect Designs ...................................................................... 6-21 6.4.4 Compression and a Heterogeneous Population of Decision Makers .................................... 6-33
6.5 APPLICATIONS AND IMPLICATIONS .................................................................................................. 6-39 6.5.1 Compression as a Tool for Debiasing ................................................................................... 6-42
FINAL THOUGHTS & FUTURE DIRECTIONS .................................................................................... 7-1
7.1 SECURITIES DESIGN........................................................................................................................... 7-1 7.2 REGULATORY INTERMEDIATION........................................................................................................ 7-3 7.3 TOWARD A COMPUTATIONAL APPROACH TO BEHAVIORAL DECISION THEORY ................................ 7-7 7.4 PRESENTATION COMPRESSION IN RISKY INTERTEMPORAL CHOICE ................................................... 7-8
REFERENCES ............................................................................................................................................. R-1
iii
Table of Exhibits
Exhibit A: Catastrophe-Risk Financial Instruments as of 1997 ......................................... 3-5 Exhibit B: Three Examples of Early Catastrophe Bond Offerings .................................... 3-7 Exhibit C: Natural Catastrophes in 1995, 1996, and 1997 ............................................... 3-13 Exhibit D: Contingent Surplus Note (CSN) Cash Flow Structure ................................... 3-15 Exhibit E: Catastrophe Insurance Index Basis Comparisons ........................................... 3-17 Exhibit F: GCCI versus PCS Coverage in Index-based Contracts .................................. 3-18 Exhibit G: Behavioral Anomalies and Implications for Catastrophe Bonds ................... 3-20 Exhibit H: The St. Paul Re Pro Rata Capital Note Cash-Flow Structure ........................ 3-22 Exhibit I: Comparing and Contrasting Cat Bonds with Conventional Securities ............ 3-26 Exhibit J: Development Scenario for ART Products ....................................................... 3-30 Exhibit K: Actual vs. Projected Issuance of Cat Bonds ................................................... 3-34 Exhibit L: Contemporary Standardized Cat Bond Structure............................................ 3-34 Exhibit M: Qualifying Rate Cases Filed per State for the Ten Most Frequently- and Least
Frequently-Filing States, 1980-2015 .......................................................................... 4-7 Exhibit N: Risk Premium Growth by Frequency of Case Filing. Gaps in the series reflect
years in which no rate cases were filed for the subject group .................................... 4-7 Exhibit O: Range of Risk Premium Growth Across States. States with highest and lowest
rates of growth (among states with at least 5 rate cases) are highlighted .................. 4-8 Exhibit P: Filing Frequency by Year ................................................................................. 4-9 Exhibit Q: Authorized Return on Equity vs U.S. Treasury and Investment Grade Corporate
Bond Rates ............................................................................................................... 4-10 Exhibit R: Distribution of Premium Across All Years .................................................... 4-10 Exhibit S: Authorized Return on Equity Premium, 1980-2015 ....................................... 4-11 Exhibit T: Authorized Rate-of-Return Premium vs. Utility Leverage ............................. 4-13 Exhibit U: Authorized Rate-of-Return Premium vs. the Hamada Capital Structure
Parameter ................................................................................................................. 4-14 Exhibit V: Authorized Rate-of-Return Premium vs. Industry Average Asset Beta ........ 4-15 Exhibit W: Market Risk Premium Trends over Time by Historical Window ................. 4-16 Exhibit X: Authorized Rate-of-Return Premium vs. Ex Ante Estimated Market Risk
Premium ................................................................................................................... 4-17 Exhibit Y: Regression Results for CAPM-based Risk Premium Model.......................... 4-19 Exhibit Z: Actual vs. Regression-Model Risk Premium Spreads .................................... 4-19 Exhibit AA: Regression Results for a Two-Period CAPM-based Risk Premium Model 4-20 Exhibit BB: Actual vs. Two-Period Model-Predicted Risk Premium Spreads ................ 4-21 Exhibit CC: Wholesale Fuel and Power Price Trends, 2007-2015 .................................. 4-22 Exhibit DD: Wholesale vs. Retail Power Prices, 2007-2015 ........................................... 4-23 Exhibit EE: Comparability of Spreads Measured with Authorized and Earned Rates of
Return and Utility Net Income ................................................................................. 4-25 Exhibit FF: Authorized Rates of Return on Equity and Skewness .................................. 4-28 Exhibit GG: Rate of Return Authorized as a Percent of Rate of Return Requested ........ 4-30
iv
Exhibit HH: Summary of Studies Estimating Cumulative Prospect Theory Parameters .. 5-8 Exhibit II: Parameter Estimates by Country from Rieger, Wang, and Hens [2017] .......... 5-9 Exhibit JJ: Empirical Distributions of Cumulative Prospect Theory Parameters ............ 5-10 Exhibit KK: Descriptive Statistics and Correlations for the Cumulative Prospect Theory
Parameters ................................................................................................................ 5-11 Exhibit LL: Parameter Space for the Value Function ...................................................... 5-12 Exhibit MM: Parameter Space for the Probability Weighting Function (Gains) ............. 5-12 Exhibit NN: Parameter Space for the Probability Weighting Function (Losses) ............ 5-13 Exhibit OO: Time Discounting Model Hierarchy ............................................................ 5-15 Exhibit PP: Distribution of Hyperbolic Present Bias Parameters .................................... 5-16 Exhibit QQ: Summary of Decision Maker Parameter Bounds ........................................ 5-16 Exhibit RR: Basic Taxonomy of Decision Processes ...................................................... 5-17 Exhibit SS: Probabilistic Representation of the Probability Weighting Function for Gains
for the Simulated Population of Decision Makers ................................................... 5-21 Exhibit TT: Probabilistic Representation of the Probability Weighting Function for Losses
for the Simulated Population of Decision Makers ................................................... 5-21 Exhibit UU: Probabilistic Representation of the Value Function for the Simulated
Population of Decision Makers ................................................................................ 5-22 Exhibit VV: Population Decision Demographics for a Choice Problem ......................... 5-22 Exhibit WW: Percent of Population Exhibiting Certainty-Risk Asymmetry .................. 5-25 Exhibit XX: Percent of Population Exhibiting Certainty-Risk Asymmetry with State 1
Adjustment at Zero, Using “Uncertainty First” Decision Sequence, and Varying Decision Process ...................................................................................................... 5-26
Exhibit YY: Logistic Regression Results for Certainty-Risk Asymmetry. Case used is p = 0.5, Δ = 3, and = 0 for HDPT, Uncertainty First. ................................................ 5-27
Exhibit ZZ: Percent of Population Exhibiting Short-Long Asymmetry .......................... 5-29 Exhibit AAA: Classification Tree Analysis Illustrating Quasi-Complete Separation. Case
used is p = 0.05, Δ = 10, t = 1, and Premium = 100 for HDPT, Uncertainty First. . 5-31 Exhibit BBB: Bayesian Logistic Regression Results for Short-Long Asymmetry. Case used
is p = 0.05, Δ = 10, t = 1, and Premium = 100 for HDPT, Uncertainty First. .......... 5-32 Exhibit CCC: Bayesian Logistic Regression Results for Short-Long Asymmetry. Case used
is p = 0.05, Δ = 10, t = 1, and Premium = 100 for HDPT, Time First. .................... 5-33 Exhibit DDD: Diagram of Choice Rule and Presentation Relations ................................. 6-8 Exhibit EEE: Presentation-Heuristic Pairing Agreement Classification ......................... 6-13 Exhibit FFF: Full and Compressed Presentation Forms .................................................. 6-13 Exhibit GGG: Typical Power Project Cash Flow Derivation .......................................... 6-15 Exhibit HHH: Four Example Cash Flow Profiles ............................................................ 6-16 Exhibit III: Four Power Plant Classes and Class-Specific Information ........................... 6-17 Exhibit JJJ: Capacity-Factor Levels for Each Type of Power Plant Included ................. 6-18 Exhibit KKK: Capacity Factor and Energy Price Multiplier Curves ............................... 6-19 Exhibit LLL: Energy-Price Trajectories with Multipliers Relative to Average Price ..... 6-20 Exhibit MMM: Comparison of Two Stylized Projects with Different Lifespans for which
Compression is a Useful Heuristic ........................................................................... 6-24 Exhibit NNN: Interpretation of Hypothesis Tables ......................................................... 6-24 Exhibit OOO: Relative Life Hypotheses and Short-Long Asymmetry ........................... 6-26
v
Exhibit PPP: Population Classification According to the Relative Lifespans of the Prospects, Measured as Longer Lifespan Divided by Shorter Lifespan .................. 6-26
Exhibit QQQ: Cash Flow Framing Hypotheses and Gain-Loss Asymmetry .................. 6-27 Exhibit RRR: Population Classification According to Cash Flow Profile Framing ........ 6-27 Exhibit SSS: Incentive Hypotheses and Short-Long Asymmetry ................................... 6-28 Exhibit TTT: Population Classification According to Use of Incentives ........................ 6-28 Exhibit UUU: Project Type Hypothesis and Gain-Loss and Magnitude Asymmetries ... 6-29 Exhibit VVV: Population Classification Across vs. Within Project Types ..................... 6-29 Exhibit WWW: Refining the Population Classification Across Project Types by
Controlling for Renewable Power Types ................................................................. 6-31 Exhibit XXX: Influence of Prospect Characteristics on Decision Problems with
Compression ............................................................................................................. 6-32 Exhibit YYY: Percentage Variation of Subgroup Mean Coefficient Value from Population
Coefficient Means .................................................................................................... 6-34 Exhibit ZZZ: Example of De Minimis Impact on from Further Dividing the Group ... 6-34 Exhibit AAAA: Connection of Decision Maker Characteristics ( ) to Project
Characteristics for the Simplifying and Useful Outcome Classes ........................... 6-36 Exhibit BBBB: Connection of Decision Maker Characteristics ( ) to Project
Characteristics for All Outcome Classes.................................................................. 6-37 Exhibit CCCC: Connection of Decision Maker Characteristics ( ) to Project
Characteristics for All Outcome Classes.................................................................. 6-38 Exhibit DDDD: Information Consumer Choice Strategy Outcomes ............................... 6-43 Exhibit EEEE: Heuristic Map of Change-in-Choice Consequences ................................ 6-46 Exhibit FFFF: Rule Accuracy by Prospect and Decision Maker Characteristics ............ 6-47 Exhibit GGGG: Gross Margin Trends in California, Pennsylvania, and Texas ................ 7-4 Exhibit HHHH: Regression Results Examining Commission Reputation on Spreads ...... 7-6
vi
Acknowledgements
Forsan et haec olim meminisse iuvabit Virgil (The Aeneid)
Although perhaps Aeneas’s journey was longer, both it and this process have stretched to epic proportions and have involved adventures with a myriad of characters. These acknowledgements do little justice to the number or influence of that group. I can but therefore offer, with deep humility, my gratitude to the following people, who must serve as proxies for the larger, unnamed cast.
There is easily no person who has had a greater influence on my arrival at this moment than Paul Fischbeck. I owe Paul thanks for seeding my path with interesting problems and clearing the roadblocks so that I could—eventually—reach my destination. He has been a trusted advisor, an influential mentor, a favorite research co-conspirator, and a good friend. Paul provided me with opportunities that literally changed my life (and my views on hyphenation) and did so while making the whole process (or maybe 95% of it) fun.
I am grateful, also, to my committee members: Baruch Fischhoff, Jay Apt, and Eric Gold. Baruch was officially my introduction to SDS, having called to inform me that I was admitted many, many years ago. I was fortunate to have been his TA during my first semester in the department and to have benefitted from his influence on my qualifying exam committee. Despite our seemingly different interests and approaches, I have been impressed frequently over the years—such is the breadth of his influence—by the number of times Baruch has said something or I have read something he’s written that has struck a chord, resonated, and provided the spark that lit the way down an interesting path. Jay has been an encouraging supporter of mine since his installation as director of CEIC, and has generously provided me numerous opportunities to share my work with CEIC and its advisory board over the years. With Eric, my connection to SDS spans the entire history of the department. Eric was SDS’s first graduate student, and defended the year I arrived in the department. To each of them, I am deeply appreciative of their support, their insight, and their patience.
Two teachers from my distant past must also be acknowledged, as my arrival at this moment is due in part to their influence. The late Jim Robinson was my Latin and English teacher in high school. A Renaissance man in the truest sense, his gift to me was a classical education. It was only with age that I have come to see, and continue to see, the expansiveness of this gift, as items he planted in my head decades ago appear with surprising regularity. I have a wider, richer, and more profound appreciation for our Western inheritance than I could have imagined three decades ago, and he is to thank for it. Subsequently, Jim Laing, now emeritus at Wharton, changed my thinking as an undergraduate. Yes, finance has always remained of central interest to me, but Jim Laing
vii
was responsible for my introduction to decision sciences, to behavioral decision theory, to game theory, to social choice theory, and to evolutionary computation methods. He is the one who inspired me to add Decision Processes as a second concentration. He is the one who suggested that perhaps considering graduate school would be of interest. He is the one who encouraged the finance practitioner in my head to make room for the researcher, and I am better for it. I truly would not be who I am today without their influence and support.
I was grateful to have a job that, for 15½ out of 16 years, was deeply satisfying. That was in no small part due to it providing long relationships with valued colleagues. Tony, Alex, Jen, Pete, and Lisa made for a great team over the years, and I was lucky enough to get to know and appreciate their families as well. Of the many decisions I made throughout that time, bringing on Tony and Alex were two of the best. I have been (and continue to be) so proud of them both. I couldn’t have asked for better colleagues or friends.
I am not—let’s be honest—the easiest person to get to know. I am thus grateful to a small, loyal group whose friendship has meant so much to me: my very long-time friends, Robert Rissman and John Forté, with whom I share a history of unmatched duration; my Pittsburgh family, Lisie (my “spark”) and Jeff Lipsitz; my soulmate, John Nulty; and last, but certainly not least, the Waldorf to my Statler, the Liz to my Karen, and the one who got away, Tim Zinn. Kuckleffel
There has been no greater surprise in my life than my discovery that I am so deeply in love with two wonderful kids. Tony and Sarah Páez could have given me no greater gift than sharing their wonderful children, Claire and Antonio, with me. They stole my heart early on and never gave it back. No one brightens my day the way they can. And no one is ever so excited to see me (and I them). They lift my heart. Claire and Antonio: I love you, I love you, I love you! Never forget that. I hope I can help to provide you both with bright futures as partial repayment for how each of you has added to the substance of my life.
Finally, I must offer my thanks to my family, which has supported my endeavors, whichever direction they took me. I have benefited from a mother and father whose support and encouragement have been complete and unconditional. I only wish my father could have lived to see my various graduations. I know he would have been proud of me, but they are still moments I’d have liked to have shared with him. I owe my sister thanks as well, for putting up with a curmudgeonly brother and a polar opposite. She’s seen the good, the bad, and the ugly, and yet she still talks to me. It is only with the passage of so much time that I have come to appreciate the entirety of my family’s legacy in the traits and behaviors and culture for which they were (and are) the guardians. I am certainly myself, but I have become what I am through their collective influence. And I am glad of it.
To each of these people, and to those unnamed behind them, I offer my profound thanks.
Remember me, but forget my fate Dido’s Lament (Henry Purcell, Dido & Aeneas)
viii
Précis Modern financial theory and practice make extensive use of the concepts of
abstraction and intermediation. Securities are designed, valued, and allocated within
portfolios based on characteristics that contain little of the richness of the actual assets
underlying their claims. Corporate financial decisions and policy analyses evaluate choices
based on comparing abstract metrics that ignore or minimize the myriad attributes belonging
to such choices. In many cases, these evaluations and choices are made by intermediaries
(whether human experts or heuristic rules) that either offer specialized information or serve
as a proxy for competitive markets. While economic theory has addressed issues in
abstraction and intermediation that arise from incentive problems and information
asymmetries directly, most approaches have maintained the traditional assumption of
rational agents free of cognitive constraints and biases. This dissertation illustrates, through
three applications, that abstraction and intermediation are at significant risk of suboptimal
operation when faced with certain limitations prevalent in the behavioral decision-making
literature.
The three diverse applications include the design of financial securities, the rate-
making practices of utility regulators, and the use of metrics and heuristics by intermediaries
in corporate finance and policy analysis. Well-established areas in the behavioral decision-
making literature, such as risk perception, communication about uncertainty, and heuristic
choice and expert judgment are shown to challenge the application of these core principles
of economics in practice. This dissertation develops a methodology for evaluation and
demonstrates the practical consequences of behavioral decision-making in each of the three
applications and examines possible responses and interventions to mitigate their effect.
1-1
1 Introduction
The Oxford English Dictionary defines “abstraction” as “the process of considering
something independently of its associations or attributes.” In finance, “intermediation”
generally refers to the role played by banks or brokers acting as a bridge between investors
and companies or lenders and borrowers [Allen and Santomero, 1997]. In a similar vein, in
regulatory economics, regulators are tasked with standing in for competitive market forces
in instances where market failures are likely or the government has permitted a monopoly to
exist. Although modern economic theory has come to see intermediation as an issue
primarily of information asymmetries in a principal-agent setting (e.g., Diamond [1984]),
Schumpeter, had a more expansive definition:
“…the banker must not only know what the transaction is which he is asked to finance and how it is likely to turn out but he must also know the customer, his business and even his private habits, and get, by frequently ‘talking things over with him’, a clear picture of the situation.” [Schumpeter, 1939: p. 116]
That is, successful intermediation involves a richness of detail regarding not just the
information itself, but how it is communicated and how it is received by transaction
stakeholders who are subject to cognitive constraints and behavioral biases. These two
concepts—abstraction and intermediation—play an enormous role in modern financial
theory. They also pose significant challenges, however, to a world faced with the numerous
findings of the behavioral decision-making literature.
Abstraction can, in theory, allow diverse securities to be priced using common
theoretical constructs built up from the treatment of Arrow-Debreu securities—highly-
stylized state-contingent claims—that contain none of the richness of actual securities, which
reflect specific firms and are targeted at specific clienteles. Intermediation, or the placing of
a person or rule in between an actor and an investment or policy, is commonplace in
corporate finance and regulation, where various metrics (e.g., Net Present Value (“NPV”),
1-2
Levelized Cost of Electricity1 (“LCOE”)) serve as proxies for projects and policies, and
regulators are called upon to take the place of competitive markets. But metrics, like heuristic
decision rules in general, are imperfect, and regulators can be subject to biases in judgment.
Given the importance of these concepts to modern financial practice, it is important to
identify what implications the behavioral decision-making literature has for them.
In this dissertation, I explore three specific areas in finance that represent archetypal
examples of abstraction and intermediation, I develop a method for exploring decision-
making behavior in novel environments, and I illustrate the challenges to their practical
application introduced by widely-known behavioral decision-making findings.
The first essay addresses abstraction in the context of securities design. Finance
theory generally, and structured finance practice in particular, operate on the premise that
cash flows can be extensively manipulated and “reconfigured” to exploit clientele effects
(or, as Merton [1989] described it, the bundling and unbundling of risks). Behavioral
decision theory suggests that there are limits to such reconfiguration, in particular when
involving high consequence, low probability events, and when violations of extensionality
and the multidimensional nature of perceived risk confound risky choice. My first essay
“Catastrophic Risk and Securities Design,” published in the Journal of Psychology and
Financial Markets2, analyzes a novel financial instrument (so-called “cat bonds”3) by
examining the complications that arise from attempting through abstraction to design
securities for investor clienteles without acknowledging their cognitive limitations and
biases in dealing with the richness of detail actually present in real examples of such
securities.
1 The LCOE is a calculation that spreads the fixed and variable costs of generating electricity from a particular source evenly over a period of time, such that the NPV of the actual cash flows and the NPV of the “levelized” cash flows are equal. In this formulation, as is traditional in finance, uncertainties are intended to be captured in the selection of a discount rate. 2 Now the Journal of Behavioral Finance. 3 A catastrophe (or “cat”) bond is a security issued by an insurance company whose principal and/or interest is reduced in the event of a pre-defined catastrophic loss (e.g., hurricane, earthquake).
1-3
The second essay addresses intermediation in the context of regulation. Regulators
are often called upon to stand in for the discipline of a competitive market in setting rates of
return for regulated industries. Governments, for example, have granted monopolies (or
“franchises”) to electric utilities to prevent costly duplication of services [Kahn, 1970]. In
return for the grant of an exclusive franchise, however, utilities are subjected to rate
regulation by regulators that seek to impose the discipline of competitive markets (by
limiting the rates that utilities are allowed to charge to those deemed to be competitive or
“just and reasonable”) in spite of their de jure monopoly status. Regulators seek to estimate,
therefore, the rate of return that such a business would earn if it were not granted a monopoly,
and therefore if its investors were compensated solely on the basis of the risks they would
assume under competitive operation. Although financial theory provides clear guidance in
this area in the form of normative asset pricing models, the evidence involving electric
utilities suggests that the judgment of regulators in this area deviates from those outcomes
and also from the decisions of regulators that default to formulaic approaches. In particular,
a “money illusion” effect appears to be present.4 The second essay uses a unique data set of
electric utility rate cases to test competing hypotheses about the existence of regulatory
biases and finds that the normative finance model is deficient in explaining observed rate-
setting behavior and that utilities appear to be earning excess returns as a result.
The third essay develops a method for exploring decision-making behavior in novel
or complex environments under uncertainty and over time. Behavioral decision-making
research has made extensive use of experiments with human subjects in identifying a variety
of anomalies regarding risky intertemporal choice. These models, at a very general level,
have tended to converge to a small set of functional forms, separated only by varying
parameterization. There remains a difficulty, however, in applying these models and their
various parameterizations to many complex and domain-specific problems. Human subject
pools have typically involved students, and more recently have involved online labor
markets, such as Amazon.com’s Mechanical Turk (“MTurk”). As such, many experimental
tasks must be reduced to very generic or “abstract” levels, resulting in a general difficulty in
4 In behavioral economics, the “money illusion” refers to the tendency of people to misperceive nominal price changes as real price changes.
1-4
exploring many complex, or domain-specific decision tasks. The third essay develops a
computational framework for extending behavioral decision-making research in risky
intertemporal choice into such domains that is based on the general decision-making
characteristics identified in the literature.
The fourth essay addresses both abstraction and intermediation in the use of metrics
by heuristic investment decision makers and policymakers. Finance theory operates on the
basis that certain decision problem presentations I shall refer to as “metrics” (such as NPV
or LCOE) are extensionally equivalent for choice purposes to the cash flows from which
they are derived. For example, Project A is preferred to Project B if NPV(Project A) >
NPV(Project B) or Policy X is preferred to Policy Y if LCOE(Policy X) < LCOE(Policy Y).
But these metrics are the product of “lossy compression.” Information is “lost” or abstracted
away in the process of “compressing” project details to create the metric. In particular,
lotteries are replaced with expected values (compression across risk) and events unfolding
over time are replaced with present values (compression across time). This lost information
is deemed to have no normative value beyond what is captured in the compressed forms.
This information may, however, contain attributes that influence actual decision making.
One such metric, the Levelized Cost of Energy, is commonly used to represent
competing choices in policy analysis (e.g., for ranking tasks, such as determining the optimal
response to environmental regulation [Bruckner et al., 2014; Anderson et al., 2016]),
regulatory economics, forecasting [EIA, 2016], and even investment decision-making (such
as determining the preferred long-term capital investment strategy for regulated utilities
[USAID, 2009; Tucson Electric Power, 2016]). LCOE is perceived as a normative means of
spreading the cost (or value) of a project over its estimated life. Notwithstanding its
normative positioning, Joskow’s [2011] critique of LCOE focused on its exclusion of certain
practical considerations (such as time-of-use) for some technology types. Restated in
different terms: Joskow argues that the use of LCOE as a heuristic could lead to preference
reversals because the compression process omitted information that did, in fact, have
normative relevance. That is, it is flawed as a normative measure.
1-5
The premise of the fourth essay has an analogous basis: the act of compression that
creates LCOE can lead to a violation of extensionality in that decision makers make different
choices than they would if presented with the uncompressed information. If the use of LCOE
as a presentation form in certain environments impairs decision making (in the sense that
compressing a project cash flow profile triggers heuristics that bias the choices made by
decision makers), then there is value in being able to identify those environments and
understand how such biases can be mitigated. However, if the information lost to
compression would otherwise have itself triggered a bias, then its removal in the process of
levelizing may actually improve decision making. Although this work identifies classes of
project types where the use of LCOE leads heuristics to become unreliable, it also illustrates
the value of LCOE as a debiasing technique for cognitively-constrained decision makers. In
many cases, having less detailed information results in such decision makers making better
choices.
In Chapter 2, I review the literature on abstraction and intermediation in finance in
the context of relevant previous work in the behavioral decision-making literature and
outline the research questions examined herein. In Chapter 3, I present the first application,
which involves securities design for risks involving catastrophic insurance losses. In Chapter
4, I present the second application, which is an empirical examination of electric utility rate
case data between 1980 and 2015. In Chapter 5, I develop a computational approach to
investigate novel decision-making environments for which reliance on human subjects may
be problematic. In Chapter 6, I present the third application, which uses the approach
developed in Chapter 5 to identify classes of decision problems and types of decision makers
for which compression through the use of LCOE creates violations of extensionality in
choice and to illustrate the usefulness of compression as a debiasing technique for
cognitively-constrained decision makers. In Chapter 7, I provide concluding thoughts on the
three applications and discuss suggestions for future research in this area.
2-1
2 Literature Review 2.1 Abstraction
Financial theory and practice have often proceeded along two parallel tracks. While
considering a security or an investment abstractly, or independent of its associations and
attributes has been at the core of financial theory, innovation in the design of financial
securities has increased steadily over time. The process of valuation may be based on
abstraction, but Wall Street sells (and buys) descriptive richness.
The theory of general equilibrium was built up in part from Arrow-Debreu securities,
which are state-contingent claims that pay one unit of the numeraire commodity in one state
of the world and zero in all other states [Arrow, 1964]. The dynamic spanning of these
securities ensures market completeness and the existence of an equilibrium [Arrow and
Debreu, 1954]. As such, these securities abstract away from corporate details, projects,
government policies, investor preferences, and other descriptive factors. This lack of
descriptive richness continued into modern portfolio theory, as Markowitz [1952] and Tobin
[1958] showed that the von Neumann-Morgenstern [1947] axioms were consistent with
mean-variance preferences if utility functions were quadratic. That is, if investors cared only
about the mean and variance of wealth (or returns), then optimal portfolios could be
determined. But if quadratic utility admits only mean and variance preferences, then no other
attributes could be influential to decision makers (at least within the von Neumann-
Morgenstern axiomatization).
The consequences of this development of portfolio theory were that securities were
reduced to a probabilistic characterization of their expected return, its variance, and its
covariance with other returns. Tobin [1958] further extended this result by demonstrating
what is now known as one-fund separation: that all investors should hold a combination of
the market portfolio and the riskless asset. In such a world, investors need only hold an index
2-2
fund and need not concern themselves with the makeup of the underlying index or the nature
of the companies composing it. Abstraction provides simplicity. Security analysis, or the
evaluation of individual investments based on their unique characteristics, is largely
irrelevant in such an environment [Fama, 1970; Malkiel, 1973].5 The preeminence of
abstraction continued into the world of corporate finance as well. Projects could be reduced
to a simple metric (NPV) and decisions made solely based on that metric were shown to
produce optimal choices in the sense of maximizing firm value [Hirshleifer, 1958]. Discount
rates could be determined by simple formulas (such as the Capital Asset Pricing Model
(“CAPM”)) that would result in asset market clearing [Sharpe, 1964; Lintner, 1965; Mossin,
1966].
The securities we observe in the real world, however, look nothing like the theoretical
Arrow-Debreu securities. Over time, more complex and descriptively rich securities have
emerged, prompting a search for an explanation of this emergence. In an early and
exhaustive study of historical security issuance, Dewing [1934: pp 236-237] noted that our
basic security concepts—debt and equity—arose from distinctions in Anglo-Saxon law
regarding debtors and creditors. But even among those two classes, Dewing catalogued a
diverse array of variations and hybrid securities. Although many of the securities he
identified at the time eventually fell out of use, financial innovation resulted in a steady
stream of new security types over time [Allen and Gale, 1994].6
The process of “securitization,” or “structured finance,” that began in the late 1970s
has only accelerated these trends. If cash flows and risks are fungible, they can be—in
theory—almost infinitely bundled and unbundled [Merton, 1989]. To put this evolution in
context, the total market capitalization of all global stock markets is approximately $69
trillion (World Federation of Exchanges, 2015) and the total face value of all outstanding
public and private debt is approximately $223 trillion (ING, 2013). However, the total
5 But see the paradox created by this assumption identified by Grossman and Stiglitz [1980]. They show that it is technically impossible for markets to be informationally efficient if no one does security analysis. 6 Many early securities fell prey to changes in tax law or regulation. Others, such as “guaranteed stock” and “participating bonds,” fell out of common usage because they were unattractive hybrids of other existing securities [Dewing, 1934].
2-3
notional value of all derivative securities (that is structured securities derived from
underlying equity and/or debt holdings) is more than $550 trillion (Bank for International
Settlements, 2015). The collective value of “innovated” securities is nearly twice that of all
“basic” securities (on which they are theoretically based). Clearly, there is a strong demand
for financial innovation. This demand exists to adapt to changing regulations (e.g., changing
definitions of bank capital), to create markets for risks that were previously not easily traded
(e.g., cat bonds), and to attract new sources of capital by attempting to appeal to investors
not previously interested (e.g., market-linked certificates of deposit).
This returns us to the two parallel tracks of finance and their reconciliation. The
descriptive sparseness of asset pricing theory and the creative efficiency of financial
engineering have given rise, ironically, to a diverse ecosystem of complex securities. The
ease with which many securities, on paper, can be modified and redesigned has resulted in
their proliferation. In light of market frictions and in seeking to appeal to as many investor
clienteles as possible, this process is reasonable and (largely) harmless. For many structured
securities, their replicability from simpler securities provides arbitrage bounds on pricing
and liquidity because they can always be “unbundled” into simpler instruments. Many
options can be decomposed into stocks or bonds and cash [Black and Scholes, 1973], zero-
coupon bonds can be created from coupons “stripped” from traditional government bonds
[Fabozzi and Fabozzi, 1995], and index funds can be synthetically created by acquiring all
of the index components.
Many other structured securities, however, have no directly-traded constituent
components. In such cases, the assembly of these instruments from their abstract components
may work on paper, but are not guaranteed of price or liquidity in practice. In those cases,
the structuring entities seek to maintain a matched book of buyers and sellers, but imbalances
in those matched books can create sudden and dramatic swings in price and liquidity (see
Salmon [2009] for an example related to mortgage-backed securities and the financial crisis
of 2007-2009). These securities, then, depend crucially on the ability of investors to
understand their characteristics, to perceive their risks, and to interpret communications
2-4
about them clearly and “rationally.” Abstraction, absent arbitrage, creates a direct reliance
on investor rationality—a subject to which we shall return in Section 2.3 below.
2.2 Intermediation
The classical model of intermediation was that of the banker, who collected deposits
from customers and lent those deposits out to businesses. The banker assumed the credit risk
and profited from scale economies and informational asymmetries by charging the borrower
more than the banker paid the depositors. It was this world that Schumpeter had in mind in
the description of intermediation that I cited previously.
This model was about resolving informational asymmetries [Diamond, 1984] and
minimizing transaction costs [Gurley and Shaw, 1960] in a world of imperfect markets.
Frictions in the form of transaction costs could be minimized by an intermediary dividing
fixed costs over individuals. Frictions caused by information asymmetries could be resolved
by intermediaries with specialized knowledge that were subjected to “delegated monitoring”
in a principal-agent setting.
Since that original research, however, transaction costs have fallen dramatically and
large amounts of information are increasingly available. Contrary to those theories, however,
the quantity of, and activity by, intermediaries has increased substantially. As a result, there
has been a significant re-thinking of intermediation theory in finance. Merton [1989] and
Allen and Santomero [1997] highlight two previously overlooked functions of
intermediaries: risk transfer and complexity reduction. These functions involve the ability to
evaluate risks, the willingness to accept and manage risks, and the ability to communicate
risks and complex structures to less-informed market participants. In many respects, the
increased complexity borne of structured finance and discussed in Section 2.1 has created a
demand for more and more intermediation.
These demands have also expanded the role of intermediaries far beyond the
commercial banks identified nearly a century ago. Allen and Santomero [1997] in particular
2-5
note the significant growth in non-bank intermediaries, such as insurance companies,
pension funds, investment banks, advisors and consultants, and institutional investors and
asset managers. I focus on this expanded treatment to consider as an intermediary “any
evaluator or rule-processor that stands between actors.” This definition is purposefully
cumbersome so as to include non-human intermediaries, such as automated rules.
Let us consider the two functions advanced by Allen and Santomero [1997]. First,
the risk transfer function can take two forms: risk minimization and risk allocation. The role
of an intermediary can be to allow smaller investors or lenders to diversify their exposures
across multiple investments at low cost. Even corporations may wish to diversify
investments across multiple parties so as to avoid the costs of financial distress [Warner,
1977] or purely out of managerial self-interest [Ross, 1977; Stulz, 1984; DeMarzo and
Duffie, 1995]. The second form, risk allocation, is that of intermediary as risk “traffic cop.”
Once risk has been minimized, a decision as to who is best able to bear the remaining risk
must be made. This is a function performed, in my interest here, by regulators who are tasked
with both providing utilities (for example) a reasonable profit so as “to preserve the financial
soundness of the utility” [Bluefield Water Works & Improvement Co. v. Public Service
Commission of West Virginia (262 U.S. 679 (1923))] and also “protecting the public interest”
[Federal Power Commission v. Hope Natural Gas Company, 320 U.S. 591 (1944)]. This
twin mandate requires regulators to apportion costs and risks of certain investments to the
various stakeholders involved as they act as proxies for the discipline of competitive markets
in mediating between firms and consumers.
The second function advanced by Allen and Santomero [1997], complexity
reduction, can also be divided into two forms: delegation and simplification. The traditional
role of intermediaries involved delegation. In other words, complexity was addressed by
delegating decisions entirely to parties with specialized information or abilities [Diamond,
1984]. Complexity reduction can also be achieved by intermediaries that simplify the
decisions to be made. This is what may be considered the “satisficing” approach [Simon,
1947; 1956] and involves choices informed by (simple) heuristic rules [Simon, 1979].
2-6
2.3 Research Questions & Previous Work
In both abstraction and intermediation in finance, there is a tension between
simplification and “complexification,” between theoretical purity and practical relevance.
As Simon [1979] noted nearly forty years ago in his Nobel Prize lecture: “[d]ecision makers
can satisfice either by finding optimum solutions for a simplified world, or by finding
satisfactory solutions for a more realistic world.” My interest is in the latter: how can
financial decision-making be improved while maintaining a practical level of descriptive
richness? I explore three questions with real-world applications:
1. How can descriptive models of choice be used to improve securities design
and guide financial innovation?
2. How can the expert judgment of regulators be improved in the performance
of their role as risk-allocating mediators?
3. How does problem presentation affect the performance of heuristics used in
risky intertemporal choice?
In each of these questions, the tension is manifest in the struggle to find a balance between
seeking missing (or hidden) information that would make for “better” (and different)
decisions (either normatively better or better justified or communicated) and suppressing
information that would make decision-making “worse” because of biases or cognitive
limitations.
2.3.1 Descriptive Models of Choice and Securities Design
Often, financial securities are designed exclusively to meet the needs of the issuer or
seller. However, if insufficient attention is given to the preferences of potential buyers, the
issuer may not realize the level of funding anticipated, may face a higher cost, or after-market
liquidity may be nonexistent. While not every security will appeal to every investor, I focus
here on two other issues: (i) potential investors may not understand the novel security, and
(ii) potential investors may not believe the information they’re given about the novel
security.
2-7
As I noted previously, the design of new securities has progressively become more
complex over time. There is an abundant literature on the challenges of decision-making in
novel and complex environments [Simon et al., 1987]. In such environments, individuals
tend to default to heuristic reasoning and automated “rule-based” responses [Newell and
Simon, 1972; Cross, 1983; Albers and Laing, 1991; Becker, 1993]. They may misapply these
decision schemata because they disagree even about what the attendant risks are [Fischhoff,
1985]. They may suffer from cognitive dissonance when presented with normatively optimal
choices that appear “unpalatable” [Fisher and Statman, 1997].
Koonce, McAnally, and Mercer [2005] provide experimental evidence concerning
how investors actually judge the risks of financial instruments, showing that additional
explanatory power is provided by incorporating “behavioral” variables in addition to what
they term “decision-theory” variables (information about probabilities and payoffs). In
attempting to reason about complex investment decisions, even trained experts have
exhibited high rates of requests for useless information [Kroll, Levi, and Rapoport, 1988].
This finding is often confounded by regulatory requirements (e.g., the Security and
Exchange Commission’s Regulation FD) for increased disclosure of information, resulting
in many firms providing “kitchen sink” disclosures out of an abundance of caution. Lo,
Mamaysky, and Wang [2000] also note that the jargon and “patois” common to securities
disclosures often distort meaning and can produce disagreement even when the intent of the
disclosures is the same. MacGregor [1989] demonstrated (in reference to consumer
products) that inferences about risks were linked closely to interpretations of words and
phrases contained in product warnings. All of these factors challenge potential investors’
understanding of novel securities, and that lack of understanding may either dissuade them
from investing, or prompt investment under mistaken premises.
Even for those investors who do understand the securities in question, they may not
trust the information they’re given, or may simply have different estimates of the risks and
values involved. If additional discounting is undertaken by prospective investors to
compensate for such doubts or differences of belief, the result is likely to be a bid-ask spread
2-8
that is sufficiently wide so as to limit liquidity or lead to a breakdown in trading altogether.
Duffie and Rahi [1995] have noted that even the perception of large information
asymmetries can lead investors to reject certain investments, feeling that they would be
hopelessly disadvantaged. But trust issues may not arise solely from intentional
asymmetries. They can also emerge from different estimates of the problem parameters.
Biases in judgment have been exhibited when dealing with probabilities, including the
overweighting of small probabilities [Kahneman and Tversky, 1979] and overestimating
probabilities involving especially “vivid” or relatively publicized events [Tversky and
Kahneman, 1973; Lichtenstein et al., 1978; Combs and Slovic, 1979].7 Even the values
involved may be perceived differently. Slovic [1964] and Fischhoff et al. [1978] have
illustrated that perceptions of risk may vary widely from the single-factor ideals so common
in finance.
As a result of these findings, the presumption of understanding should be questioned
even among domain experts, and care should be taken not only in the provision of
information, but also in how such information is communicated to potential investors. A
central question in this line of inquiry is whether or not the provision of additional or
different information leads to better decisions. Would investors make “better” decisions
(about novel securities, for example) if they were provided with greater information about
probabilities and diverse risks in an attempt at debiasing, or when they are not distracted by
the additional information provided and focus instead on a simplified presentation?
2.3.2 Expert Judgment and Regulatory Decision-Making
Regulators, despite having domain expertise, are not immune from many of the
decision-making biases identified above. Slovic [1972] noted that even with accurate models
and good judgment, experts are unable to apply what they know consistently. Imperfect
expert judgment has also been shown to permit suboptimal choices in difficult and
7 For example, attempts to sell contingent convertible notes (known as “CoCos” or “bail-in bonds”), which are used to bolter bank regulatory capital positions, often follow well-publicized concerns about bank solvency. The disproportionate availability of information about banking systems risks may lead potential investors to overestimate the risks associated with a particular bank’s issuance.
2-9
consequential tasks [Dawes, Faust, and Meehl, 1989; Mowen, 1994; Thaler, 1994; Kammen
and Hassenzahl, 1999]. Dawes [1979], in particular, has demonstrated that even simple
linear models tend to outperform experts on issues of “clinical intuition.”
One of the activities that regulators are tasked with performing is setting rates of
return for regulated industries. Their twin mandate is to provide for rates that are sufficient
to maintain the utility in a financially-sound state, while also protecting the public interest
from excessive costs. In the U.S., rates of return have typically been set in nominal terms.
For example, a utility may be granted the approval to charge rates intended to provide a
return on equity of 10%. The underlying cost of that utility’s capital, however, varies not
only with the perceived risk of the firm, but also macroeconomic conditions, including
inflation. In the 1970s and early 1980s, many utilities suffered from “regulatory lag” because
regulators were unable to adjust their nominally-denominated authorized rates of return
quickly enough to accommodate rising inflation.
This problem is an example of money illusion, or the tendency to misperceive
nominal price changes as real price changes [Shafir, Diamond, and Tversky, 1997]. Thaler
and Tversky [1996] note that the presence of money illusion can also impact risk aversion.
Kahneman, Knetsch, and Thaler [1986] show that judgments of fairness are based on
evaluation of nominal (as opposed to real) changes as well. Because evaluation of risk and
fairness are both solidly within the purview of regulators, there is concern that decisions by
regulators may be subject to such biases in a way that results in suboptimal regulatory
decision-making—a consequence that may be exceedingly costly for consumers and/or
firms. In this situation, the public interest may be better served by advocating for the use of
heuristic rules by regulators, in lieu of detailed evaluation and fact-finding during rate cases.
2.3.3 Computational Modeling of Risky Intertemporal Choice
Almost all the major findings in the behavioral decision-making literature have
emerged from experiments with human subjects. These experiments, often out of necessity,
have typically involved highly-stylized decision-making tasks. Although “natural”
2-10
experiments have been conducted when feasible (for example, Lichtenstein and Slovic
[1973] studied actual behavior in casinos, Odean [1999] examined actual brokerage records
for individual investors, and Chetty et al. [2014] studied labor market and retirement saving
behavior using detailed household balance sheet and employment data), the ability of
researchers to test decision making experimentally in complex or novel environments has
been limited.
Notwithstanding these limitations, a large and robust literature on human decision-
making behavior under uncertainty and over time has emerged. The behavioral decision-
making literature has established a collection of anomalies and alternative theories of risky
choice [Kahneman and Tversky, 1979; Tversky and Kahneman, 1992] and of intertemporal
choice [Thaler, 1981; Loewenstein, 1988; Kirby and Maraković, 1995; Read, 2001]
supported by extensive experimental results. These alternative theories have been codified
into certain generally-accepted stylized models of decision making. Further, a significant
effort has been made to estimate the parameters of these models and evaluate differences
between various populations of decision makers.
Most of the experimental work, however, has been carried out using college students
and, more recently, using online labor markets such as MTurk. Scholars have rightly
questioned the generalizability of results obtained from these populations. Although for
many generic decision tasks, these groups are representative of the population as a whole,
there are also clear exceptions. Shuptrine [1975] compared experimental results from student
to those from actual consumers and found the results “inconclusive,” “strongly suggest[ing]
that investigators attempt in every way possible to test the population that they are interested
in studying.” Similarly, Peterson [2001] concluded that “caution must be exercised” when
attempting to extrapolate results to non-student populations. More recently, Goodman,
Cryder, and Cheema [2013] noted that MTurk participants exhibit many similarities to the
general population, but have important differences as well. In particular, they are “less likely
to pay attention to experimental materials.”
2-11
Even samples of experts or professionals exhibit behavior not perfectly aligned with
the general population. Albert [1967] noted that businesspeople differ “significantly” in
experimental research. Shanteau [1988] said that the experimental evidence for expert
decision making is weak and mixed; that experts were equally prone to certain errors, but
also often used better decision processes.
The notion that different groups tend to exhibit different behaviors is not necessarily
surprising. However, it does present a complicating factor in attempting to evaluate decision
making in complex or novel environments. Tasks that require domain-specific knowledge
or expert reasoning, for example, can be difficult to examine both because general
populations may not have such knowledge and because of the difficulty or cost in accessing
statistically-significant groups possessing such knowledge or experience. It is, I believe,
equally difficult to draw meaningful inferences from overly-simplified tasks given to more
readily-available populations that are intended to represent some complex or novel “real-
world” task.
One such environment is the combination of choice under uncertainty and choice
over time: risky intertemporal choice. Although the normative approach is essentially settled
with the recursive utility framework of Kreps and Porteus [1978], as extended by Epstein
and Zin [1989], the descriptive theory of risky intertemporal choice is “not settled” [Albrecht
and Weber, 1997], is “complex and not easily understood” [Weber and Chapman, 2005], is
“relatively limited” [Anderson and Stafford, 2009], and is “quite heterogeneous in methods
and findings” [Hardisty and Pfeffer, 2017].
There are four general types of anomalies in risky intertemporal choice that have
been identified:
1. Certainty-Risk Asymmetry (discount rates are higher for certain outcomes than for risky outcomes)
2. Short-Long Asymmetry (violations of stationarity, preference reversals induced by shifting risky choices forward in time)
3. Gain-Loss Asymmetry (discount rates are smaller for gains than for losses) 4. Magnitude Asymmetry (discount rates are higher for larger payoffs)
2-12
The experimental work examining their findings, however, is limited and often
contradictory. Albrecht and Weber [1997] find support for short-long asymmetry in
matching tasks, but not choice tasks. Öncüler and Onay [2009] support violations of
stationarity, but find the evidence on direction of the effect mixed. Keren and Roelofsma
[1995] find support for certainty-risk asymmetry. Anderson and Stafford [2009] find no
support for certainty-risk asymmetry. Shelley [1994] supports gain-loss asymmetry;
Blackburn and El-Deredy [2013] do not. Disagreements involving the interactions between
risk and time emerge from whether subjects evaluate risk first or time first [Weber and
Chapman, 2005] and whether there is a single underlying mechanism [Keren and Roelofsma,
1995] or many mechanisms [Read, 2003]. Albrecht and Weber [1997] note that the empirical
results are sensitive to the elicitation procedure used.
Mindful of these challenges, the objective, then, is to adapt the knowledge and insight
already developed to such an environment. One means by which to do so is to develop a
computational model of a population of decision makers and perform “virtual experiments”
on that population. The role of computational simulation in the social sciences is now well-
accepted (see Epstein and Axtell [1996] and Gaylord and D’Andria [1998] for early
examples). Applying computational simulation methods to behavioral decision-making
questions, however, has been relatively unexplored, but may provide a platform for more
extensive exploration of this complex area.
2.3.4 Presentation Compression and Complexity Reduction
The use of heuristic rules as a means of complexity reduction has been widely studied
for decades [Simon, 1955; Simon, 1956; Polya, 1957; Gigerenzer and Todd, 1999]. Apart
from the problem of insufficient information, decision makers can also suffer from having
to process too much information or too many choices [Camerer, Loewenstein, and Weber,
1989; Kleinmuntz and Schkade, 1993; Iyengar and Lepper, 1999], useless or irrelevant
information [Kroll, Levi, and Rapoport, 1988; Redelmeier, Shafir, and Aujla, 2001], or
useful information absent a context (even a normatively irrelevant context) [Dawes, 1999].
2-13
One objective, in such situations, is to seek to focus the decision maker on the most
relevant information, and then present that information in a format that allows for clear
interpretation. But presentation itself can be a confounding factor. Funtowicz and Ravetz
[1987] write at length about the importance of how quantitative information is presented, if
one wishes the recipient to have a useful appreciation of its strengths and weaknesses and
context, and propose a paradigm for presenting such information.
The degree to which decision tasks can be simplified has been studied in a variety of
contexts. Axtell [1992] wrote extensively about model aggregation as a means of simplifying
complex models. The literature in reduced-form modeling has encompassed numerous areas
including climate modeling [Mendelsohn, Nordhaus, and Shaw, 1994], population dynamics
[Lutz et al., 2002], environmental planning [Webster et al., 2002], and credit risk modeling
[Duffie and Singleton, 2003]. One form of simplifying presentation may be considered
“compression.” Compression refers to the collapsing of information about a decision task
into a single number. For example, in finance many decision problems use compression in
the form of metrics (such as NPV and LCOE) that collapse future cash flows into present
values and uncertainties into expected values. Such compressed presentations are deemed to
be extensionally equivalent to the full presentations in theory.
However, in many settings, such actions are shown to exhibit violations of
extensionality in practice. The non-linear weighting of probabilities, for example, has been
shown to result in preference reversals not present in normative theory [Kahneman and
Tversky, 1979]. Likewise, the time-inconsistent discounting of outcomes, such that one’s
future self would not make the same choice that one would make in the future today, has
also been shown to result in preference reversals [Thaler, 1981]. As a result, financial
metrics—which are forms of decision problem presentation—that collapse across these
dimensions (risk and time) may either (or both) bias or aid decision making.
With regard to practical application, decision makers that succumb to biases related
to risk and time may be benefitted from using a metric that in effect “bypasses” such biases
2-14
by employing a simpler presentation. That is, by focusing the decision maker on the most
relevant information, “less” (information) may actually be “more” (useful). As a result, use
of the LCOE presentation could play an important role as a debiasing technique for
cognitively-constrained decision makers. However, users of the LCOE presentation must
also understand where the opposite effect may occur and biases are introduced by the process
of compression. Joskow [2011] has shown that LCOE is “not a useful way to compare
generating technologies [that have] different production profiles” (e.g., baseload vs.
intermittent) and thus is normatively flawed.
A consequent line of inquiry here asks whether LCOE’s flaws in a normative setting
are more than offset by its benefits in mitigating the impact of biases in a descriptive setting.
Compression, for example, obscures the presence of risk. Certainty-risk asymmetry results
in decision makers applying higher discount rates to certain outcomes. Likewise,
compression explicitly “levelizes” cash flows over time, and short-long asymmetry results
in discount rates declining as events are pushed into the future. Although Joskow’s [2011]
recommendation was to “abandon” levelized cost comparisons, a less drastic response may
be warranted if classes of problem characteristics and decision maker characteristics where
the use of LCOE leads heuristics to become unreliable can be identified, so that its use can
be limited to those larger areas where its reliability is preserved. In the end: how reliable are
heuristics based on compression?
3-1
3 Catastrophic Risk and Securities Design
3.1 Introduction†
Insuring catastrophic risks represents a significant challenge not only for insurance
companies, but also for government agencies, banks, financial institutions, and individuals.
Traditionally, risk management has been handled by a network of insurance companies, with
some government oversight. A policy holder purchases coverage for some event from a
property and casualty (PC) insurer. The PC insurer then typically divests much of that risk
by obtaining coverage from reinsurers. The reinsurers transact with one another in order to
diversify their own exposures.
When the risk of loss from any individual event is both well-known and small, this
system works quite well (e.g., automobile and health risks). Historically, default rates among
insurers have been like those for corporations in general.8 However, with catastrophic losses,
the consequences are more concentrated and the probability of occurrence harder to assess.
These problems have drawn increasing attention as catastrophic losses have increased in
their number and scope. In some cases, like riverine floods, human activities (e.g., land-use
changes) have increased the number of severe events. In other cases, like earthquakes and
hurricanes, more people and insured property are in harm’s way, even if the number of events
has remained the same. Despite advances in hazard prediction, the models are still quite
† A version of this paper appears as Rode, D., B. Fischhoff, and P. Fischbeck. Catastrophic Risk and Securities Design. Journal of Psychology and Financial Markets 1:2 (2000): 111-126. Thanks go to John Miller, Carter Butts, and an anonymous reviewer for their thoughtful comments on earlier drafts, and to Hadi Dowlatabadi and James Risbey for their assistance in providing background information on climate modeling. The authors also appreciate the valuable comments and suggestions of Steven Goldberg and other participants at the NBER Insurance Project. Partial support was provided by the NSF-sponsored Center for Integrated Assessment of Human Dimensions of Global Change. All errors remain our own. 8 Matthews et al. [1999] report that failure frequency in the property and casualty insurance industry has averaged roughly less than 1% per year since 1969 and never exceeded 2.5%. Over the entire period that they studied, only 8% of insolvencies were attributed to catastrophic losses. At the same time, they report that in the period from 1989 to 1993, the percentage of defaults attributable to catastrophic losses increases from 8% to 56% of all defaults.
3-2
imperfect, particularly for limited time periods and geographic areas.9 Helbling, Fallegger,
and Hill [1996] also note increasing tendencies towards litigation, decreasing burdens of
proof in determining liability, and new regulations regarding old, pre-existing risks as
additional reasons for the increasing costs of catastrophic losses. All of these trends
complicate predicting insurance exposures.
Other uncertainties arise from the structure of the insurance industry itself. Cummins
and Doherty [1997] assess the ability of insurers to pay for the “Big One” and find that,
while recent losses of $10 billion to $15 billion may seem manageable compared to the
industry’s total capital of over $300 billion, there is a mismatch in the distribution of that
capital (across firms) and the distribution of claims following a catastrophic loss. In this
context, the capacity of individual insurers is actually quite limited. According to Haag
[1995], it is “unusual” for any insurer to obtain more than $100 million in catastrophe
reinsurance per policy. If insurers wanted more, they would find that worldwide reinsurance
capacity (circa 1994) is approximately $7.2 billion, much smaller than the possible demand.
Claims resulting from some single major catastrophes could reach $70 billion to $100 billion
[Palm, 1995; Wharton Alumni Magazine, 1998]. A repeat of the earthquake that destroyed
Tokyo in 1923 could result in damages of between $900 billion and $1.4 trillion [Valery,
1995].
The gap between the catastrophic coverage that the industry could and does provide
might be traced to failures in the market for coverage. One familiar problem is the cyclical
nature of the insurance business, reflecting its profit incentives. Some analysts have
attributed the laggard performance of insurance and reinsurance firms, relative to other
financial companies and the market as a whole, to “excess capital” and “underleveraging”
9 As Jenkins [1998] notes, there are problems combining the forecasts of global and regional climate models. Global climate models lack the resolution of the regional models in predicting changes to specific areas (such as coastlines and mountain ranges). Although regional models can accommodate such predictions, they depend on global models for input. Jenkins [1998] writes that “problems arise when interpolating to increase the number of data points [to move from one scale to another]. . . . [C]oupled global and regional models can not [sic] give the best results yet because coupled atmosphere-ocean models which provide boundary conditions for the regional climate model are still early in their development.” In essence, the “grave concerns” arise from the newness and extraordinary complexity of most climate models built on a scale useful (to insurers or insurance investors) for hazard prediction.
3-3
[Moody’s, 1997; Standard & Poor’s, 1997; Wehrly and Friedheim, 1998]. However, in fact,
the net written premium-to-surplus ratio declined from 1.82:1 in 1984 to 1.13:1 in 1995 to
0.90:1 at year-end 1997 [Wehrly and Friedheim, 1998]. Regulators allow insurers to
leverage their capital twice (2:1), in terms of which, the industry had “excess” capital of over
$120 billion. Although the premium-to-surplus ratio is only one measure of capitalization
efficiency, its decline suggests that surplus is accumulating at a greater rate than premiums
are increasing. That is, firms are generating more cash than they can efficiently manage.10
The resultant pressure on policy prices has diminished profits. In 1997, industry surplus
stood at $308.1 billion, up 20.1% from the previous year.
The definition of “excess” capital depends on the time frame used. Given the
certainty of eventual catastrophic losses, the capital may not be excessive. Rather, it may
eventually be required to cover losses. However, those losses may be decades into the future.
Only from a short-term perspective is such retained capital excess. Accounting conventions
and the U.S. tax code limit reserves to actual losses or to those losses that may reasonably
be expected within a year [USAA, 1998]. Any additional premium capital set-aside (e.g., for
longer term, less frequent losses) must be transferred to the insurer’s balance sheet, hence is
subject to tax (and possible distribution to shareholders). Liquid capital on an insurer’s
balance sheet reduces the firm’s return on equity. In the long run, this reduced return is
“fair,” in the sense of covering shareholders against expected future losses. However,
myopic [Mossin, 1968] or self-serving [Babcock and Loewenstein, 1997] managers may see
short-term advantage in reallocating that capital to more profitable short-term projects, such
as writing additional insurance coverage in non-catastrophic lines or returning it to
shareholders. By reducing provisions for losses, they will thereby increase future financial
risk. Shareholders should then expect increased returns as compensation for bearing more
risk (even if that risk is overlooked by current shareholders and managers).
Some smaller firms are underreserved, especially those with asbestos and
environmental liabilities [Standard & Poor’s, 1997]. However, the industry as a whole
10 Of course, this situation could be transiently justifiable after a disaster when firms raise premiums to address a decline in surplus. A fuller account would look at multiple measures of financial health.
3-4
appears to face a capitalization paradox: it is over-capitalized from a short-term profit
perspective, while lacking the long-term ability to cover catastrophic losses when they occur.
The paradox could be resolved if the industry made better use of its short-term capacity. As
mentioned, the infrequent nature of catastrophic claims means that insurers who prepare
adequately will suffer long periods of perceived excess capital (in the eyes of myopic
stakeholders) as well as diminished profits (from taxation of capital held against future
liabilities, but not treated as official reserves). Conversely, those who do not prepare
adequately have higher short-term profits, along with higher likelihoods of failure
(bankruptcy). Large catastrophic losses have periodically prompted reorganizations and the
tightening of capital. Then, for some (usually brief) period, perceived excess capital
disappears, prices increase, and profits remain stable. However, as capital flows back to the
industry, firms compete for profits, prices drop, competition intensifies, profits disappear,
and the industry finds itself awash in capital (which it cannot invest in projects profitable
enough to satisfy shareholders and managers) [Standard & Poor’s, 1997; Wehrly and
Friedheim, 1998].
Several recent initiatives have attempted to increase the industry’s capacity by
providing insurers access to capital markets [Lewis and Davis, 1998; Osterland, 1998].
Losses of even catastrophic proportion remain almost negligible, compared to the size of
global capital markets [Cummins and Doherty, 1997; Jaffee and Russell, 1997]. For
example, a $100 billion catastrophe would consume nearly one third of the PC industry’s
total capital and surplus, perhaps pushing some firms into insolvency. Yet, that amount is
less than the average daily variation in global equity wealth.
Catastrophe bonds (“cat” bonds) are one possible way to access capital markets,
building on the wild popularity of securitization for other asset classes and exposures (e.g.,
mortgages, credit card receivables, real estate, David Bowie11). However, although
11 In 1997, rock star David Bowie issued $55 million worth of bonds backed by future royalties from 25 of his albums. The bonds, which were rated AAA by Moody’s, were purchased by the Prudential Insurance Company of America and provide a return of 7.9% over 10 years. Although experts were quick to dismiss the issue as “a glamour investment” and claim that they would only be “attractive to people who want to associate themselves with show-business personalities,” the entire issue was in fact, purchased by a major insurance company [Wall Street Journal, 1998]. However, recent reports [Financial Times, 1999] suggest that the success of the “Bowie
3-5
insurance companies have created a variety of offerings (summarized in Exhibit A), they
have generally experienced but modest success in raising the desired capital at a cost
commensurate with the portfolio risk of the security [Lewis and Davis, 1998].
Reinsurer Instrument Amount (in millions) Status Date
Nationwide CSN $400.0 Closed August 1995
Hannover Re CSN $100.0 Closed 1995
Arkwright CSN $100.0 Closed May 1996
AIG Cat Bond $25.0 Closed May 1996
CAT, Ltd. Cat Bond $50.0 Withdrawn 1996
ACE, Ltd. Cat Bond $35.0 Withdrawn 1996
USAA Cat Bond $500.0 Withdrawn August 1996
Calif. Earthq. Auth. Cat Bond $1,500.0 Withdrawn 1996
RLI CatEPut $50.0 Closed October 1996
Hannover Re Cat Bond $100.0 Closed December 1996
St. Paul Re Pro Rata Bond $68.5 Closed December 1996
Winterthur Cat Bond $282.0 Closed January 1997
Reliance Cat Bond $40.0 Closed April 1997
Horace Mann CatEPut $100.0 Closed April 1997
USAA Cat Bond $477.0 Closed June 1997
Swiss Re Cat Bond $137.0 Closed July 1997
LaSalle Re CatEPut $100.0 Closed August 1997
Tokio-Marine Cat Bond $100.0 Closed November 1997
Exhibit A: Catastrophe-Risk Financial Instruments as of 1997 Source: Lewis and Davis [1998]
A frequently voiced objection by insurance companies is that such structured finance
costs too much for them to use it [Penalva-Zuasti, 1997]. For example, catastrophe bonds
have typically carried premiums of 300 to 500 (or more) basis points over the LIBOR for
medium-term securities with investment-grade ratings (see Exhibit B below). Even with
Bonds” may have been unique. Apart from various structural impediments in the music business (few recording artists own their own master tapes), the illiquid aftermarket for such securities raises many of the same questions discussed here in the context of catastrophe bonds [Silverman, Sparks, and Osterland, 1998].
3-6
these added inducements, buyers have sometimes been scarce. As a result, the critical
question facing the creators of this market is whether potential investors are reluctant to
purchase any security backed by catastrophic risk insurance or just the current offerings. In
an excellent overview, Froot [1997] advances eight possible reasons for the appearance of
insufficient risk sharing:
(1) Actuarially insufficient reinsurance capital12 (2) Undue reinsurer market power (3) Inefficient corporate form for reinsurance13 (4) High frictional costs of reinsurance (5) Moral hazard and adverse selection at the insurer level14 (6) Regulatory interference15 (7) ex post third-party financing16 (8) Behavioral factors
12 By nearly any measure, reinsurance appears very expensive. One problem in measuring the cost of insurance, however, is that measurement requires some estimate of the actuarial value of the insurance. Given the limited historical data available on catastrophic losses, such estimates are automatically suspect. However, capital-market reactions to announcements of reinsurance contracts are often positive (viz. Berkshire Hathaway’s obtaining the reinsurance contract for earthquake losses in California), indicating that the market as a whole believes that the premiums collected substantially exceed what the market believes the actuarially expected loss to be [Froot, 1997]. Indeed, Berkshire uses its financial capacity to tremendous competitive advantage by emphasizing that it can guarantee access to reinsurance capital where other firms cannot. This may explain Berkshire’s strategic response (q.v.), which precluded the issuance of catastrophe bonds in California for earthquake risks, an event that would rob them of what they see as a very lucrative business. 13 Two reasons are possible: (1) Managers of reinsurance firms regard their capital as equity-based and thus require returns in excess of the riskless rate. Writing reinsurance policies for catastrophic risks at actuarially fair rates is seen as being against the shareholders’ interests. However, given the uncorrelated nature of those risks with most other financial assets (although Dong, Shah, and Wong [1996] question this claim, it is generally supported [Guy Carpenter & Co., 1997]), shareholders’ required returns on catastrophic risks should be low. Agency costs may be one factor forcing up required returns - a factor that would not be present in some other (non-corporate) organizational form. (2) Many reinsurers’ shares strongly covary with the general stock market. This would indicate that there exists some systematic risk that would demand compensation. 14 That would occur, for example, if insurers only ceded coverage of risks that they had reason to believe would be bad investments. If the original insurer didn’t want the risk, why would anyone else want it? 15 This does not explain the high prices for catastrophic reinsurance, but it does explain why so little of it is purchased. Froot [1997] uses the analogy of rent control: the only way for insurance companies to increase their profits is to cut expenses - such as reinsurance. Policyholders get what they pay for, because the cheaper insurance is also worth less (given that the firm has an increased likelihood of default). This shifts more of the burden to government insurance pools, financed by taxpayers. 16 This refers to the government compensating losses after the occurrence of some catastrophe, without receiving a premium in advance [Kunreuther, 1996].
3-7
ACE Limited USAA CAT Limited
Principal $25 million $500 million $50 million
Coupon 6-Month T-bill + 550bp 1-Month LIBOR + 300bp 6.72%
Risk Period Last 5 months of 1996 8/96 to 7/97 Last 5 months of 1996
Trigger Level $25 billion industry loss $1 billion of firm losses $55 million of firm losses
Region Covered US US East Coast US Northeast
Risk Index PCS Company’s loss experience Company’s loss experience
Exhibit B: Three Examples of Early Catastrophe Bond Offerings Source: Lewis and Davis [1998]
The remainder of this paper examines catastrophic risk insurance financing through
the lens of behavioral factors, expanding on Froot’s eighth point and related accounts [e.g.,
Lewis and Davis, 1998] by drawing on general processes identified in the psychology of
investment and decision-making behavior. Section 3.2 further describes the insurance and
investment environment for securitization. Section 3.3 proposes that catastrophe bond
offerings cannot be sold at prices that insurance companies find acceptable unless they
address important behavior patterns. Section 3.4 proposes a market-level equilibration
hypothesis, namely that the current problems with catastrophe bond offerings are a function
of the novelty of the product and the psychology of market participants. Section 3.5
addresses tests of these hypotheses and concludes.
3.2 The PC Industry and the Nature of Catastrophic Risks
Although we assert in this paper that investor psychology plays a crucial role in the
market for catastrophic insurance risk, we certainly recognize that it is not the only factor.
There are several institutional and regulatory factors that also influence the development of
the market for securitized insurance risk and risk transfer products. In addition, because there
is likely to be interplay between investor psychology and institutional and regulatory factors,
it is important to comment briefly on the role such factors play in the overall problem.
3-8
3.2.1 The Insurance Market
A. M. Best, an insurance industry research firm, tracks 2,430 property and casualty
insurers [Standard & Poor’s, 1997]. Together, they wrote $259.8 billion in premiums for
1995.17 Although several large players dominate the industry, particularly for catastrophic
risk policies, the industry is still so fragmented that they lack pricing power. In fact, price
pressure, the fragmented distribution of capital across firms, and the increasing needs for a
multinational presence are spurring consolidation in the industry [Standard & Poor’s, 1997].
In 1997, 91 PC mergers were announced [Wehrly and Friedheim, 1998]. However, Wehrly
and Friedheim [1998] predict that competition will continue to escalate, increasing the gap
between weak and strong companies.
One measure of the strength of an insurance company is its surplus or reserves,
relative to its exposures. As mentioned, maintaining sufficiently high reserves is
prohibitively expensive, from both an accounting and strategic standpoint [Helbling,
Fallegger, and Hill, 1996].18 Additionally, firms that write more catastrophic risk insurance
are also more likely to have insufficient capital and surplus [Cummins and Doherty, 1997].
Thus, although the industry as a whole appears to have sufficient capital for even very large
catastrophic losses, that capital cannot actually be “pooled” across firms. Securitization
could offer contingent access to capital, with firms paying for the option value of that capital,
rather than maintaining their own standing reserves.
17 Financial results for PC insurers are reported one year in arrears (hence, a 1997 publication reports the accounting results for fiscal year 1996, which rely, in turn, on 1995 earned premium data). Premiums collected in a given accounting period are not recognized as revenue until the following year. Until then, they are considered “unearned premium income” and are held in reserve against losses that might accrue during the interim period. The length of this “vesting” period varies around the world. 18 Standard accounting practice (U.S. GAAP, namely FASB 5) prohibits companies from keeping “hidden” reserves. As a result, firms must leave large amounts of capital uninvested in the core business. Despite the fact that such assets are “reserved” for (possibly distant) future losses, any earnings from this “non-investment capital” are fully taxable. From a strategic perspective, maintaining large amounts of cash or marketable securities on a balance sheet not only provokes the ire of shareholders (who want their funds invested more profitably), but also can attract the eye of other firms or corporate raiders interested in putting such capital to “better” use. Also, Jaffee and Russell [1997] note that the only potential tax benefits of setting aside reserves (tax-loss carryforwards and backward tax-code provisions) are worthless in the event the firm goes bankrupt under a catastrophic loss, which is precisely when they would be used.
3-9
3.2.2 Regulation
As Klein [1997] notes, “insurance is perceived to be ‘vested with the public
interest.’” As a result, the PC insurance industry is subject to heavy government
involvement. Every state has an insurance commissioner (elected in thirteen), charged with
regulating the nature and premium level of coverage for firms operating in the state. State
and federal government institutions enforce entry and exit restrictions on firms (e.g.,
preventing non-insurance firms from indirectly restricting capital sources). They also limit
firms’ hedging behavior [Klein, 1997].
Such government intervention can impose costs on insurers and policyholders.
Penalva-Zuasti [1997] analyzed a simulated market for catastrophe bonds and earthquake
insurance in California and found that catastrophe bonds carried a 3.7% excess premium (in
1997) relative to estimated competitive prices. The implied efficient market prices were one
order of magnitude lower ($0.29 versus $3.29 per thousand dollars of coverage) than those
proposed by the California Earthquake Authority (CEA, a quasigovernmental agency).19 Not
only was CEA-sponsored insurance highly noncompetitive, but also catastrophe bonds
designed to improve market efficiency traded at a premium to competitive prices. Penalva-
Zuasti [1997] attributed this premium to the novelty of the product and to the “impact of the
highly regulated environment surrounding current insurance markets.”
3.2.3 Intra-Industry Competition
Santomero [quoted in Wharton Alumni Magazine, 1998] notes that insurers are
divided over such non-traditional methods of insurance coverage. For strong firms, with
sufficient capital to endure industry cycles, structured products such as cat bonds put
downward pressure on prices for catastrophic risk coverage, thereby reducing their profits.
For firms chronically short of capital, however, structured finance can strengthen their
19 Penalva-Zuasti [1997] estimated, for example, that for households living in earthquake risk-prone areas (mostly the Los Angeles and San Francisco basins), the average cost of full coverage (no deductible) was $0.29 per thousand dollars. The cost of the proposed CEA coverage (with a 15% deductible) was $3.29 per thousand dollars. Note also that the efficient rate ($0.29 per $1,000) is itself an overestimation of the cost of coverage, because it includes only households in risk-prone areas - not in the whole of California.
3-10
balance sheets at a reasonable price. This is summarized by the view that “the strong thrive,
the weak issue cat bonds.”20 Access to capital is frequently used to competitive advantage
by strong, well-capitalized insurers.
Several firms, in fact, have attempted to thwart the issuance of catastrophe bonds.
For example, the California Earthquake Authority (CEA) decided to issue bonds after
concluding that catastrophe insurance offered by Berkshire Hathaway (through its National
Indemnity subsidiary) would be too costly. However, at the last minute, Berkshire
underwrote the entire issue for approximately $650 million [Osterland, 1998].21 In Berkshire
Hathaway’s 1997 Annual Report, Chairman Warren Buffett [1997] spelled out his dislike of
catastrophe bonds in terms of his perception of their exploitation of investor psychology:
“The second word in this term [catastrophe bonds], though, is an Orwellian misnomer: A true bond obliges the issuer to pay; these bonds, in effect, are contracts that lay a provisional promise to pay on the purchaser. . . . This convoluted agreement came into being because the promoters of the contracts wished to circumvent laws that prohibit the writing of insurance by entities that haven’t been licensed by the state. . . . A side benefit is that calling the insurance contract a ‘bond’ may also cause unsophisticated buyers to assume that these instruments involve far less risk than is actually the case. . . . The influx of ‘investor’ money into catastrophe bonds -- which may well live up to their name -- has caused super-cat prices to deteriorate materially. Therefore, we will write less business in 1998.” [emphasis added]
Although the premium for catastrophe bonds estimated by Penalva-Zuasti’s [1997]
model runs contrary to Buffett’s assertion, his opinions and actions carry considerable
weight. It is unclear whether his public disdain reflects a desire to preserve the market as is
or is, instead, the sort of strategic behavior described by Borch [1962] and others. The
20 In fact, the real value of securities such as these comes from the increase in debt capacity that they provide (allowing firms a larger tax benefit without proportionately larger expected bankruptcy costs). If firms had decided instead to issue conventional debt securities, the expected bankruptcy costs may have prevented them from borrowing at all. Thus, even with a higher interest rate, from a traditional corporate finance perspective, catastrophe bonds may be less expensive than equity. 21 The actual contract provides a fourth layer of coverage to the CEA. Berkshire Hathaway agreed to provide $1.5 billion in coverage at a cost of $161 million per year for four years. The CEA’s financing was structured such that, in addition to working capital, the first layer ($3 billion) was provided by assessments on insurers, the second layer ($2 billion) was provided by reinsurance contracts at a cost of $148 million per year under a two-year contract, the third layer ($1 billion) was provided by possible assessments on policyholders, and the fifth layer again by insurer assessments. It is noteworthy that the fourth layer (Berkshire’s coverage) is more costly than the second, even though it is less likely to be used [California Legislative Analyst’s Office, 1997].
3-11
possibility of such behavior further complicates life for investors and issuers of such
securities, struggling to understand what they are buying or selling.
3.2.4 Definitions
Adam Smith [1776] noted that the insurance premium must compensate for the
expected losses, the expenses of insurer operation, and an appropriate profit on invested
capital. Most insured events happen with sufficient frequency to allow accurate estimates of
expected losses [e.g., Borch, 1969; 1990]. However, catastrophic losses come from
infrequent and unfamiliar events. As such, any model for predicting losses from them is
bound to be critically sensitive to the downside tail of the distributions of dollar losses. It is
precisely in those areas, however, that estimation is most difficult. Moreover, it encounters
the sorts of ambiguities that threaten the usefulness of forecasts [Fischhoff, 1994].
Estimating a probability distribution for “catastrophic events” requires a definition
of that term. Unfortunately, there is little standardization. Standard & Poor’s [1997] defines
a catastrophe as “an event or series of related events that causes insured losses of $5 million
or more.” Swiss Re [1996; 1997; 1998] defines a natural “catastrophe” as, “event caused by
natural forces. The following categories are used: flood, storm, earthquake (including
seaquake/tsunami), drought/bushfire/heat wave, cold/frost, and other (including hail and
avalanche).”
The Guy Carpenter Catastrophe Index (GCCI), used in many insurance-related
contracts (such as catastrophe derivatives), measures “atmospheric damage,” defined as
“hurricanes, tornadoes, windstorms, hail, and freezing temperatures.” It specifically
excludes all other perils, including fire, flood, lightning, earthquake, and riot. There is an
Event GCCI, for a single catastrophe, and an Aggregate GCCI, for a time period. This
method is used for all states except Texas [IndexCo, 1997]. The other major catastrophe
monitoring firm, PCS, defines a catastrophe as an “event resulting in excess of $5 million in
insured property damages and having an effect upon a large number of insurers and insured”
[Ray, 1993].
3-12
Epstein [1996] describes the difficulty of defining “catastrophe” in a contractually
meaningful way. He notes that it must consider the risk relative to the pool of potential
insured individuals. It must determine whether to consider small, but highly correlated losses
(such as asbestos damage) as catastrophes. If it is specific enough to be contractually
satisfying, a definition may be too idiosyncratic to allow generalization. Epstein concludes
by defining a catastrophic event as one “capable of laying devastation to vast numbers of
individuals at a single blow.” However, even he leaves “devastation,” “vast numbers,” and
“single blow” undefined (see also Zeckhauser [1996]).
A workable definition of catastrophic losses must also treat “collateral” or
“collocation” damage, such as debris from a collapsed building damaging neighboring
buildings or fires in structures next to those hit by lightning [Dong, Shah, and Wong, 1996].
If a hurricane (considered a catastrophic event) caused damage that led to a fire (not
considered a catastrophic event), how would the composite event be classified? Whether
these losses are included may determine whether the total insurable loss passes the threshold
for qualifying as a catastrophic loss. A clear definition is essential to investors in catastrophe
bonds.
3.2.5 Estimating the Risks of Catastrophic Events
However defined, the expected losses from catastrophes must be estimated, even if
they do not happen often enough to establish a track record in the actuarial sense [Dong et
al., 1996]. To illustrate this difficulty, Exhibit C shows three years of catastrophic losses as
reflected in Swiss Re’s annual sigma research reports. They show the variability over even
this short period, including the impact of an unusual event, the January 17, 1995, earthquake
in Kobe, Japan. Swiss Re [1996] estimated insured losses at $2.5 billion, but total damage
at $82.4 billion. In contrast, EQE [1995], a catastrophe monitoring firm, estimated total Kobe
losses at $95–$147 billion, not including building contents (such as equipment and
inventory), and insured losses at $6 billion. By either estimate, a much smaller fraction of
losses was insured in Kobe than for equivalent American events. In the 1994 Northridge
3-13
earthquake, $12.5 billion out of the total of approximately $20 billion in damage was insured
[EQE, 1994; U.S. Geological Survey, 1996]. Palm and Hodgson [1992] note similar
coverage rates for the 1989 Loma Prieta earthquake in California. Thus, uncertainty about
current coverage rates further complicate evaluating cat bonds, not to mention possible
future changes.
Exhibit C: Natural Catastrophes in 1995, 1996, and 1997
The dollar figures of these loss estimates are reported in the dollars of the year of the loss, unadjusted for inflation, and converted to U.S. dollars at the time of the loss (using the appropriate market or official exchange rate).
Although it has improved significantly over the past twenty years, the reporting
method for catastrophes still creates uncertainty for investors. Changes in the dollar value or
number of losses may reflect changes in reporting practices (e.g., pressure to reduce
fraudulent claims) as well as changes in the world. Furthermore, Property Claims Services
(PCS), the organization which estimates catastrophe losses, “does not release exact figures
in order to prevent their use by unauthorized third parties.” As a result, even Swiss Re obtains
the insured loss figures from PCS only within relatively wide ranges: the 1998 sigma report
includes the ranges $25 to $100 million and $101 to $300 million. These uncertainties
compound the inherent problems of estimating the probabilities and consequences of
catastrophic events causing catastrophes.
The frequency distributions of past catastrophic natural events are often well-
documented, with future estimates being further refined using basic science. For example,
1995 1996 1997 Total 1995 1996 1997 Total 1995 1996 1997 Total
Floods 45 44 478 567 5,835 6,853 4,950 17,638 $367 $233 $1,420 $2,021Storms 47 50 42 139 3,826 5,385 5,315 14,526 $7,452 $5,252 $2,460 $15,165
Earthquakes 13 8 16 37 8,406 544 2,878 11,828 $2,472 $0 $12 $2,484Drought, bushfires
8 7 5 20 1,452 97 667 2,216 $0 $0 $0 $0
Cold, frost 7 10 7 24 421 779 417 1,617 $536 $2,360 $168 $3,064Other 7 10 1 18 305 292 157 754 $1,602 $61 $80 $1,742
Total 127 129 549 805 20,245 13,950 14,384 48,579 $12,429 $7,906 $4,141 $24,475
Number of Events VictimsInsured Losses
(in millions)
3-14
there is a great deal of (physical) understanding22 of the ENSO (El Niño/Southern
Oscillation) phenomenon which has been described (Zebiak and Cane, 1998) as “second
only to the seasons themselves in driving worldwide weather patterns.” The U.S. Geological
Survey [Michael et al., 1996] notes that “[a]lthough quake forecasting is still maturing, it is
now reliable enough to make official earthquake warnings possible.” Nonetheless,
catastrophic events fall in the tails of these distributions, hence are the least predictable.
Furthermore, predictions often lack the spatial resolution needed by insurers of specific
properties. For example, Gray et al. [1998] caution readers of their hurricane forecasts that
“landfall probability estimates at any one location along the coast are very low . . . no matter
how active an individual season is.” Michael et al. [1996] voice similar warnings concerning
earthquake forecasts.
Estimating the consequences of these events requires understanding a complex web
of related events. Flooding causes landslides, earthquakes cause fires, winter storms often
lead to flooding, and so on [Swiss Re, 1996]. As mentioned, the available information is
disproportionately in the hands of the insurers. That imbalance is a barrier to investors, even
if they realize that the insurers themselves lack confidence in what they know.
Paradoxically, insurers may also avoid situations where their information is
sufficiently good that they could be held legally responsible for model errors. For example,
the State of Florida spent $1 million on certifying models of catastrophic exposure as
scientifically valid. Nonetheless, Florida’s Insurance Commissioner refused to use the
certified models and filed legal action to block state adoption of them [Florida Department
of Insurance, 1997]. Although other models (such as EQECAT’s USWINDTM and E. W.
Blanch’s Catalyst 3.0) have been certified by Florida [e.g., EQECAT, 1998; Property and
Casualty Online, 1998], regulators are still reluctant to use catastrophe models [Kibbee,
1997]. Katten [1997] writes “to a great extent, catastrophe modeling, while a useful tool, has
caused an unnecessary paranoia that adversely affects the market.”
22 For example, Gray et al. [1998] reported correlations of predictions of number of named storms of between r = 0.65 and r = 0.85 over the past 25 years.
3-15
3.3 Structure of Catastrophe Bonds
A cat bond is typically structured as a conventional corporate bond with an embedded
option. Other securities share this basic structure (e.g., convertible bonds, mortgage-backed
bonds, and even U.S. Savings Bonds).23 Lewis and Davis [1998] describe the 18 securities
offerings incorporating catastrophic risk exposure issued to that date. These 18 securities
have three common structures (see Exhibit A): (a) contingent surplus notes (CSNs), (b)
CatEPuts, and (c) catastrophe bonds. In addition to these structures, there was also the unique
St. Paul Re Pro Rata bond which will be described in detail in Section 3.4. CSNs are fully-
collateralized securities. The issuing insurance firm invests the proceeds of the offering in
Treasury securities. If a catastrophic event occurs, the insurance firm can substitute its own
corporate bonds for the Treasury bonds (effectively issuing debt at a prearranged price).
Investors receive the interest from the Treasury securities plus the premium (paid as yield)
for selling to the issuers the option to substitute its own debt. Exhibit D (adapted from Lewis
and Davis [1998]) illustrates the structure of CSN cash flows.
Exhibit D: Contingent Surplus Note (CSN) Cash Flow Structure
23 Convertible bonds obviously have the option (which belongs to either the issuer or the investor) of converting (into equity). Mortgages and mortgage-backed bonds typically have the option of prepayment. Savings bonds have the embedded option to extend the maturity of the bond or cash out early at a predetermined rate.
Treasury Security Collateral Pool
Principal
P&I
P&I, Option Premium
Principal
Trust
Investor
Insurer
Option Premium
"Substitutability" Put Option
3-16
CatEPuts, a contraction of Catastrophic Equity Puts, allow insurance firms to “put”
or sell new equity (in contrast to the new debt provided by CSNs) to investors at a
prearranged price, if a catastrophe occurs [Jewett, 1997]. This new equity typically takes the
form of newly issued shares (thereby diluting earnings for pre-existing shares). These
CatEPuts are a basic option transaction: the insurer pays a premium to investors in return for
the right to put its own shares to them should a catastrophic loss occur.
Cat bonds typically involve a standard corporate bond with a provision for reducing
its principal or interest (or both) in the event of a catastrophic loss. Thus, they provide capital
(or reduced debt) to the issuer when it is needed most. There are two classes of cat bonds:
indemnity cat bonds and index (or recapitalization) cat bonds. Indemnity bonds base contract
threshold payouts on the issuer’s own loss experience. Index-based catastrophe bonds use
indices such as PCS and GCCI (see Exhibits E and F). Using indices encourages a liquid
market in the bonds, but increases basis risk (insofar as the indices deviate from the issuer’s
actual loss exposure). Cat bonds were not issued until 1996, despite years of intense interest.
Even then, they have been hampered by lack of standardization in both the structure of the
bonds and the reference indices [Lewis and Davis, 1998]. (Exhibit B illustrated differences
in three early catastrophe bond offerings.) As a result, securities based on one index can only
be imperfectly substituted for securities based on the other, introducing additional basis risk.
Catastrophe derivatives address some of these liquidity problems. Although they can
be used for risk management in isolation, they are more frequently embedded in other
securities and contracts. The Chicago Board of Trade (CBOT) now makes a market in both
futures and options for which the underlying asset is catastrophic insurance risk in various
regions of the United States [CBOT, 1998].
These securities can be classified in terms of their relative exposure to moral hazard,
credit risk, and basis risk. For example, although reinsurance has a low risk of moral-hazard
problems, it carries a high credit risk. Index-based cat bonds have a low credit risk, but carry
substantial basis risk. Moral hazard refers to the possibility that purchasing an insurance
contract may change firms’ behavior. Reinsurance has greater moral hazard than CBOT
3-17
derivatives because the losses are more exactly matched to the exposure (compared to the
index-based derivatives). Issuers left with residual risk (as a result of the imperfect hedge)
have a greater incentive to reduce such risks.
Features PCS sigma RMS GCCI
Geographic Area
State Country ZIP code ZIP code
Insured Property
All major lines All lines All major lines Homeowners
Perils All significant perils
All perils Earthquakes and hurricanes
Hurricanes, hailstorms, tornadoes, thunderstorms, winter storms, and freezing conditions
Index Value Dollars of loss Dollars of loss Dollars of loss Paid loss-to-insured value ratio
Source of Estimate
Insurer survey, computer model, and ground survey
News and other sources
Computer model 39 companies’ insurance and paid loss records
Other Information Provided
None Number of casualties
None Premiums, deductibles, amounts of insurance, claim counts, paid losses, and construction types
Published 3 to 5 days after event, updated as necessary
Annually 7 days after event, with final value after 28 days
Quarterly
Exhibit E: Catastrophe Insurance Index Basis Comparisons
Source: Thomas [1997]
3-18
GCCI Regions PCS Regions Differences
National 50 States and DC National 50 States and DC -
Northeast CT, DE, DC, ME, MD, MA, NH, NJ, NY, PA, RI, VT, WV
Northeastern ME, NH, VT, MA, CT, RI, NY, NJ, PA, DE, MD (and DC)
WV
Southeast FL, GA, NC, SC, VA Southeastern VA, WV, NC, SC, GA, FL, AL, MS, LA
WV, AL, MS, LA
Gulf AL, AR, FL, LA, MS Eastern ME, NH, VT, MA, CT, RI, NY, NJ, PA, DE, MD (and DC), VA, WV, NC, SC, GA, FL, AL, MS, LA
Not comparable
Midwest /West
AZ, CA, CO, ID, IL, IN, IA, KS, KY, MI, MN, MO, MT, NE, NV, NM, ND, OH, OK, OR, SD, TN, UT, WA, WI, WY
Midwestern OK, AR, TN, KY, OH, MI, IN, IL, WI, MN, ND, SD, IA, NE, KS, MO
AR, AK, and HI are not included by GCCI
Western HI, AK, WA, OR, CA, NV, AZ, NM, UT, CO, WY, MT, ID
Florida FL Florida FL -
Texas TX Texas TX -
California CA GCCI does not offer isolated CA coverage
Exhibit F: GCCI versus PCS Coverage in Index-based Contracts
Source: IndexCo [1997]; Chicago Board of Trade [1998]
If the components of a structured product have liquid markets, then the composite
should trade at parity with the market value of those components. If not, then arbitrage is
possible. The conventional approach to pricing complex or “exotic” securities is to replicate
their cash flow streams with simpler securities that are easier to price [Black and Scholes,
1973]. Thus, the value of a cat bond equals the sum of the values of its constituent
components: a corporate bond and an option. The theories for evaluating both corporate
bonds [e.g., Fabozzi and Fabozzi, 1995] and options [e.g., Hull, 1993] are widely known.
Such arbitrage-supported replication presupposes: (1) that investors approach the valuation
of all securities identically (e.g., in estimating volatility) and (2) that the building blocks
needed to construct the replicating portfolio exist. Should the true value of a security, such
as a cat bond, not be preserved through such arbitrage bounds, investors’ perceptions of that
true value become a relevant factor in the pricing of such securities.
3-19
3.4 Human Behavior and Catastrophe Bonds
Following Olsen [1998] and Froot [1997], we consider how individuals might
perform the cognitively complex task of evaluating such investment products. Depending
on the circumstances, these processes could lead investors to overvalue, undervalue, or
refuse to value these offerings. Exhibit G briefly summarized eight phenomena and their
predicted impact on the market for catastrophe bonds.
Our analyses extrapolate from the research literature of behavioral decision making
to a domain in which no direct studies have been conducted. One constraint on such
extrapolation is that most behavioral decision-making studies have been conducted on
individuals without the financial experience of potential cat-security investors. Currently,
individual investors have no (direct) access to the market for these securities, which, for
regulatory reasons, is limited to institutional investors. Moreover, even these institutional
investors must be licensed to sell insurance by state insurance regulators before they can
transact in the catastrophe bond market [Klein, 1997]. However, the absence of arbitrage-
based pricing models means that these experts must rely on their own judgment. One would
like to believe that (1) experts can avoid mistakes on tasks central to their expertise and (2)
the stakes would motivate them to do so. Unfortunately, there is enough evidence of
imperfect expert judgment to allow the possibility for suboptimal choices in these novel,
difficult, and consequential tasks [e.g., Dawes, 1988; Dawes, Faust, and Meehl, 1989;
Mowen, 1994; Thaler, 1994; Kammen and Hassenzahl, 1999].
3-20
Stylized Facts About Human Decision Making
Features of Catastrophe Bonds
Implications for the Catastrophe-Bond Market
Understanding Facts of Transaction
Cognitively-constrained individuals tend to resort to sub-optimal heuristic decision rules in new and complex environments
Past catastrophe bond offerings have often exhibited very complex structures making hedging and comprehension difficult
May limit market depth if investors are “overwhelmed.” Those who stay may require higher returns as compensation for learning costs
Exaggerated comprehensiveness; individuals tend to overestimate the completeness of their pictures of complex problems
Catastrophe bonds are linked to insurance portfolios of the issuing firms that involve very complicated risk structures and events which are difficult to forecast
Investors (and issuers) may inaccurately estimate risks and thus costs, leading to different expectations of what an appropriate return would be
Illusion of control; the perception of risk varies inversely with the perception of control
Catastrophic events are frequently viewed as “acts of God”
Investors (and issuers) may inaccurately estimate risks and thus costs
Reliance on availability; the likelihood of highly available or vivid events tends to be overestimated
Catastrophic events are infrequent and dramatic
Investors (and issuers) may overestimate risks, and thus costs
Understanding Relevant Values
Small probabilities are overweighted Catastrophic events are, by their nature, very infrequent
Investors (and issuers) may overestimate risks, and thus costs
Violations of extensionality; the structure of a decision problem often influences individuals’ judgments.
Catastrophe bonds have features of both debt and equity, as well as the contingent exposure of options
Inability to think about cat investments with a pre-existing schema raises the cost of entering the market, inhibiting liquidity and increasing required returns
Dimensions of Risk
Individuals often perceive and react to differences between psychological and actuarial perceptions of risk
Events that have an actuarial nature to the firm have a visceral impact on individuals
Investors may be pricing (attaching value to) features of risks that are ignored by traditional financial theory
Asymmetric Information
Individuals tend to improperly weight different pieces of information; aversion to the risks of asymmetric information
Vast resources and the ability to collect primary data provide a substantial informational advantage to insurers
Liquidity of market is substantially constrained
Exhibit G: Behavioral Anomalies and Implications for Catastrophe Bonds
3-21
3.4.1 Cognitive Complexity
There are cognitive limits to individuals’ ability to function effectively in complex
decision-making environments [Simon et al., 1987]. Even where financial theory can distill
a problem to cash-flow streams, people often cannot. Instead, they rely on heuristic
reasoning and “automated” rule-based responses [Newell and Simon, 1972; Cross, 1983;
Albers and Laing, 1991; Becker, 1993]. In financial thinking, Neftci [1991] has
characterized reliance on technical analysis (essentially a collection of heuristic decision
rules) among experienced market participants as showing the pervasiveness of rule-based
decision processes.24
As a result, it would not be surprising to find, as observed by Lewis and Davis [1998],
that “investors sent a clear message to the insurance industry - complexity is a liability,”
after St. Paul Re placed only $68.5 million of $204 million in pro rata capital notes. The St.
Paul Re security, as depicted in Exhibit H, was remarkably complicated. It involved not only
St. Paul Re, but also two special purpose reinsurers (SPRs) created specifically for the deal,
a swap transaction, and two distinct collateral accounts. The pro rata capital notes combined
debt issuance with participation in returns on the reinsurance provided to St. Paul Re, which
was also to cede reinsurance business from five excess-of-loss classes, on a proportional
basis, in two pre-specified layers that adjust over time to reflect claims experience [Lewis
and Davis, 1998]. The bonds also offered options on two Class B St. Paul Re common shares
per $1 million in invested principal. This structure was actually intended to offer additional
protection for investors and, thus, make the bonds more desirable. However, the complexity
and attendant uncertainty over hidden risks overwhelmed investors.
24 Schwager [1992] contains the transcripts of interviews with several professional traders and investment managers. These interviews frequently include long discussions about the value of technical analysis, ex post rationalization of trading strategies, and descriptions of the use of visceral factors or “gut” instincts in making investment decisions—decisions frequently involving tens and hundreds of millions of dollars. It is important to note that even those traders who claimed not to “believe” in technical analysis personally still paid attention to technical trading statistics “because other traders use them.” One trader said the following of his trading system: “There was no system to it. It was nothing more than, ‘I think the market is going up, so I’m going to buy.’ ‘It’s gone up enough, so I’m going to sell.’ It was completely impulsive. I didn’t sit down and formulate any trading plan. I don’t know where the intuition comes from, and there are times when it goes away.”
3-22
Exhibit H: The St. Paul Re Pro Rata Capital Note Cash-Flow Structure Source: Lewis and Davis [1998]
In addition to prompting rule-based rejection, complexity can increase rational
rejection by accentuating investors’ informational disadvantage [Duffie and Rahi, 1995].
Investors must understand the relations in Exhibit H, track the various cash-flow streams,
and then evaluate the probabilities of different outcomes. They must do this well enough to
identify and hedge all sources of risk in the security. Here, that includes not only the risk of
catastrophic loss, but also (1) the counterparty risk of the swap, (2) the tax, accounting, legal,
and regulatory risks of the SPRs, including their offshore (UK) foreign-exchange risks, (3)
the interest rate risk of the zero-coupon securities held in the collateral account, (4) the basis
risk and moral hazard attached to the insurance contract itself, the correlations among these
risks, and possibly many others. At each stage, St. Paul Re is ahead of the investors. It may
just not be worth the transaction costs of catching up.
SwapCounterparty
Investor
Fixed Rate
Floating Rate
$204 millionAvailable Net Income
$204 - $304 million
$204 millionat maturity
$104 million
Reinsurance
Net Income
$100 millionSecurityInterest
Available NetIncome andCompanyExpense
Coverage LimitLimited to Value ofCollateral Account
St. Paul ReUK, plc
Collateral Account
Georgetown Re, Ltd.
10-Year Zero CouponAgency Securities
3-23
3.4.2 Exaggerated Comprehensiveness
Scholes [1996] criticized stock market and accounting regulators for failing to see
the “big picture” of corporate risk-management needs. He pointed specifically at regulators
limiting the use of derivatives in hedging, even when two highly volatile instruments should
hedge one another, thereby reducing overall risk. If his claim is correct, then such behavior
would be a special case of the general tendency to overestimate the completeness of one’s
picture of complex problems [Fischhoff, Slovic, and Lichtenstein, 1978], one source of the
overconfidence that has been found with many difficult tasks [Yates and Stone, 1992].
Such exaggeration could also cause those who do invest in catastrophe bonds to
overlook the risks involved. If their investments fare surprisingly poorly, then they may be
dissuaded from such investments, not understanding just what went wrong. For example,
they might not realize the full implications of a catastrophic event on the issuing firm, such
as the interconnectedness of losses when a hurricane causes landslides, strikes two insured
locations, or triggers associated health and life insurance-related claims. Such possibilities
can reduce a catastrophe bond’s value both by increasing the chances of triggering the
embedded option and by diminishing the issuing firm’s financial capacity.
3.4.3 Illusion of Control
Olsen [1998] summarizes financial evidence of the general phenomenon of people
exaggerating their control over uncertain events, leading, in turn, to underestimating risks.
In a classic study, subjects playing a game of pure chance were more aggressive with an
opponent who appeared naive than with one who appeared sophisticated [Langer, 1975].
The novelty of these securities, highlighted by the term “Act of God” bonds (used by USAA)
may have allowed little room for illusions of control. Inexperienced investors have had little
chance to develop heuristics or confidence needed to convince themselves that they could
beat the averages. If so, then, ironically, the success of these securities may have been limited
by their inability to take advantages of natural biases.
3-24
3.4.4 Reliance on Availability
Tversky and Kahneman [1973] proposed that, when individuals do not know the
frequency or probability of an event, they judge its frequency by its availability to memory.
Although often helpful, this heuristic can exaggerate the likelihood of disproportionately
salient events. For example, Lichtenstein et al. [1978] and also Combs and Slovic [1979]
found that subjects overestimated the frequency of deaths from relatively publicized causes
(e.g., tornadoes and floods), while underestimating the frequency of deaths from less visible
ones (e.g., cancer or heart disease). Potential investors in catastrophe bonds must assess the
probability of adverse events. Unless they accept the modelers’ claims, they must rely on
their own judgments. If catastrophes are disproportionately available, then investors will
overestimate their probability. If so, then investors would demand greater compensation for
bearing these risks than insurers, relying on model estimates, see reason to pay.
3.4.5 Overweighting Small Probabilities
According to prospect theory’s decision weighting function, people pay undue
attention to small probabilities, above and beyond any errors in their estimation (e.g., due to
availability). As statistically rare events, catastrophes might receive such weighting (above
and beyond any tendency to misjudge their probabilities). For example, Kahneman and
Tversky [1979] had subjects choose between a sure loss of $5 and a 0.001 chance at losing
$5,000. Of 72 respondents, 83% preferred the sure loss, even though the two options had the
same expected value. One in a thousand (0.001) is a plausible probability for catastrophe
bonds to exceed the preset threshold and trigger the embedded option payout leading to loss
of principal.25 In a normative pricing model, this small probability would imply a
comparably small premium. A prospect theory weighting function would lead investors to
overweight that admittedly small probability of loss and, thus, demand a higher return.
25 No catastrophe bond (including all securities in Exhibit A) had ever (as of 1999) had its loss threshold triggered yet. Even events the magnitude of the recent Hurricane Floyd have failed to generate losses for cat-bond investors. One analyst noted “everyone’s been wondering what will the capital markets do when there is a real loss” [Wall Street Journal, September 15, 1999].
3-25
Prospect theory’s probability weighting function is discontinuous at the endpoints,
with very small or large probabilities rounded to certainty. Thus, investors who saw the
probability fall below a certain threshold might treat cat bonds as risk free. However, that
would mean ignoring the whole point of the bonds. Rather, it seems likely that some small
overweighted probability would remain.
3.4.6 Violations of Extensionality
Psychologists have long known how formally equivalent ways of describing the
same tasks can affect people’s choices [Turner and Martin, 1984; Poulton, 1989; Fischhoff,
1991; Schwarz, 1999]. Tversky and Kahneman’s [1981] prospect theory provided an
integrated account of such effects, cast in terms of rational actor models. For example, they
showed that patients are less likely to choose surgery framed in terms of the probability of
death rather than the complementary probability of surviving.
In an ideal world, investors would analyze securities on their own merits, in terms of
first principles of finance. However, in fact, they are often trained to analyze types of
investments: bonds, stocks, options, etc. Thus, standard textbooks [e.g., Bodie, Kane, and
Marcus, 1993; Sears and Trennepohl, 1993] discuss the pricing and trading of fixed income
securities, corporate equities, and derivatives in distinct sections, with little overlap. Only
advanced books, directed toward more mathematically-sophisticated audiences [e.g., Park
and Sharp-Bette, 1990; Luenberger, 1998], focus on evaluating pure risky cash-flow
streams.
Although called catastrophe bonds, these securities’ exposure to the equity of the
insurance firm (as subordinated, unsecured debt) means that they also can behave like equity.
However, they are not really equity because they have no ownership interest in the firm.
Exhibit I summarizes ways in which cat bonds exhibit both bond and stock behavior. An
investor using heuristic rules for either bonds or stocks would miss vital features unique to
this hybrid asset class, or erroneously assume features not actually found in catastrophe
bonds. An investor who realized that neither frame of reference worked entirely might shy
3-26
away from the investment, not knowing how to think about it. Consider, for example, an
investor who thought of catastrophe bonds as traditional fixed-income securities. Redeeming
a traditional bond below par means that the issuer was in such poor financial health that it
cannot make full repayment. However, with catastrophe bonds, below-par redemption
reflects the occurrence of a catastrophe. The firm might actually be stronger financially
because the cat bond reduced its exposure.
Characteristic Traditional Corporate Bonds
Common Equity Catastrophe Bonds
Type of Return Fixed, known return “Random” return Partially fixed, partially random, contingent on catastrophe
Ability to Hedge Payoff
Certain, hedgable payoff; liquid markets
Hedgable payoff; liquid markets
Contingent payoff; illiquid markets for hedging; imperfect ability to hedge (basis risk)
Performance Model
Known distribution (lognormal)
Generally known distribution (normal, lognormal)
Discrete, low-frequency data; distribution not precisely known and subject to change
Certainty of Payoff
Unequivocal payoff determination
Unequivocal payoff determination
“Event” definition varies widely between indices and issuers
Default Claims Hierarchy, Managerial Interests
Secured/Guaranteed; Top-tier; No managerial interest
Lowest priority, but managerial discretion
Top-tier structure and appearance, but most are subordinated; no managerial discretion
Risk in the Context of the Firm
Moderate correlated risk
Moderate correlated risk
High correlation risk: ability to repay weakest precisely when catastrophe depresses the bond price (because of the embedded option)
Exhibit I: Comparing and Contrasting Catastrophe Bonds with Conventional Securities
When cat bonds trade below par, it indicates high expected losses that will be paid by the
bondholders. Without paying significant transaction costs, investors may be unable to
decode the unique nature of these securities, adding to their cost, confusion, and chance of
poor choices [Allen and Gale, 1994].
3-27
3.4.7 Dimensions of Risk
The Capital Asset Pricing Models (CAPMs) of Sharpe [1964], Lintner [1965], and
Mossin [1966] state that investors should only be compensated for bearing systematic (or
undiversifiable, market) risk. The psychological literature indicates, however, that
individuals often have rather different notions of risk. In early work, Slovic [1964] found
that perceived risk could not be measured by a single index, noting that a “large amount of
evidence bearing on the convergent validity of [methods assessing risk-taking propensity] is
negative.” As complicating factors, he pointed to emotional arousal and cognitive concerns
outside of classical financial theory.
One common approach to studying the multidimensional character of perceived risk
is to have subjects judge the riskiness of activities along dimensions like those in Fischhoff
et al. [1978]. Factor analyses of these judgments typically find that two factors explain much
of the variance in subjects’ judgments: Unknown risk measures the extent to which potential
effects are delayed, unobservable, new, and unknown to science. Dread risk measures the
extent to which the activity is seen as uncontrollable, inequitable, involuntary, catastrophic,
or potentially fatal [Slovic, 1987; Jenni, 1997]. Slovic et al. [1984] found that, other things
being equal, people are more inclined to accept chronic risks (e.g., auto accidents) than
catastrophic ones (e.g., accidents involving nuclear power), and are particularly averse to
ones evoking a feeling of dread.
Although cat bonds (and the like) have not been studied in these terms, some
speculation seems possible. The dramatic, unpredictable nature of catastrophes may create a
visceral response among investors afraid of losing much (or all) of their investments, perhaps
akin to the feeling of dread in these studies of life-threatening risks. Investors might want
compensation for such feelings, even if they officially subscribe to normative asset pricing
models.
Within those models, expected return might be related to dread risk while variance
captures some of unknown risk. Higher moments might play a role as well, with skewness
3-28
capturing some of dread risk and kurtosis related to unknown risk, in the sense of the
prevalence of extreme (tail) events. However, Payne [1973] found that the moments were
“unacceptable as variables for the theory of risky decision making” because the interactions
terms could not be independently estimated in real-life choices. Coombs and Lehner [1981]
reached the same conclusion for experiments. Preferences over higher-order and partial
moments have been extensively studied in finance [Hogan and Warren, 1974; Kraus and
Litzenberger, 1976; Bawa and Lindenberg, 1977; Fishburn, 1977; Holthausen, 1984; Sortino
and van der Meer, 1991], with mostly disappointing results. Sharpe [1964] found that partial
moments (semivariance) seemed to provide a better fit to the observed data, but rejected
them because of computational complexity that continues to make them impractical for real-
time use. Thus, formal financial models ignore these features, so central to the experience of
risk.
3.4.8 Asymmetric Information
Asymmetric information (as illustrated by Akerlof [1970]) is particularly important
in insurance and risk transfer [Nachman and Noe, 1994]. “Given the potential for adverse
selection […] we might expect markets to collapse if the issuer’s information is sufficiently
large relative to that of potential investors” [Duffie and Rahi, 1995, p. 2]. As mentioned,
such asymmetry exists with catastrophic information, which the insurers both collect and
disseminate. As a result, insurers hoping to create a market should provide as much
information as possible to potential investors, at minimal cost.
Credible communication is particularly important when there seem to be incentives
for strategic reporting. For example, the loss threshold for many early catastrophe bonds
(e.g., the USAA and CAT Limited issues) depended on the issuing firm’s own loss estimates.
Such informational disadvantages could further discourage investors and increase demands
for risk premiums.
3-29
3.5 Market Equilibration
Securities linked to catastrophic risks challenge investors to perform tasks that
research has shown to seem and be difficult, as well as creating mismatches between the
perspectives of investors and issuers. As a result, individuals may shun the market altogether
or perform poorly in it, thereby discouraging future investments. Such discomfort
characterizes many new markets, until investors become more familiar with them. Then,
trading volume and liquidity increase, and prices move towards an equilibrium. Whether this
happens depends on whether investors merely need time and information to understand a
market, or if they need fundamental help with comprehension and debiasing. The issues
raised here suggest there exist barriers to the learning process for cat securities capable of
preventing them from ever leading to a stable and optimal equilibrium. In a situation too
novel to allow immediate comprehension and too complex to allow trial-and-error learning,
the “invisible hand” may be stilled.
3.5.1 Immature Market Structure
Pricing inefficiencies are eliminated most reliably in liquid markets with many active
traders and prompt delivery of detailed, accurate information. Even though neither element
is present in the current market for catastrophe reinsurance-based products, Chichilnisky and
Heal [1998] state confidently that because the “underlying pressure [from insurers to tap
new capital] is relentless,” securitized offerings will eventually predominate.
Swiss Re [1996] makes similar claims, citing parallels to the securitization process
in the banking sector. When the worldwide debt crisis of the early 1980s constrained
commercial lending capacity, direct issuance of securities increasingly complemented and
partly substituted for traditional corporate financing. Trading in U.S. Treasury futures, often
considered the most successful financial innovation in recent memory, began in 1977, but
took 15 years to reach its current volume. Swiss Re’s [1996] projections (Exhibit J) assume
similar market growth for cat bonds.26
26 Although volume in the Treasury futures market continues to grow over time, recent growth levels are much lower than in earlier periods. In fact, the logistic curve tends to provide a good model of development, as
3-30
Exhibit J: Development Scenario for ART Products Source: Swiss Re [1996]
One potential flaw in this analogy is that most investors had investment experience
with Treasury bonds long before Treasury futures were introduced. The same cannot be said
for cat bonds, which were specifically created to access a new capital market. Thus, rather
than “repackaging” an existing product, cat bonds are entirely new for most investors.
Resolving pricing inefficiencies in the cat-bond market and increasing liquidity could take
more time, perhaps even longer than issuers will be willing to tolerate
3.5.2 A Dual Equilibrium
One obstacle to acceptance is a liquidity “Catch-22”: the risk characteristics of
insurance-related securities should be attractive to many investors, but only if they can sell
them, should the need arise. However, the limited liquidity of the cat-bond market
growth rates eventually slow once a market becomes saturated. At this point, further growth is limited by the growth rate of the market itself. Swiss Re developed its projections for cat-bond market development by modeling catastrophic insurance risk as an asset class and determining its weight in an optimally diversified portfolio. This figure determined the maximum possible market share and it was then assumed that market penetration would proceed according to a standard logistic-shaped development cycle.
$0
$5
$10
$15
$20
$25
$30
$35
$40
1992 1995 1998 2001 2004 2007 2010
Bill
ions
of D
olla
rs
Year
New Cat Capacity via Derivative ExchangesNew Cat Capacity via SecuritizationTotal Alternative Cat Capacity
3-31
invalidates standard no-arbitrage arguments and models. At the extreme, the market value
of a security with no liquidity is zero, regardless of what some pricing model claims.
Silverman, Sparks, and Osterland [1998] point to an “illusion of liquidity” in the
securitization market: “[j]ust because an asset is tradable today doesn’t mean it will be
tradable tomorrow.” Without liquidity, traders lack the prices needed to use their models.
Without traders, there is, in turn, no liquidity.
Embrechts [1996] identifies an unfortunate “dual equilibrium”: investors are either
all in or all out of the market.27 Given the abundance of other investments, it is easy enough
simply to go elsewhere with one’s money. One barrier that the issuers are trying to reduce is
the lack of hedging instruments needed by investors pursuing analytical strategies. Indeed,
some progress is being made in developing securities usable for hedging cat bonds (e.g.,
exchange-traded catastrophe options). However, there is still too much variability in the
structure of these different options and too little volume in those markets to cover any sizable
risks. The idiosyncratic structure of most cat-bond offerings makes general hedging
instruments hard to create.
Economides [1995] notes that in financial exchange markets, high liquidity is a
positive externality: it increases the willingness of all participants to trade and is provided to
traders without cost. Economides [1992; 1993; 1995] has demonstrated that any degree of
participation in a market can be sustained as an equilibrium, including none at all. Of course,
equilibria involving greater participation are more beneficial to those trying to create a model
(e.g., insurers and, ultimately, those purchasing insurance).
Embrechts [1995] proposes an institutional approach to inducing participation,
whose details depend on the design of the security in question. According to his analysis,
investor behavior will affect prices (and participation) as long as strict arbitrage bounds are
unavailable. Unless investors can understand securities appropriately, there will be no
externalities, and markets will remain at a low-liquidity equilibrium. Federal Reserve
27 This would explain the “logistic” shape of the product development/market saturation curves: slow initial acceptance → sudden, broad acceptance → sustained tapering off as saturation occurs.
3-32
Chairman Greenspan was quoted as saying “market discipline appears far more draconian
and less forgiving than 20 or 30 years ago. Capital, in times of stress . . . flees more readily
to securities and markets of unquestioned [emphasis added] quality” [Wall Street Journal,
1998]. Thus, while securitization can provide new sources of capital, it can also make the
availability of that capital more volatile.
3.6 Conclusions and Implications for Securities Design
Insurers have an enormous interest in tapping capital markets by securitizing
catastrophic insurance risk. Additional capital could help to cover future losses that
potentially run in the hundreds of billions of dollars. Creating a market for catastrophic risk
would repeat the successes securitization has enjoyed with other cash flows.
In other markets, the required equilibration process involves merely allowing
investors to adjust to new information. However, with significantly novel products like
catastrophic risk, that may never happen because investors are often constrained in their
ability to understand and process the necessary information as it is commonly found. With
vast, complex, and interconnected sources of uncertainty, new schemata might not be created
and existing ones may be misused.
We outlined eight well-known behavioral decision-making results, suggesting how
they might affect investor behavior in this arena. When investors cannot muster proper
information use, they generally resort to heuristic decision making, attempting to use
“similar” schemata to tackle new problems. Unfortunately, such decision making can result
in errors of sufficient magnitude to prevent widespread participation in such new markets.
For example, Bantwal and Kunreuther [2000], building, in part, on an earlier version of this
paper, illustrate how suboptimal decision-making processes could lead to an excessive risk
premium for cat bonds. These processes are likely to play a role in the securitization of any
significantly new (to investors) cash-flow stream, not just cat bonds.
3-33
This analysis is but the beginning of the research process. The next step is to evaluate
these hypotheses in the actual market for catastrophic insurance risk (or analogous novel
securities). Once the contours of these problems are better understood, methods of solving
them must be developed. Our account suggests that widespread acceptance will require
limiting new offerings to relatively simple and standardized products. These will provide
investors with a common and cognitively tractable schema for thinking about the nature of
their investments. That way, they can focus their attentions on the still challenging task of
understanding the world of catastrophes. The design of new securities merits consideration
of human behavior equivalent to that devoted to the design and marketing of consumer
products.
3.7 Epilogue
In the time that has elapsed since this paper was published, the cat bond market has
undergone significant change. Issuance remained largely flat for years after the paper was
published, before beginning to increase in 2006, and again in 2010 (see Exhibit K). Among
the notable events impacting the market subsequent to our paper was the first loss event
experienced by investors. A $190 million issue by KAMP Re 2005 Ltd suffered a total loss
of principal in 2005 due to Hurricane Katrina. The resulting resolution of the issue proceeded
smoothly, giving confidence to potential investors as to the handling of loss events
[Cummins, 2008]. (By 2014, cumulative losses across $51 billion in cat bonds issued totaled
only $682 million (1.3%)—a figure much lower than the average premium paid on the bonds
[Yoon and Scism, 2014].)
Writing in 2005, Hommel and Ritter [2005] acknowledged these high premiums at
the time, noting that the excess premiums paid on early cat bond issues were slowly
declining, but that a lack of understanding persisted in the market. Citing our paper, they
acknowledged that many of the issues we identified were still barriers for investors. Then,
three years later, Cummins [2008] wrote that the market began to experience steady growth
and falling premiums, with (then) recent data suggesting “broad market interest […] among
institutional investors.”
3-34
Exhibit K: Actual vs. Projected Issuance of Cat Bonds Source: Projected issuance is from SwissRe [1996]; actual issuance is from the Artemis Catastrophe Bond &
Insurance-Linked Securities Deal Directory as of February 2017
Exhibit L: Contemporary Standardized Cat Bond Structure
$0
$2,000
$4,000
$6,000
$8,000
$10,000
$12,000
$14,000
$16,000
$18,000
0
10
20
30
40
50
60
1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015
Mill
ions
of D
olla
rs
Number of Issues (left axis) Total Size (right axis)
Projected Cat Securitization (right axis)
Hurricane CharleyHurricane Ivan
Hurricane KatrinaHurricane Wilma Hurricane Sandy
Hurricane Ike Hurricane Matthew
Special Purpose Vehicle
InvestorInsurer
Trust Account
Interest Rate Hedge
Fixed Rate
Floating Rate
Principal
Floating Rate+ Premium
PrincipalPrincipal +
Floating Rate
Premium
Proceeds if loss
triggered
Proceeds if loss not triggered
3-35
It was during this period that the cat bond market underwent fundamental change.
Carayannopoulos and Perez [2015] note that the new structures used beginning in 2009 were
positively received by the market. A new “standardized” structure emerged (Exhibit L) that
addressed many of the issues raised in my original paper: (i) the structure was simplified and
standardized to reduce complexity and mitigate concerns about it being a “black box,” (ii)
the treatment of collateral was made more rigorous and third parties were brought in to hold
collateral in trust, allowing the securities to be seen as a “pure play” on catastrophe risk and
diminishing the tendency toward exaggerated comprehensiveness, and (iii) the loss triggers
were simplified and externalized to mitigate concerns about asymmetric information. As a
result, the new structure (Exhibit L) resembled more closely Exhibit D than Exhibit H.
Today, the cat bond market is viewed largely as successful [Scism and Das, 2016].
Nevertheless, the current market for cat bonds is significantly smaller than the projections
made in 1996 for its market share (illustrated in Exhibit K). The industry continues to push
forward though, and has attempted to securitize a broad range of new catastrophic risks, such
as rogue traders, cyber hacking, and accounting fraud [Scism and Das, 2016]. Richards
[2016] summarizes this experience as a market heading toward greater standardization, but
unable to shake an attachment to complexity. In the process, market participants have revised
their beliefs about what is and is not an attractive feature. For example, one of the primary
advantages touted of cat bonds was their lack of correlation to other assets classes. Ironically,
as Richards [2016] notes, some market participants now see the lack of correlation as an
added risk in the form of unpredictability (here, as a type of ambiguity aversion).
Overall, the market has learned a number of important lessons both with regard to
securitizing catastrophic risk in particular and in regard to behavioral factors related to
securities design more generally. This learning process notwithstanding, issuers have been
persistently unable to shake a tendency toward complexity in design. On one hand, this is
not a new observation. Scitovsky [1950] noted that complexity (or consumer ignorance) was
a source of oligopoly power for firms. More specifically, Carlin [2009] noted that firms add
complexity, rather than transparency, as a means of increasing market power and
3-36
maintaining supernormal profits. Henderson and Pearson [2011] make this same case
specifically in regard to innovation in securities design. Insurers, then, may want to improve
transparency (and reduce complexity) in order to attract new investors in cat bonds, but not
by so much that it cedes their pricing power. Maintaining his longstanding opposition to cat
bonds, Berkshire Hathaway’s Warren Buffet lamented that reinsurance had become a
“fashionable asset class” [Scism and Das, 2016], and that the resulting lower prices were
leading Berkshire’s reinsurance subsidiaries to enter into fewer deals.
Some firms, it seems, were all too happy to maintain their arcane aura, and therefore
their profit margins. As a result, encouraging less complex, and more “palatable” design of
securities may not only increase capital to the insurance industry, but may also mitigate
market power and improve insurance pricing to consumers.
4-1
4 Regulated Equity Returns: A Puzzle
4.1 Introduction†
In economics, the equity-premium puzzle refers to the empirical phenomenon that
some equity returns have exceeded the riskless rate of return by more than can be explained
by traditional models of compensation for bearing risk. Since Mehra and Prescott’s [1985]
initial paper on the subject, a large body of research has attempted to explain away the
puzzle, but without much success [Mehra and Prescott, 2003]. The most likely explanation
for the premium appears to be the presence of certain frictions not reflected in classical
equilibrium models (always a less-then-fully-satisfying answer).
It is not our intent, however, to rehash the general equity-premium puzzle discussion.
Rather, we call the reader’s attention to it as a means of introducing our instant problem,
which may be considered a special case of the Mehra-Prescott puzzle. Simply put: why are
the equity returns authorized by electric utility regulators so high, given that riskless rates
are so low? Our “puzzle” involves the large and growing spread, or risk premium, observed
between the rates of return on equity authorized by utility regulators and the riskless rate of
return. In addition to questioning the magnitude of the spread, we also question why it has
grown over time.
This regulated equity return puzzle is important not just from a theoretical asset-
pricing perspective, but also for very practical reasons. The database of more than 1,400 rate
cases since 1980 used in this paper reflects approximately $2 trillion in cumulative rate base
exposure.28 An error or bias of merely one percentage point in the allowed return would
† The author gratefully acknowledges the thoughtful comments received from Tony Páez and from participants at the Carnegie Mellon Electricity Industry Center’s weekly seminar. All responsibility for any errors remains with the author. 28 This figure reflects the simple cumulative sum of authorized rate bases across all cases. Because rate-base decisions may remain in place for several years, this sum most likely underestimates the actual figure, which should be the authorized rate base in each year examined, whether or not a new case was decided. We cite this figure merely as evidence of the substantial magnitude of the costs at stake.
4-2
involve tens of billions of dollars in additional cost for ratepayers in the form of higher retail
power prices and could play a profound role in the allocation of investment capital. Coupled
with utilities’ tendency toward excessive capital accumulation or “gold plating” under rate
regulation [Averch and Johnson, 1962; Spann, 1974; Courville, 1974; Hayashi and Trapani,
1976; Tapon and van der Weide, 1979], the magnitude of the problem makes it incumbent
on the industry to get it right.
There are also policy implications for market design and regulation. A recent PJM
Interconnection [2016] study compared and contrasted entry and exit decisions in
competitive and regulated markets in an attempt to evaluate the efficiency of competitive
markets for power. One finding that emerged from the study was that regulated utilities
appeared to be “overearning” and had generated positive alpha, but competitive firms had
not generated positive alpha.29 Although the study used a limited time window of rate case
data, its findings are consistent with those we explore in more detail here.
As an old joke goes, an economist is someone who sees something work in practice
and asks whether it can work in theory. Undoubtedly, the utility sector has been successful
in attracting capital over the past three decades. We cannot necessarily say, however, that
had returns been consistent with theoretical models (and thus lower), this would still have
been the case. Accordingly, this paper also raises the question of whether our theoretical
models of required return and asset pricing must be refined. Or, at the very least, whether
there are important considerations that must be accounted for in the application of those
models to the regulated electric utility industry.
In this paper, therefore, we examine the historical data on authorized rates of return
on equity in electric utility rate cases. We compare these rates of return to a number of
conventional benchmarks and a theoretical asset-pricing model. We demonstrate that the
spread between authorized equity returns (and also actual “earned” equity returns) and the
riskless rate has grown steadily over time. We investigate whether this growing spread can
29 In asset pricing models, positive alpha is evidence of non-equilibrium returns, meaning that investors are receiving compensation in excess of what would be required for bearing the risks they have assumed.
4-3
be explained by conventional asset pricing parameters and conclude that it cannot. Finally,
we conclude by suggesting alternative reasons for the growing spread and provide some
suggestions for future research and policy-making in the area.
4.2 Regulated Equity Returns and the CAPM
At the outset, let us make clear that we are addressing only regulated rates of return
on equity in this paper. We draw no conclusions or inferences about the behavior of returns
on non-regulated assets. Such returns may be considered a priori appropriate in that they are
subject to the discipline of competitive markets for investment. Our focus is limited to
regulated returns because in such cases it is regulators who are tasked with standing in for
the discipline of a competitive market and ensuring that returns are just and reasonable. For
more than a century, courts have ruled consistently in support of this objective, while
recognizing that achieving it requires consideration of numerous factors that are subject to
change over time.30 The task set to regulators, then, is to approximate what a competitive
market would provide, if one existed.
Mindful of this mandate, two U.S. Supreme Court decisions are commonly thought
to provide the conceptual foundation for utility rate-of-return regulation. In Bluefield Water
Works & Improvement Co. v. Public Service Commission of West Virginia, the Court ruled
that:
“The return should be reasonable, sufficient to assure confidence in the financial soundness of the utility, and should be adequate, under efficient and economic management, to maintain and support its credit and enable it to raise money necessary for the proper discharge of its public duties.”31
30 In Bluefield Water Works & Improvement Co. v. Public Service Commission of West Virginia (262 U.S. 679 (1923)), the Court identified eight factors that were to be considered in determining a fair rate of return. In United Railways & Electric Company of Baltimore v. West (280 U.S. 234 (1930)), the Court concluded: “What is a fair return […] cannot be settled by invoking decisions of this Court made years ago […]. The problem is one to be tested primarily by present day conditions.” 31 Bluefield Water Works & Improvement Co. v. Public Service Commission of West Virginia (262 U.S. 679 (1923)).
4-4
Then, in Federal Power Commission v. Hope Natural Gas Company, the Court ruled:
“…return to the equity owner should be commensurate with returns on investments in other enterprises having corresponding risks.”32
In both Bluefield and Hope, the Court sought to balance the need for utilities to attract capital
sufficient to discharge their duties with the need for regulators to protect ratepayers from
what would otherwise be rent-seeking monopolists.
These efforts in determining “just and reasonable” returns received significant
assistance in the 1960s when groundbreaking advances in asset pricing theory were made in
finance. Specifically, the development of the Capital Asset Pricing Model (“CAPM”)
[Sharpe, 1964; Lintner, 1965; Mossin, 1966] provided a rigorous framework within which
the question of the “appropriate” rate of return could be addressed in an objective fashion.
The security market line representation of the CAPM set out the equilibrium rate of return
on equity, , as the sum of the rate of return on a riskless asset, , and a premium related
to the level of risk being assumed, : .
It is outside of the scope of this paper to delve too deeply into the foundations of
asset pricing. We note, also, that the CAPM methodology is not the sole candidate for rate-
of-return determination in utility rate cases. Morin [2006, p. 13] identifies four main
approaches used in the determination of the “fair return to the equity holder of a public
utility’s common stock,” of which the CAPM is but one.33 Nevertheless, the concept of the
appropriate rate of return on equity being a combination of a riskless rate of return and a
premium for risk-bearing has since become widely accepted as a means of determining the
appropriate authorized return on equity in utility rate cases [Phillips, 1993] and the CAPM
32 Federal Power Commission v. Hope Natural Gas Company, 320 U.S. 591 (1944). 33 The other three approaches identified by Morin are: Risk Premium (which is an attempt to derive empirically what the CAPM derives theoretically), Discounted Cash Flows (or “DCF,” which is a Gordon-type dividend capitalization model), and Comparable Earnings (which is an empirical approach to deriving cost of capital from market comparables based on the Hope ruling cited above). A typical example of these approaches’ use in practice is found with the Pennsylvania Public Utility Commission [2016]: “The Commission determines the ROE […] based on the range of reasonableness from the DCF barometer group data, CAPM data, recent ROEs adjudicated by the Commission, and informed judgment.”
4-5
in particular has been seen as the “preferred methodology” in regulatory proceedings [Roll
and Ross, 1983].
Further, in Hope, the Court advocated the “end results doctrine.” By that, the Court
acknowledged that regulatory methods were immaterial so long as the end result was
reasonable to the consumer and investor. In other words, there is no single formula for
determining rates. Although we will generally be content in this paper to follow the Hope
doctrine and examine returns at a very general level as riskless rate plus risk premium, we
will call upon the CAPM approach in order to make quantitative assessments of the data
examined.
To the extent that the spread, – , has grown under any of the theoretical models,
it must be because compensation for risk-bearing has increased. Although the CAPM is but
one model of compensation for risk-bearing, it will be useful in this paper as a means of
categorizing our various investigations into the growth of that compensation. In the CAPM,
spread growth is related directly to two types of factors: (i) , the measure of the market-
relative riskiness of an asset, and (ii) the market risk premium ( ), which is the market
price of risk (and thus the competition for capital).
Before we turn to the data, however, let us dispense with an alternate formulation of
the underlying question. In questioning the size of the spread and why equity returns are so
high, one might also ask instead why the riskless rate is so low. Indeed, Mehra and Prescott
[1985] ask this very question, before dismissing it on theoretical grounds. We may revisit
this question in light of recent data and ask whether the observed spread growth is more a
function of riskless rates being forced down by the Federal Reserve’s intervention, than of
equity returns increasing (since the manifest intent of quantitative easing is to lower riskless
rates).34 Our historical data, as the next section indicates, does not support that hypothesis.
34 This has been an ongoing issue of contention in recent regulatory proceedings. In Opinion 531-B (issued March 3, 2015), FERC found that “anomalous capital market conditions” caused the traditional discount rate determination methods not to satisfy the Hope and Bluefield requirements (150 FERC ¶ 61,165 at 7). But in a related decision only eighteen months later (Order on Complaint, Establishing Hearing and Settlement Judge Procedures, issued September 20, 2016), FERC acknowledged that expert witnesses disagreed as to whether any market conditions were “anomalous” (156 FERC ¶ 61,198 at 10).
4-6
The spread growth has persisted since the beginning of our data series in 1980 and has
persisted across a variety of monetary and fiscal policy regimes.
4.3 Regulated Electric Utility Returns on Equity Since 1980
4.3.1 Historical Authorized Return on Equity Data
The data used in this paper reflect the collective experience of 1,421 electric utility
rate cases from almost every state in the country from January 1980 through December 2015.
We examine the returns on equity authorized by the regulatory agencies, not the returns
requested by the utilities.35 Out of all rate cases filed in the United States, the data only
include rate cases in which the utility has requested a rate change of at least $5 million or a
regulator has authorized a rate change of at least $3 million. We further exclude cases in
which no decision was made with regard to the return on equity and capital structure. Lastly,
only cases that have been fully litigated or settled are included; cases pending decision or
under appeal at the time of our analysis are not considered.
Nearly all fifty states and Washington D.C. are represented in the data set.36 Twenty-
eight electric utility rate cases satisfying the qualifications listed above were filed in the
average state over the past thirty-five years, with the most being filed in Wisconsin (115)
and the fewest being filed in Alaska, Alabama, and Tennessee with one each (as shown in
Exhibit M). Although Wisconsin is the most common state in the data, the frequency of
filing in a state does not appear to have any relationship to spread growth. The premium has
grown in both the ten states that completed the most rate cases and the ten states that
completed the fewest rate cases, and has grown at very similar rates (see Exhibit N). In fact,
as Exhibit O illustrates, the general trend across all states is similar.
35 To be clear, we refer to the rates set by regulators as the “authorized” rates. These may be contrasted with the utilities’ “requested” rates and also with the actual “earned” rates of return realized by utilities. Regulatory authorization of a rate is not a guarantee that a utility will actually earn such a rate. We address this issue in further detail in a later section. 36 Only Nebraska did not have a reported rate case meeting the parameters of the data set.
4-7
Exhibit M: Qualifying Rate Cases Filed per State for the Ten Most Frequently- and Least Frequently-Filing States, 1980-2015
Exhibit N: Risk Premium Growth by Frequency of Case Filing. Gaps in the series reflect years in which no rate cases were filed for the subject group
In the early 1980s there were over 100 rate cases filed each year. By the late 1990s,
in the midst of widespread deregulation of the electric power industry, the number of filings
reached its lowest point (with five in 1999). Since then, filing frequency has increased to an
average of more than thirty-five per year over the last three years (see Exhibit P). The decline
in rate case activity in many instances was the direct result of rate moratoria related to the
transition to competitive markets in the late 1990s, as well as to moratorium-like concessions
0
100
200
300
400
500
600
700
800
900
1,000
1980 1985 1990 1995 2000 2005 2010 2015
Spr
ead
of R
ate
of R
etur
n ov
er U
.S.
Trea
sury
Rat
es (
in b
asis
poi
nts)
Average of 10 Least Frequently-Filing StatesAverage of 10 Most Frequently-Filing States
4-8
made to regulators related to merger approvals over the last decade. Many of these moratoria
will expire over the next several years, suggesting a new increase in rate case activity is
likely.37 Finally, no individual utility had an outsized influence on the sample. One hundred
forty-five different companies filed rate cases, but many have since merged or otherwise
stopped filing.38 The average firm filed ten rate cases in our sample. Within our sample the
most frequently-filing entity was PacifiCorp, which filed seventy-one rate cases, or 5% of
the sample.
Exhibit O: Range of Risk Premium Growth Across States. States with highest (VA) and lowest (SD) rates of growth (among states with at least 5 rate cases) are
highlighted
37 Although these new rate cases may tend to be dominated by transmission and distribution utilities as opposed to generation utilities, given the structural shifts that have occurred within the industry. 38 The level of analysis is at the regulated utility level. We recognize that many holding companies have multiple ring-fenced regulated utility subsidiaries.
4-9
Exhibit P: Filing Frequency by Year
4.3.2 The Regulated Equity Premium
Generally speaking, regulated equity returns are equal to the sum of the riskless rate
of return and a premium for risk-bearing. In the CAPM, the premium for risk-bearing is
given by . Rearranging the security market line equation, we define the regulated
equity premium as – . Presented thus, we first note that the existence of
a (positive) regulated equity premium is not, by itself, evidence of irrational investor
behavior or model failure. Neither is the existence of a growing regulated equity premium.
We take no position here on what the “correct” premium should be in any particular instance.
Rather, we shall be content in this paper simply to determine whether or not the behavior of
the risk premium in practice is consistent with existing financial theory.
Of the 1,421 rate cases used in the analysis, all but nineteen (1.3%) had approved
rates of return greater than the riskless rate at the time.39 On average, the authorized return
on equity is 4.8% higher than the riskless rate. The standard deviation on the spread is 2.1%.
Exhibit Q illustrates the average authorized return on equity over the period against the
average annual riskless rate and investment-grade corporate bond rate.40 For avoidance of
39 The small number of exceptions is exclusively from the early 1980s when a rapid increase in interest rates made the problem of regulatory lag particularly acute. 40 We used the 10-year constant maturity U.S. Treasury note yield as a proxy for the riskless rate and the yield on the Moody’s Seasoned Baa Corporate Bond Index as a proxy for investment-grade corporate bond rates. Board of Governors of the Federal Reserve System (US), 10-Year Treasury Constant Maturity Rate [DGS10], retrieved from FRED, Federal Reserve Bank of St. Louis https://research.stlouisfed.org/fred2/series/DGS10,
4-10
doubt, we note that only the U.S. Treasury note rate should be considered the riskless rate.
We include corporate bond rates solely to assess whether the trend in riskless rates is
materially different from the trend in risky debt. Exhibit R shows the distribution of the
premium for all of the rate cases in our sample.
Exhibit Q: Authorized Return on Equity vs U.S. Treasury and Investment Grade Corporate Bond Rates
Exhibit R: Distribution of Premium Across All Years
March 7, 2016; Board of Governors of the Federal Reserve System (US), Moody’s Seasoned Baa Corporate Bond Yield© [BAA], retrieved from FRED, Federal Reserve Bank of St. Louis https://research.stlouisfed.org/fred2/series/BAA, March 7, 2016.
0%
2%
4%
6%
8%
10%
12%
14%
16%
18%
1980 1985 1990 1995 2000 2005 2010 2015
Rat
e of
Ret
urn
Authorized Return on EquityAverage 10-Year U.S. Treasury NoteAverage Moody's Seasoned Baa Corporate Bond
4-11
While the regulated equity premium has averaged 483 basis points across the entire
time period, in 1980 the average premium was only 278 basis points, whereas in 2015 it
averaged 771 basis points. Exhibit S shows the difference between the authorized return on
equity and the riskless rate for each case in the data over the past thirty-five years. Although
the spread is determined against the riskless rate of return (represented here as the yield on
a 10-year U.S. Treasury note), we also present for comparison the spreads determined
against the yield on the Moody’s Seasoned Baa Corporate Bond Index to illustrate that the
effect is not an artifact of recent monetary policy on Treasury rates. The trends of the two
series are quite similar (and both have statistically-significant positive slopes41);
accordingly, we shall present only the Treasury rate-determined spreads throughout the
remainder of this paper.
Exhibit S: Authorized Return on Equity Premium, 1980-2015
Given that a large and growing regulated equity premium exists, our question is
whether or not it can be explained within an equilibrium asset-pricing framework such as
the CAPM. If were to have increased during the time period in question, for example, the
41 With both U.S. Treasury rates and corporate bond rates, the slopes are reliably positive (p < 0.001).
4-12
growth of the regulated equity premium may well be explained by the increasing (relative)
riskiness of utility equity. As the next section demonstrates, however, in fact it cannot.
4.4 Potential Theoretical Explanations for the Premium
Having demonstrated the existence of a large and growing regulated equity premium,
we investigate various potential explanations. As we indicated above, we proceed with our
investigation of explanations for the premium via the Capital Asset Pricing Model. The
CAPM allows three basic mechanisms of action for a change in the risk premium: (i) the
manner in which the underlying assets are financed has changed, (ii) the risk of the
underlying assets themselves has changed, and/or (iii) the rate at which the market in general
prices risk has changed. We explore each in turn.
4.4.1 Capital Structure Effects
As corporate leverage increases, the underlying equity becomes riskier and thus
deserving of higher expected returns. In finance, the Hamada equation decomposes the
CAPM equity beta ( ) into an underlying asset beta ( ) and the impact of capital structure
[Hamada, 1969; 1972]. Specifically, the Hamada equation states that 1 1 ,
where is the tax rate and D and E are the debt and equity in the firm’s capital structure,
respectively.42
One explanation for a growing risk premium would be steadily increasing leverage
among regulated utilities. However, regulators also generally approve of specific capital
structures as part of the rate-making process. As a result, our database also contains the
authorized capital structures for each utility.43 In fact, utilities are less leveraged today than
42 We use the marginal corporate federal income tax rate for the highest bracket, as provided in the SOI Tax Stats, Historical Table 24. Downloaded from https://www.irs.gov/pub/irs-soi/histabb.xls on March 8, 2016. 43 To be clear, the authorized capital structures evaluated here apply to the regulated utility subsidiaries, and not necessarily to any holding companies to which they belong. The holding companies themselves may utilize more or less leverage, but typically the regulated utility subsidiaries are “ring-fenced” so as to isolate them from holding company-level risks. Similarly, rate-of-return regulation would apply only to the regulated subsidiaries, not to the parent holding company. As a result, the capitalization of the regulated entity (studied here) is often different from the capitalization of the publicly-traded entity that owns it.
4-13
they were in 1980. The average debt-to-equity ratio in the first five years of the data set
(1980-1984) was 1.73; in 2011-2015 it was 1.04. More generally, we can observe the impact
of leverage moving in the opposite direction of what one may expect, whether we examine
the debt-to-equity ratio exclusively or the Hamada capital structure parameter (i.e., the
portion of the Hamada equation multiplied by , or 1 1 ) in its entirety. Exhibits
T and U illustrate these results. As a result, it does not appear as if capital structure itself can
explain the behavior of the risk premium.
Exhibit T: Authorized Rate-of-Return Premium vs. Utility Leverage
4-14
Exhibit U: Authorized Rate-of-Return Premium vs. the Hamada Capital Structure
Parameter
4.4.2 Asset-Specific Risk
As noted above, the Hamada equation decomposes returns into compensation for
bearing asset-specific risks and for bearing capital structure-specific risks. Even if a firm’s
capital structure remains unchanged, the riskiness of its underlying assets may change. This
risk is represented by the unlevered asset beta, . An increase in the asset beta applicable
to such investments would, all else held equal, justify an increase in the risk premium.
To examine such a hypothesis, we used the fifteen members of the Dow Jones Utility
Average between 1980 and 2015 as a proxy for “utility asset risk.” We calculated five-year
equity betas for each firm by regressing their monthly total returns against the total return
on the S&P 500 index.44 The equity betas calculated were then converted to asset betas using
Hamada’s equation and corrected for firm cash holdings using firm-specific balance sheet
information. We then averaged the fifteen asset betas calculated in each year as our proxy
44 We evaluated the composition of the Dow Jones Utility Average index at the end of each year and used a rolling five-year window to perform the regressions. For example, the 1980 regression betas were calculated based on monthly returns from 1975-1979, the 1981 regression betas were calculated based on monthly returns from 1976-1980, and so on.
4-15
for utility asset risk.45 The results remain substantively unchanged whether an equal-
weighted or a capitalization-weighted average is used.
Exhibit V: Authorized Rate-of-Return Premium vs. Industry Average Asset Beta
Although there is, of course, variation in the industry average asset beta across the
thirty-five years, the general trend is down. Exhibit V presents the risk premium in
comparison to the industry average asset beta. As a result, the asset beta is moving in the
opposite direction from what one might expect, given a steadily-increasing risk premium,
and therefore does not appear to explain the observed behavior of the risk premium.
4.4.3 The Market Risk Premium
The last CAPM-derived explanation for a changing risk premium relates to the
pricing of risky assets in general. If investors require greater compensation for bearing the
45 The balance sheet and total return data are taken from the Standard & Poor’s COMPUSTAT database. We calculate 1 1⁄ and 1⁄ , where C equals the amount of cash and cash equivalents held by each firm and D and E represent, respectively, the debt and equity of each firm. We measure D as the sum of Current Liabilities, Long-Term Debt, and Liabilities–Other in the COMPUSTAT data.
4-16
systematic risk of the market in general, then the risk premium across all assets would
increase as well (all else held equal) as a result of the average risk aversion coefficient of
investors increasing. The market risk premium reflects this risk-bearing cost in the CAPM.
Although we can observe the ex post market risk premium, investors’ assessment of
the ex ante market risk premium is generally based on assuming that historical experience
provides a meaningful guide to future experience.46 It is customary to examine the actual
market risk premium over some historical time period and base one’s estimate of the
expected future market risk premium on that historical experience (see, for example, Sears
and Trennepohl [1993]).
Exhibit W: Market Risk Premium Trends over Time by Historical Window
Exhibit W illustrates the ex ante market risk premium determined in each year by
examining a historical window of n years and using the average ex post premium calculated
46 We do not dwell here on the issue of the “observability” of the market portfolio as it relates to testability of the CAPM. We shall assume that the S&P 500 index is a reasonable proxy for the market portfolio.
-10%
-5%
0%
5%
10%
15%
20%
25%
1980 1985 1990 1995 2000 2005 2010
Ave
rage
Ann
ual
Pre
miu
m
5 Years
25 Years
50 Years
Linear (5 Years)
Linear (25 Years)
Linear (50 Years)
4-17
within that window.47 It is sufficient for our purposes to observe simply that the slope of the
market risk premium over time has been negative irrespective of the historical window used.
Throughout the remainder of this paper, we use a fifty-year historical window for calculation
purposes. As Exhibit X illustrates, that declining trend in the ex ante market risk premium
appears inconsistent with the increasing risk premium exhibited by the rates of return
authorized by regulators.
Exhibit X: Authorized Rate-of-Return Premium vs. Ex Ante Estimated Market Risk Premium
4.4.4 Testing a Theoretical Model of the Risk Premium
In this section, we examine each component composing the risk premium under the
CAPM and observe that the historical experience of each of these series is at odds with what
the CAPM would imply about the risk premium derived from rates of return authorized by
regulators. We can go further, however, than making these general observations. By
47 The market risk premium data used here are taken from data on the S&P 500 and 10-year U.S. Treasury notes collected from the Federal Reserve and maintained by Prof. Aswath Damodaran. Downloaded from http://pages.stern.nyu.edu/~adamodar/New_Home_Page/datafile/histretSP.html on March 8, 2016.
4-18
combining the security market line representation of the CAPM and the Hamada equation,
we can quantitatively identify the risk premium, .
1 1 [1]
In [1], is the risk premium, or the difference between the authorized rate of return on
equity and the riskless rate, taken as the average of the yield on the 10-year U.S. Treasury
note during the year in which the equity return was authorized. The asset beta, , is
calculated as in Section 4.4.2. The middle component is taken from the Hamada equation
and reflects the marginal corporate income tax rate in effect in the year in which the equity
return was authorized and the authorized debt-to-equity ratio reflected in the regulators’
decisions. Lastly, MRP is the ex ante estimate of the market risk premium based on a fifty-
year historical window as of the year in which each equity return was authorized.
By using a logarithmic transform of [1], we arrive at an equation that can be tested
via linear regression:
ln ln ln 1 1 ln [2]
Finance theory would hypothesize that should be zero (or not significant) and , , and
should be positive and significant. What we find, however, is exactly the opposite of that.
The intercept is negative and significant and the coefficients are negative and strongly
significant (Exhibit Y).48 Further, a comparison of the actual risk premium spreads to the
regression-model spreads reveals a good fit (Exhibit Z).
48 We removed 19 data points (approximately 1.3% of the data set) that had negative spreads from use in the regression analysis.
4-19
ln , Constant -5.638****
(0.151)
, Asset beta,ln -0.184**** (0.027)
, Capital structure,ln 1 1 -0.487**** (0.116)
, Market risk premium,ln -0.957**** (0.039) R-squared 42.6% Adjusted R-squared 42.4% F statistic 345.3**** No. of observations 1,402
Standard errors are reported in parentheses. *, **, ***, and **** indicate significance at the 90%, 95%, 99%, and 99.9% levels, respectively.
Exhibit Y: Regression Results for CAPM-based Risk Premium Model
Exhibit Z: Actual vs. Regression-Model Risk Premium Spreads
-600
-400
-200
0
200
400
600
800
1,000
1,200
1980 1985 1990 1995 2000 2005 2010 2015
Spr
ead
of R
ate
of R
etur
n ov
er T
reas
ury
Rat
es(in
bas
is p
oint
s)
Actual Risk Premium
Regression Model
4-20
Exhibit Z also reveals a distinct shift in the predicted trend of the risk premium
beginning in 1999. This is notable because for many parts of the U.S., 1999 represented the
year that implementation of deregulation began, with wholesale markets commencing
operation and several divestiture transactions of formerly-regulated generating assets
occurring. To examine this point in time, we divided the data into two sets, 1980-1998 and
1999-2015, and estimated separate regression models for each subset (Exhibit AA).
Although the results in both cases are consistent with our earlier finding that the standard
finance model appears at odds with the empirical data, the two regression models are
noticeably different from one another and appear to better represent the data (Exhibit BB).
We find this result suggestive that deregulatory activity may have an influence even on still-
regulated utilities, although we leave further exploration to future work.
1980-1998 1999-2015 ln ln
, Constant -15.552**** -4.156**** (0.725) (0.101)
, Asset beta,ln -0.894**** -0.084**** (0.133) (0.009)
, Capital structure,ln 1 1 -0.131 -0.333****
(0.153) (0.056)
, Market risk premium,ln -4.585**** -0.497**** (0.263) (0.027) R-squared 27.2% 45.5% Adjusted R-squared 27.0% 45.2% F statistic 113.6**** 134.2**** No. of observations 916 486
Standard errors are reported in parentheses. *, **, ***, and **** indicate significance at the 90%, 95%, 99%, and 99.9% levels, respectively.
Exhibit AA: Regression Results for a Two-Period CAPM-based Risk Premium Model
4-21
Exhibit BB: Actual vs. Two-Period Model-Predicted Risk Premium Spreads
4.5 Possible Implications
The consequences of these findings are significant, both because of the cumulative
amount of capital and ratepayer expense at stake, and also because of what they may imply
for finance theory and public policy. For example, although fuel costs and wholesale power
prices have declined since 2007 (see Exhibit CC), the retail price of power has increased
over the same period (see Exhibit DD). One explanation for this divergence in wholesale
and retail rates may be the presence of a growing premium attached to regulated equity
returns and therefore embedded into rates. To be sure, other forces may also be at work (for
example, recovery of transmission and distribution system investments is consuming a
greater portion of retail bills—a circumstance potentially exacerbated by excessive risk
premiums). Further, even if the growing divergence between wholesale and retail rates is
related to a growing risk premium, it does not necessarily follow that such growth is
inappropriate or inconsistent with economic theory. Nevertheless, the potential for
embedding of such quasi-fixed costs into the cost structure of electricity production may be
significant for end users.
-600
-400
-200
0
200
400
600
800
1,000
1,200
1980 1985 1990 1995 2000 2005 2010 2015
Spr
ead
of R
ate
of R
etur
n ov
er T
reas
ury
Rat
es (
in b
asis
poi
nts)
Actual Risk Premium
Two-Period Regression Model
4-22
To the extent that the size and growth of the risk premium over time cannot be
explained by traditional asset pricing models, we examine potential causes and consequences
of this behavior of regulated rates of return in four distinct areas as a guide toward further
study and an indicator of potential policy implications.
Exhibit CC: Wholesale Fuel and Power Price Trends, 2007-2015 Data Source: Intercontinental Exchange; U.S. Department of Energy, Energy Information Administration
4-23
Exhibit DD: Wholesale vs. Retail Power Prices, 2007-2015
Data Source: Intercontinental Exchange; U.S. Department of Energy, Energy Information Administration 4.5.1 Potential Alternative Finance Explanations
In Mehra and Prescott’s [2003] review of the equity premium puzzle literature, the
authors acknowledge that uncertainty about changes in the prevailing tax and regulatory
regimes may explain the premium. Such forces may also be at work with regard to regulated
rates of return. To the extent that investors require higher current rates of return because they
are concerned about future shocks to the tax or regulatory structure of investments in
regulated electric utilities (e.g., EPA’s promulgation of the Clean Power Plan, the U.S.
Supreme Court’s stay of the Clean Power Plan), such concern may be manifest in a higher
degree of risk aversion unique to the electric utility sector than would otherwise be obtained.
A separate line of inquiry concerns a criticism of the Hamada equation in the
presence of risky debt (Hamada [1972] excluded default from consideration). Conine [1980]
extended the Hamada equation to accommodate risky debt by applying a debt beta.
Subsequently, Cohen [2008] sought to extend the Hamada equation by adjusting the debt-
to-equity parameter to incorporate risky debt. We view neither of these proposed solutions
as entirely satisfying, and note that they tend to be material only for high leverage, which is
not common to regulated utilities. Nevertheless, we acknowledge that adjustments to the
4-24
capital structure may influence the risk premium. However, applying the Cohen [2008]
modification49 and using the Moody’s Seasoned Baa Corporate Bond Yield as a proxy for
the cost of risky debt ( ), we note that our regression results are substantively unchanged.
Lastly, some researchers have suggested that the Arbitrage Pricing Theory (“APT”)
[Ross, 1976] is preferable to the CAPM because the CAPM produces a “shortfall” in
estimated returns [Roll and Ross, 1983] and “underestimates” actual returns in utility
settings [Pettway and Jordan, 1987]. While the works of Roll and Ross and of Pettway and
Jordan are suggestively similar to the analysis contained in this paper, we note that those
authors were examining the actual returns on utility common stocks, rather than the rates of
return authorized by regulators for assets held in utility rate bases.
To address a similar point, we also examined the actual earned rates of return on
equity for the 15 utilities in the Dow Jones Utility Average over our historical window. We
used each firm’s actual return on equity, calculated annually as Net Income divided by Total
Equity, as reported in the COMPUSTAT database. This measure of firm profitability
examines how successful the firms were at converting their authorized returns into earned
returns. In general, the earned returns closely tracked the authorized returns, suggesting that
the decisions of regulators are significantly influencing the actual earnings of regulated
utilities. Exhibit EE compares the spread of authorized rates of return over riskless rates to
the spread of earned rates of return over riskless rates and to the median net income of
utilities in constant dollars.50 The steadily increasing risk premium we have identified is
present in both series.51 Further, the “capture rate” (the percentage of authorized rates
actually earned by the utilities) averaged 96% over the entire time period. As a result, we
49 Cohen’s modified Hamada equation is 1 1 . 50 We used the median earned rate of return over the 15 Dow Jones utilities. The results are substantively equivalent if the average earned rate of return is used, but are more volatile due to the impact on earnings of the California energy crisis of 2000-2001 and the collapse of Enron in 2001. 51 These measures of firm performance must be interpreted with caution. The authorized rates of return apply to jurisdictional utilities, while the earned rates of return are calculated based on holding company performance, which in many cases are not strictly equivalent. Further, increasing net income may be due to industry consolidation producing larger firms (with income increasingly only proportionally to size), rather than an increase in profitability itself. Nevertheless, the results are suggestive.
4-25
conclude that the trend of increasing risk premiums is not an abstract anomaly occurring in
a regulatory vacuum, but rather a direct contributor to the earnings of regulated utilities.
Exhibit EE: Comparability of Spreads Measured with Authorized and Earned Rates of Return and Utility Net Income
We have not taken that analysis the final step and examined the relationship between
firm performance and stock performance. However, the findings of Roll and Ross [1983]
and Pettway and Jordan [1987] stated in different terms suggest that regulated utilities have
realized higher stock returns than can be explained by the CAPM—a finding congruent with
our work and suggestive of other factors being priced by the market. This does not entirely
explain, however, why regulators appearing to use a CAPM approach provide utilities with
returns that also appear to be excessive.
4.5.2 Potential Public Policy Explanations
Public policy, or regulation itself, may be a causal factor in the observed behavior of
the risk premium. The U.S. Supreme Court acknowledged, in Duquesne Light Company et
al. v. David M. Barasch et al. (488 U.S. 299 (1989)) that “the risks a utility faces are in large
part defined by the rate methodology, because utilities are virtually always public
$0
$200
$400
$600
$800
$1,000
$1,200
$1,400
$1,600
-200
0
200
400
600
800
1,000
1,200
1980 1985 1990 1995 2000 2005 2010 2015 Med
ian
Net
Inco
me
(mill
ions
of
201
5$)
Spre
ad o
f R
ate
of R
etur
n ov
er T
reas
ury
Rat
es (i
n ba
sis
poin
ts)
Authorized Rates (left axis) Earned Rates (left axis)
Median Net Income (right axis)
4-26
monopolies dealing in an essential service, and so relatively immune to the usual market
risks.” The recognition that the very act of regulating utilities subjects them to a unique class
of risks may influence its cost of capital determination. And yet, if the purpose (or at least a
purpose) of regulating electric utilities is to prevent these quasi-monopolists from charging
excessive prices, but the practice of regulating them results in a higher cost of equity capital
than might otherwise apply, it calls into question the role of such regulation in the first place.
Similarly, regulators may also question whether the hybrid regulated and non-
regulated nature of the electric power sector plays a role as well. Has deregulation caused
risk to “leak” into the regulated world because both regulated and non-regulated firms must
compete for the same capital? Has the presence of non-regulated market participants raised
the marginal price of capital to all firms? We note that the trajectory of public policy during
the entire time period studied has been toward deregulation (beginning before 1980 with
Public Utility Regulatory Policy Act, through the Natural Gas Policy Act in the 1980s, and
electric industry deregulation in the 1990s) and that “today’s investments face market,
political and regulatory risks, many of which have no historical antecedent that might serve
as a starting point for modeling risk.” [PJM Interconnection, 2016] Has the progressive
deregulation of the industry caused a convergence in regulated and non-regulated returns
over that time period? The data do not suggest that utilities in states that have never
undertaken deregulation have meaningfully different risk premiums, but there are many
ways to evaluate the “degree” of deregulatory activity that could be explored.
Another public policy-related factor could be a change in the nature of the rate base
or of rate-making itself. Toward the beginning of our study period, most of the electric
utilities were “vertically integrated” (i.e., in the business of both generation and transmission
of power). Over time, generation became progressively exposed to deregulation, while
transmission and distribution of power has tended to remain regulated. To the extent that the
portion of the rate base comprised of transmission and distribution assets has increased at
the expense of generation assets, it may suggest a shift in the underlying risk profile of the
assets being recognized by regulators. We note, for example, that public policy has tended
to favor transmission investments with “incentive rates” in recent years in order to address
4-27
a perceived relative lack of investment in transmission within the electric power sector.
Although we do not present the results here, there does not appear to be a meaningful
difference between the average annual risk premiums for vertically integrated utilities and
non-vertically integrated utilities.
As for a change in the nature of rate-making itself, we note that the industry has
tended to move from cost-of-service rate-making to performance-based ratemaking. Has this
shift, in an attempt to increase utility operating efficiency, inadvertently raised the cost of
equity capital through the use of incentive rates, and has the net cost-benefit balance been
positive if so?
4.5.3 Potential Behavioral Economics Explanations
One curious observation in the empirical data is that the average rate of return on
regulated equity appears to have “converged” to 10% over time. Although the underlying
riskless rate has continued to drop, authorized equity returns have generally remained fixed
in the neighborhood of 10% recently. Anecdotally, we have observed a reluctance among
potential investors to accept equity returns on power investments of less than 10%—even
though those same investors readily acknowledge that debt costs have fallen. To that extent,
then, a behavioral bias may be at work.
In economics, “money illusion” refers to the misperception of nominal price changes
as real price changes [Fisher, 1928]. Shafir, Diamond, and Tversky [1997] proposed that this
type of choice anomaly arises from framing effects, in that individuals give improper
influence to the nominal representation of a choice due to the convenience and salience of
the nominal representation. The experimental results have been upheld in several subsequent
studies in the behavioral economics literature [Fehr and Tyran, 2001; Svedsäter, Gamble,
and Gärling, 2007].
The effect here may be similar: investors and regulators may conflate “nominal” rates
of return (the authorized rates) with the risk premium underlying the authorized rate. The
4-28
apparent “stickiness” of rates of return on equity around 10% is similar to the “price
stickiness” common in the money illusion (and, indeed, the rate of return is the price of
capital). If there was in fact a tendency (intentional or otherwise) to respect a 10% “floor,”
one might expect that the distribution of authorized returns within each year may “bunch
up” in the left tail at 10%, where absent such a floor one may expect them to be distributed
symmetrically around a mean. As Exhibit FF illustrates, we see precisely such behavior. As
average authorized returns decline to 10%, the skewness of the within-year distributions of
returns becomes persistently and statistically significantly positive, suggesting a longer
right-hand tail to the distributions, consistent with a lack of symmetry below the 10%
threshold.52
Exhibit FF: Authorized Rates of Return on Equity and Skewness
A related finding has been reported by Fernandez, Ortiz, and Acín [2015], where
respondents to a large survey of finance and economics professors, analysts, and corporate
managers tended, on average, to overestimate the riskless rate of return. In addition, their
estimates exhibited substantial positive skew, in that overestimates of the riskless rate far
52 Our test statistic for skewness is equal to the skewness divided by its standard error
6 1 2 1 3⁄ , where n is the sample size. The test statistic has an approximately normal distribution [Cramer and Howitt, 2004].
4-29
exceed underestimates.53 The authors found similar results not just in the U.S., but also in
Germany, Spain, and the U.K. In the U.S., the average response exceeded the
contemporaneous 10-year U.S. Treasury rate by 30 to 40 basis points. It may be that
overestimating the riskless rate is simply one way for participants in regulatory proceedings
to “rationalize” maintaining the authorized return in excess of 10%. Alternatively, it may be
an additional bias in the determination of authorized rates of return.
If such biases exist, there are clear implications for the regulatory function itself. For
example, this apparent 10% “floor” was even recognized recently in a FERC proceeding: “if
[return on equity] is set substantially below 10% for long periods […], it could negatively
impact future investment in the [New England Transmission Owners].”54 One notable
jurisdictional difference in regulatory practice is between formulaic and judgment-based
approaches to setting the cost of capital. In Canada, for example, formulaic approaches are
more prevalent than in the United States [Villadsen and Brown, 2012]. By pre-committing
to a set formula (e.g., government bond rates plus n basis points) in lieu of holding
adversarial hearings, regulators could minimize the potential for deviation from outcomes
consistent with finance theory. Villadsen and Brown [2012] note, for example, that recent
rates set by Canadian regulators have tended to be lower than those set by U.S. regulators
despite nearly equivalent riskless rates of return.
4.5.4 Potential Public Choice Explanations
The last category of potential explanations emerges from the public choice literature.
Regulators may be deliberately or inadvertently providing a “windfall” of sorts to electric
utilities. Stigler [1971], among others in the literature on regulatory capture, noted that firms
may seek out regulation as a means of protection and self-benefit. Close relationships
between regulators and the industries that they regulate have been observed repeatedly, and
one explanation for the size and growth of the risk premium is the electric utility industry’s
53 At the time of the survey, the 10-year U.S. Treasury rate was 2.0%. The average riskless rate reported by the 1,983 U.S. survey respondents was 2.4% (median 2.3%), but responses ranged from 0.0% to 8.0%. 54 Martha Coakley, et al. v. Bangor Hydro-Electric Co., et al., Initial Decision, 144 FERC ¶ 63,012 at 576 (2013).
4-30
increasing “capture” of regulatory power. For example, Hagerman and Ratchford [1978]
find that utility commissioners facing very long terms tend to be more generous to firms. In
contrast, states with elected utility commissioners [Kwoka, 2002] or commissioners whose
appointment by the executive requires approval by the legislature [Boyes and McDowell,
1989] tend to have lower electricity prices. We also note the similarities here to an Olson-
type collective action problem of concentrated benefits (excess profits to utilities may be
significant) and diffuse costs (the impact of those excess profits on each individual ratepayer
may be small) [Olson, 1965].
We are somewhat skeptical of this explanation, however, both because of the degree
of intervention in most utility rate cases by non-utility parties, and because the data do not
suggest that regulators have become progressively laxer over time. Exhibit GG compares
the rates of return on equity requested by utilities in our data set against the rates of return
ultimately authorized. As the trend line illustrates, this ratio has remained remarkably stable
over the thirty-five years of data, even as the risk premium itself has steadily increased. As
a result, the data do not suggest an obvious, growing permissiveness on the part of regulators
(although the last six years are suggestive of an increased level of accommodation among
regulators).
Exhibit GG: Rate of Return Authorized as a Percent of Rate of Return Requested
4-31
4.6 Conclusion
In this paper, we have examined a large database of electric utility rates of return
authorized by regulatory agencies over the past thirty-five years. These rates have
demonstrated a growing spread over the riskless rate of return across the time horizon
studied. The size and growth of this spread—the risk premium—does not appear to be
consistent with finance theory, as expressed by the CAPM. In fact, regression analysis of the
data suggests the opposite of what would be predicted if the CAPM holds.
The apparent persistence of large and growing risk premiums authorized by
regulators has broad implications across the electric power sector and beyond. In particular,
if rate case activity increases over the next several years as rate moratoria expire, the public
policy implications may be significant. Although our work suggests that the empirical
behavior of the risk premium is not explained by traditional asset-pricing models,
specifically the CAPM, we have identified a diverse collection of possible alternative
explanations that draws on work in finance, public policy, behavioral economics, and public
choice economics. We also suspect that these findings are not unique to electric utilities, and
therefore exploration of other regulated industries—both within and outside of the United
States—may be fruitful.
5-1
5 A Computational Framework for Risky Intertemporal Choice
5.1 Introduction
Computational simulation models have been used broadly within the social sciences
for decades. From the early efforts of Schelling [1978], more comprehensive work
blossomed in the general treatments of Epstein and Axtell [1996] and Gaylord and D’Andria
[1998], and has been complemented by more recent efforts such as Miller and Page [2007].
In many specific fields in the social sciences, computational simulation has made targeted
contributions, including to economics (e.g., Arthur et al. [1997] on the behavior of markets),
to organizational theory (e.g., Carley and Prietula [1994] on social networks and group
dynamics), to sociology (e.g., Butts [1998] on the emergence of collective beliefs), and to
political science (e.g., Kollman, Miller, and Page [1992] on the evolution of spatial voting
models). As a result of this now-vast literature, the use of computational simulation has
become widely accepted in virtually all of the social sciences.
Psychology, however, has proved more resistant to this innovation. Despite early
enthusiasm from Abelson [1968], twenty years later, it was still not widely adopted as a
methodology [Stasser, 1988]. In an effort to reverse this trend, Ostrom [1988] asserted the
advantages of computational simulation not only in testing, but also in formal theory
development and hypothesis generation. Yet again, now another thirty years later, it remains
underutilized, especially in the behavioral decision-making literature. Behavioral decision
making has remained almost exclusively dependent on experimental work. While such
efforts absolutely have a central role to play, I would propose that computational simulation
can also contribute to deepening our understanding of behavior, especially in novel and
complex environments.
For example, despite having models of decision making that address risky
intertemporal choice, the experimental results of tests of those models have been limited and
5-2
inconsistent. This essay develops a computational approach, in the spirit of Ostrom [1988],
to the generation of hypotheses on which future experimental work can be conducted. The
advantage of using a computational approach is the ability to identify quickly areas of the
problem space where further experimental exploration would be most fruitful. In addition,
computational simulation can allow for hypothesis testing in applications where
experimentation with human subjects is difficult or costly. The significant advances made to
date in developing descriptive theories of choice have enabled a computational methodology
to extend previous findings and assist with statistical inference from the results.
In this chapter, I develop a probabilistic computational simulation of risky
intertemporal choice and address two topics made possible with its use: (i) the inclusion of
probabilistic information about the population of decision makers that constitutes the
“decision demographics” of the population, and (ii) the use of computational methods to
facilitate experimental exploration in novel or complex decision environments.
In Section 5.2, I review the literatures on choice under risk and choice over time, as
well as the comparatively limited literature on risky intertemporal choice, and introduce the
building blocks of the decision models used. Section 5.3 develops a model of the decision
problem faced by a population of heuristic decision makers, as well as the measures used to
evaluate decision performance. Section 5.4 describes the simulation analysis performed and
communicates the results. Finally, Section 5.5 concludes by discussing the implications of
the methodology and their practical applications.
5.2 Literature Review
Historically, the theories of choice under risk and uncertainty and choice over time
have evolved separately. Most real-world decision-making applications, however, involve
resolving questions of both uncertainty and time. These include personal decision making
(such as retirement planning and health-related decision making), as well as business
decision making (such as capital investment and policy-related decision making), among
many others. The remainder of this section reviews the separate literatures on uncertainty
5-3
and time, before reviewing the comparatively limited literature on risky intertemporal
choice.
5.2.1 Choice Under Risk and Uncertainty
The normative literature on risky choice has existed for more than half a century and
is vast and well-settled. From the pioneering efforts of von Neumann and Morgenstern
[1947] in the axiomatic derivation of expected utility theory, to Savage’s [1954]
axiomatization of subjectively-expected utility and probability, to Arrow’s [1953]
development of the state-preference approach to uncertainty used in general equilibrium
theory, to Pratt’s [1964] work on risk aversion, the normative theory of choice under risk
and uncertainty has remained virtually unchanged. Summarized in simple terms: lotteries
can be compressed into expected utility. For outcomes x, probabilities p, and concave utility
function u, the expected utility across i states of the world is ∑ .
Notwithstanding this work on normative choice, an extensive literature subsequently
developed questioning the descriptive accuracy of the normative approach. The literature on
non-expected utility theories is vast (see the review in Starmer [2000]), but the general result
has been to transform payoffs by emphasizing losses and to transform probabilities from a
linear to an S-shaped function. These approaches were formalized by Kahneman and
Tversky [1979] into prospect theory, and subsequently into cumulative prospect theory
[Tversky and Kahneman, 1992]. Quiggin [1982] and Lattimore, Baker, and Witte [1992],
among others, have also developed alternative probability-weighting functions, but the vast
majority of this research has resulted in the general characteristics of reference points
(characterization of gains and losses as relative to a salient reference point), loss aversion,
and S-shaped probability-weighting functions.
In the prospect theory formulation, the prospect value V of a lottery is a function of
transformed values of p and x, where is a probability-weighting function and is a value
function. Accordingly, ∑ . Tversky and Kahneman [1992] later
proposed specific functional forms for and and a large literature developed to
5-4
parameterize those functions based on experimental data or to infer them from real-world
choice examples. The approaches in this section—both normative and descriptive—
however, deal solely with choice under risk and uncertainty at a single point in time.
5.2.2 Choice Over Time
For choices over time, a similar foundational scenario emerged. The normative
theory of intertemporal choice emerged from Samuelson’s [1937] introduction of the
exponential discounting of utility and Koopman’s [1960] demonstration that certain basic
assumptions lead to impatience over infinite consumption programs (i.e., preference for
immediate values over delayed values). From these models, which involved only payoffs
with certainty, a general model emerged such that the expected utility of a sequence of
certain payoffs over time periods t could be calculated as ∑ , where
1 and r is the discount rate (or rate of time preference). In other words, multi-period
payoffs could be compressed into a single discounted utility. This normative formulation has
the important characteristic of time-invariant discount rates.55
And once again, a large series of empirical results demonstrated the descriptive
inadequacy of the normative model. Thaler [1981] found evidence of both dynamic
inconsistency and gain-loss asymmetry relative to the normative exponential discounting
model in intertemporal choice. Loewenstein [1988] found that incorporation of a reference
point (as in prospect theory) led to what he referred to as delay-speed-up asymmetry, where
subjects would require more compensation for postponing consumption than they would be
willing to pay for accelerating consumption. Ainslie [1991] and Kirby and Maraković [1995]
proposed that the hyperbolic discounting model provided a better fit to the experimental data
on intertemporal choice. Read [2001] asserted that a subadditive discounting model provided
a better fit than the hyperbolic model, although both explained the experimental results. The
primary conceptual difference between the hyperbolic and subadditive models is the
55 For any discount function , f is time-invariant if ⁄ is a constant with regard to t. For example, for the exponential discount function ⁄ . However, for the Lowenstein-Prelec hyperbolic discount function 1 ⁄ 1 ⁄⁄ 1 , and therefore it remains a function of t.
5-5
motivation for the present bias as being either impulsiveness (for the hyperbolic model) or
the perception of time (for the subadditive model).
In general, the hyperbolic model became commonly accepted as the preferred
descriptive model of intertemporal choice under certainty. I use the hyperbolic model
specified in Prelec [1989] and Loewenstein and Prelec [1992]: 1 / , where
r is the traditional (exponential) discount rate and adjusts the rate at which near and distant
outcomes are compared. As → 0, converges to the exponential discounting function.
The hyperbolic formulation does not display time invariance. Rather, the discount rates
determined under hyperbolic discount functions are all functions of t. Other functional forms
are compared and contrasted in Section 5.3.1 in the context of experimental work using the
different forms.
5.2.3 Risky Intertemporal Choice
Despite significant advancements in choice under risk and choice over time, what
has been left relatively unexplored is their combination: risky intertemporal choice.
Although the normative approach is essentially settled with the recursive utility framework
of Kreps and Porteus [1978], as extended by Epstein and Zin [1989], the descriptive theory
of risky intertemporal choice is “not settled” [Albrecht and Weber, 1997], is “complex and
not easily understood” [Weber and Chapman, 2005], is “relatively limited” [Anderson and
Stafford, 2009], and is “quite heterogeneous in methods and findings” [Hardisty and Pfeffer,
2017].
Four general types of anomalies in risky intertemporal choice have been identified:
1. Certainty-Risk Asymmetry (discount rates are higher for certain outcomes than for risky outcomes)
2. Short-Long Asymmetry (violations of stationarity, preference reversals induced by shifting risky choices forward in time)
3. Gain-Loss Asymmetry (discount rates are smaller for gains than for losses) 4. Magnitude Asymmetry (discount rates are higher for larger payoffs)
5-6
The experimental work examining these findings, however, is limited and often
contradictory. Albrecht and Weber [1997] find support for short-long asymmetry in
matching tasks, but not choice tasks. Öncüler and Onay [2009] support violations of
stationarity, but find the evidence on direction of the effect mixed. Keren and Roelofsma
[1995] find support for certainty-risk asymmetry. Anderson and Stafford [2009] find no
support for certainty-risk asymmetry. Shelley [1994] supports gain-loss asymmetry;
Blackburn and El-Deredy [2013] do not. Disagreements involving the interactions between
risk and time emerge from whether subjects evaluate risk first or time first [Weber and
Chapman, 2005] and whether there is a single underlying mechanism [Keren and Roelofsma,
1995] or many mechanisms [Read, 2003]. Albrecht and Weber [1997] note that the empirical
results are sensitive to the elicitation procedure used. Wang et al. [2016] report global
experimental results that show significant correlations in time preference with cultural
variables, particularly in regard to cultural traits that are related to high levels of uncertainty
avoidance.
In short, comparatively little clarity has emerged from the experimental literature on
compression and risky intertemporal choice other than a general recognition that hyperbolic
discounting is well-supported and that resolution of uncertainty involves an S-shaped
decision-weighting function and differential treatment of gains and losses. I proceed,
therefore, with models based on those generally-accepted characterizations.
In general, the normative approach is characterized by the discounted expected utility
framework and uses the value function provided in [1]. For the descriptive approach, the
prospect theory value function is combined with the hyperbolic discounting model, which
produces the value function in [2] [Prelec and Loewenstein, 1991; Öncüler, 1999].
[1]
5-7
[2]
To parameterize the prospect theory value function, the probability-weighting function [3]
and value function [4] proposed by Tversky and Kahneman [1992] are maintained. In those
formulations, , , and represent, respectively, the curvature of the gain and loss functions
(i.e., the degree of reference point-dependent risk aversion) and the loss aversion parameter.
The in the probability-weighting function controls the degree to which the function is S-
shaped and is contingent on whether the outcome x is characterized as a gain or a loss.
1 / [3]
00
[4]
5.3 Simulating the Population of Decision Makers
5.3.1 Parameterization and Decision Demographics
Numerous experiments have been performed in the literature to estimate the
parameters required by cumulative prospect theory. In Tversky and Kahneman’s [1992]
original work, the parameters for [3] and [4] were estimated as 0.88, 2.25,
0.61, and 0.69, where and indicate, respectively, the probability
weighting function coefficients for gains and losses. However, there is certainly no
theoretical reason to believe that only a single set of parameters exists for all decision
makers. Rather, each decision maker in a population would be expected to have parameter
values drawn from a distribution of such values. Much like describing the population
demographics of a group (e.g., percent of each ethnicity, percent of age group), I am
interested in characterizing the decision demographics of the population (e.g., percent loss
averse, percent risk seeking for losses). I examined the literature to estimate the distribution
5-8
of parameter values. Exhibit HH summarizes the parameter value estimates from a variety
of such studies (recognizing that not every study estimated all parameter values). Exhibit II
provides the results of a single group of identical studies performed on populations of
decision makers worldwide [Rieger, Wang, and Hens, 2017]. The resulting distributional
information in Exhibit JJ incorporates the estimates in both Exhibits HH and II.56 Summary
statistics are provided in Exhibit KK for the combined data, with and without the country-
level detail in Rieger, Wang, and Hens [2017] (noted as “ex RWH”).57
Exhibit HH: Summary of Studies Estimating Cumulative Prospect Theory Parameters
56 To be precise, the dispersion common to the experimental results cited is such that by convention the median coefficients are typically reported. This is true for the Rieger, Wang, and Hens [2017] data, for example. Accordingly, the distributions presented in Exhibit JJ are actually the distributions of the median coefficients, not the population distribution or the distribution of mean coefficients. While the Central Limit Theorem would ensure normality if the mean coefficients were reported, additional (but not overly onerous) regularity conditions must be imposed in order to adapt the Central Limit Theorem to order statistics such as the median. That being said, for a sample of size 2 1 drawn from an infinite population with density function
, the sampling distribution of the median is nevertheless approximately normal with mean and variance 1 8⁄ [David and Nagaraja, 2003; Genton, Ma, and Parzen, 2006]. I proceed as if these distributions are normally distributed. 57 Correlations are calculated only for the Rieger, Wang, and Hens [2017] data because they are complete and consistent.
Year Study4.80 1979 Fishburn and Kochenberger [1979]
0.88 0.88 2.25 0.61 0.69 1992 Tversky and Kahneman [1992]0.22 0.56 1994 Camerer and Ho [1994]0.50 0.71 1996 Wu and Gonzalez [1996]0.39 0.84 1998 Fennema and van Assen [1998]0.49 1999 Gonzalez and Wu [1999]0.89 0.92 0.60 0.70 2000 Abdellaoui [2000]
0.67 2000 Bleichrodt and Pinto [2000]0.61 0.61 2001 Donkers et al. [2001]
2.17 2001 Bleichrodt et al. [2001]1.43 2002 Schmidt and T raub [2002]1.81 2003 Pennings and Smidts [2003]
0.97 2004 Etchart-Vincent [2004]0.94 0.96 2005 Abdellaoui et al. [2005]0.68 0.74 3.20 2005 Tu [2005]1.01 1.05 2006 Fehr-Duda et al. [2006]0.81 0.80 1.07 0.76 0.76 2006 Andersen et al. [2006]0.72 0.73 2.04 2007 Abdellaoui et al. [2007]0.86 1.06 2.61 2008 Abdellaoui et al. [2008]0.24 0.26 2008 Pitcher [2008]0.71 0.72 1.38 0.91 0.91 2009 Harrison and Rutström [2009]0.88 -0.26 1.87 2009 Booij and van de Kuilen [2009]0.86 0.83 1.58 0.62 0.59 2010 Booij et al. [2010]0.80 0.78 2.39 0.01 5.48 2010 Chow et al. [2010]0.82 0.88 1.83 0.58 0.73 2011 Nilsson et al. [2011]0.73 1.52 0.74 1.24 2012 Glöckner and Pachur [2012] (values are midpoint)0.21 0.06 1.34 2016 Harrison and Swarthout [2016]0.73 0.73 1.11 2017 Murphy and ten Brincke [2017]
5-9
Exhibit II: Parameter Estimates by Country from Rieger, Wang, and Hens [2017]
Year Country0.60 0.60 1.25 0.35 0.88 2016 Angola0.48 0.68 1.11 0.45 0.71 2016 Argentina0.41 0.45 1.08 0.62 1.00 2016 Australia0.37 0.37 1.28 0.54 0.78 2016 Austria0.56 0.59 1.08 0.41 0.98 2016 Azerbaijan0.44 0.55 1.42 0.64 0.94 2016 Belgium0.54 0.37 0.99 0.39 0.90 2016 Bosnia-Herzegovina0.42 0.83 1.62 0.44 0.60 2016 Canada0.54 0.90 1.72 0.52 0.73 2016 Chile0.54 0.55 1.43 0.52 0.94 2016 China0.42 0.37 1.26 0.40 0.78 2016 Colombia0.53 0.33 1.36 0.34 0.90 2016 Croatia0.56 0.45 1.49 0.46 1.00 2016 Czech Republic0.51 0.90 1.71 0.57 0.73 2016 Denmark0.35 0.37 1.52 0.40 0.78 2016 Estonia0.50 0.61 1.49 0.52 0.82 2016 Finland0.41 0.49 1.38 0.48 0.98 2016 France0.80 0.90 3.80 0.30 0.73 2016 Georgia0.42 0.49 1.38 0.44 0.71 2016 Germany0.50 0.30 1.29 0.44 0.82 2016 Greece0.35 0.34 1.27 0.39 0.78 2016 Hong Kong0.39 0.49 1.37 0.45 0.71 2016 Hungary0.41 0.49 1.38 0.52 0.71 2016 India0.42 0.49 1.32 0.40 0.71 2016 Ireland0.42 0.37 1.31 0.44 0.78 2016 Israel0.42 0.55 1.43 0.44 0.94 2016 Italy
0.26 0.55 1.37 0.71 0.94 2016 Japan0.30 0.21 1.10 0.38 0.81 2016 Lebanon0.48 0.31 1.29 0.36 1.00 2016 Lithuania0.46 0.46 0.94 0.55 0.75 2016 Luxembourg0.32 0.31 1.06 0.62 0.70 2016 Malaysia0.31 0.37 1.14 0.39 0.68 2016 Mexico0.71 0.60 1.95 0.52 0.88 2016 Moldova0.47 0.90 1.47 0.82 0.73 2016 Netherlands0.44 0.33 0.99 0.47 0.76 2016 New Zealand
0.71 0.37 1.05 0.30 0.78 2016 Nigeria0.39 0.45 1.27 0.55 1.00 2016 Norway0.47 0.55 1.59 0.45 0.94 2016 Poland0.48 1.72 2.31 0.50 0.34 2016 Portugal0.48 1.72 3.19 0.50 0.34 2016 Romania0.39 0.30 1.41 0.41 0.82 2016 Russia0.32 0.42 1.31 0.54 1.00 2016 Slovenia0.44 0.68 1.28 0.70 0.71 2016 South Korea0.44 0.74 1.63 0.47 0.88 2016 Spain0.48 0.82 1.60 0.58 0.80 2016 Sweden0.37 0.49 1.37 0.54 0.98 2016 Switzerland0.26 0.49 1.33 0.71 0.71 2016 Taiwan0.87 0.90 1.01 0.84 0.73 2016 Tanzania0.44 0.49 1.61 0.64 0.71 2016 Thailand0.55 1.06 1.51 0.55 0.94 2016 Turkey0.44 0.49 1.06 0.47 0.98 2016 UK0.42 0.49 1.36 0.44 0.71 2016 USA0.56 0.55 1.29 0.41 0.94 2016 Vietnam
5-10
Exhibit JJ: Empirical Distributions of Cumulative Prospect Theory Parameters
0%
20%
40%
60%
80%
100%
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 1.10
Freq
uenc
y
Alpha Parameter (risk aversion for gains)
10th percentile 90th percentile
Tversky & Kahneman [1992]
0%
20%
40%
60%
80%
100%
0%
5%
10%
15%
20%
25%
30%
35%
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
Freq
uenc
y
Gamma (Gains) Parameter
10th percentile
90th percentile
Tversky & Kahneman
[1992]
0%
20%
40%
60%
80%
100%
0%
5%
10%
15%
20%
25%
0.00 0.30 0.60 0.90 1.20 1.50 1.80
Freq
uenc
y
Beta Parameter (risk aversion for losses)
10thpercentile
90th percentile
Tversky & Kahneman
[1992]
0%
20%
40%
60%
80%
100%
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
0.00 0.20 0.40 0.60 0.80 1.00 1.20
Freq
uenc
y
Gamma (Losses) Parameter
10th percentile
90th percentile
Tversky & Kahneman [1992]
0%
20%
40%
60%
80%
100%
0%
5%
10%
15%
20%
25%
0.80 1.30 1.80 2.30 2.80
Freq
uenc
y
Lambda Parameter (loss aversion)
10th percentile
90th percentile Tversky & Kahneman [1992]
5-11
Exhibit KK: Descriptive Statistics and Correlations for the Cumulative Prospect Theory Parameters
It is worth noting that the results, generally, differ materially with the initial estimates
of Tversky and Kahneman [1992], suggesting that incorporating distributional information
is important in terms of characterizing the population of decision makers. For calculation
purposes to eliminate numerical outliers, ranges explored for each of the parameters are
truncated to the 10th and 90th percentiles of the distributions in Exhibit JJ (and indicated in
the blue-shaded rows in Exhibit KK). Based on this assumption, the space of decision makers
can be described in cumulative prospect theory terms as being contained within the gray-
shaded spaces in Exhibit LL (for the value function58), Exhibit MM (for the probability
weighting function for gains), and Exhibit NN (for the probability weighting function for
losses), although not every point in those spaces is equally likely to be represented by a
decision maker.
58 The graph presents results for the interval ∈ [-$100, $100], but obviously could be extended arbitrarily.
Count 76 72 71 66 62Count (ex RWH) 22 18 17 12 8Average 0.53 0.62 1.58 0.51 0.88Average (ex RWH) 0.68 0.74 2.02 0.59 1.39Median 0.48 0.55 1.38 0.50 0.78Median (ex RWH) 0.73 0.81 1.83 0.61 0.75Minimum 0.21 -0.26 0.94 0.01 0.34Maximum 1.01 1.72 4.80 0.91 5.4810th Percentile 0.32 0.33 1.07 0.37 0.7010th Percentile (ex RWH) 0.25 0.45 1.25 0.29 0.6690th Percentile 0.84 0.92 2.25 0.71 1.0090th Percentile (ex RWH) 0.89 0.99 2.85 0.76 2.51
RWH = Rieger, Wang, and Hens [2017]
1.00
0.32* 1.00
0.31* 0.66*** 1.00 -0.09 0.26 -0.12 1.00
0.01 -0.57*** -0.42** -0.07 1.00
*, **, and *** indicate significance at the 95%, 99%, and 99.9% levels
5-12
Exhibit LL: Parameter Space for the Value Function
Exhibit MM: Parameter Space for the Probability Weighting Function (Gains)
-200
-150
-100
-50
0
50
100
-100 -80 -60 -40 -20 0 20 40 60 80 100
Val
ue F
unct
ion
Payoff
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Dec
isio
n W
eigh
t (G
ains
)
Probability
5-13
Exhibit NN: Parameter Space for the Probability Weighting Function (Losses)
A similar analysis can, conceptually, be performed with hyperbolic discounting.
However, estimation of the degree of “present bias” found in experimental results is
significantly complicated by the variety of functional forms assumed by researchers.59 There
are also issues with separating time discounting from the payoff amounts (or “amount-
dependent discounting” [Frederick, Loewenstein, and O’Donoghue, 2002]) in experimental
work, evaluating the relative credibility of promised future payoffs between experimental
designs, and specifying subject utility functions. Takeuchi [2011] even notes that
“correction” of these factors can induce future bias instead of present bias, although both
may coexist in subjects.
Notwithstanding these issues, an attempt must be made to compare the diverse
experimental results on a common footing in order to estimate the distribution of present
bias parameters across the population. I begin by employing the most general functional
form for time discounting in the literature: the generalized Weibull with fixed costs
[Benhabib, Bisin, and Schotter, 2010]. The generalized Weibull form for the discount factor
59 To be clear, by “present bias” we mean the empirical finding underlying hyperbolic discounting that decision makers have higher short-term discount rates (or lower short-term discount factors) and lower long-term discount rates (or higher long-term discount factors) than would be expected under an exponential discounting model.
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Dec
isio
n W
eigh
t (Lo
sses
)
Probability
5-14
D [5] is useful because it allows for both present bias and future bias and nests the
exponential, hyperbolic, and quasi-hyperbolic models within it.
, ; , , , , 1 1 [5]
In [5], y is the payoff in question, t indexes time, is the curvature parameter of the discount
function ( 1 implies exponential discounting, 2 implies hyperbolic discounting), r
is the time discount rate, b is the fixed cost present bias parameter (for 0), and and q
are present bias parameters ( 1 and 1 indicate present bias).
Exhibit OO is a sort of “Rosetta Stone” that presents the functional forms employed
by previous experimental work and shows how each of the forms emerges from the
generalized Weibull model in [5]. In order to make appropriate comparison of the hyperbolic
bias parameters, each study’s functional form must be converted into the same shared form.
As a balance between modeling detail and parsimony, I use the Loewenstein and Prelec
[1992] model as a common ground for this study (the shaded row in Exhibit OO). When I
introduced this model in Section 5.2.2, I retained the notation common to the literature,
1 / , however, as the prospect theory functions already employ in the
value function (as the risk aversion parameter for gains), I will henceforth use instead to
indicate the hyperbolic bias parameter (or what is sometimes referred to as the degree of
hyperbolicity) in order to avoid confusion. The hyperbolic discounting function used,
therefore, is given in [6], where r is the time discount rate and is the hyperbolic bias
parameter (as → 0, [6] becomes exponential).
; , 1 / [6]
Each study in Exhibit OO was modeled according to its own parameters and
formulation. Then, the resulting data for each experiment was fit to [6] by minimizing the
sum of the resulting squared differences between the original functional form and [6]. The
5-15
result of this procedure was a distribution of hyperbolic bias parameters all common to a
single functional form. This distribution is illustrated in Exhibit PP.60
Common Name
Generalized Weibull with Fixed Costs parameters Discount Factor Formula Used by
y t q r b Generalized
Weibull with Fixed
Costs
y t q r b 1 1 Benhabib, Bisin, and
Schotter [2010]
Generalized Weibull y t q 1 r 1 0 1 /
Tanaka, Camerer, and Nguyen [2010]; Takeuchi
[2011]
Loewenstein and Prelec y t 1 1 r 1 0 1 /
Loewenstein and Prelec [1992]; Cairns and van
der Pol [2000]
Myerson and Green y t 1 1 s k 1 0 1
Myerson and Green [1995]; McKercher et al.
[2009]
Rachlin y 1 2 1 0 1 McKercher et al. [2009]
Harvey y t 1 1 b 1 0 1 Cairns and van der Pol [2000]; van der Pol and
Cairns [2002]
Mazur y t 1 2 k 1 0 1
Kirby and Maracović [1995]; Myerson and
Green [1995]; Kirby and Maracović [1996]; Cairns and van der Pol [2000];
Johnson and Brickel [2002]; van der Pol and
Cairns [2002]; Madden et al. [2003]; Anderson et al. [2008]; McKercher et al. [2009]; Tanaka, Camerer,
and Nguyen [2010]
Phelps-Pollak, or
the ( ) model
y t 1 → 1 ln
0
Paserman [2008]; Andreoni and Sprenger [2012]; Fang and Wang [2014]; Wang, Rieger,
and Hens [2016]; Laibson [2017]; Laibson et al.
[2017]
Exponential y t 1 → 1 r 1 0
Exhibit OO: Time Discounting Model Hierarchy
60 The data categorized under “No Fit” in Exhibit PP indicates the small portion of the sample (approximately 7%) for which the Loewenstein and Prelec [1992] model did not provide a reasonable fit. Those data points are excluded from the following analysis. The of all of the fits ranged from 87.6% to 100.0%, with an average of 97.7%.
5-16
Exhibit PP: Distribution of Hyperbolic Present Bias Parameters
Based on these results from the experimental work listed in Exhibit OO, for
calculation purposes in my study, the range explored for the hyperbolic parameter given
by the interval [0.1, 0.5]. → 0 is specifically excluded from the interval because I consider
the exponential discounting model separately. The population of decision makers, therefore,
is characterized as having the set of parameters , , , , , for the equations [2], [3],
[4], and [6] as defined in Exhibit QQ. In the next section, I turn to the assumptions made
about the “logistical” characteristics of their decision-making processes.
Parameter Purpose Lower Bound Upper Bound Risk aversion – gains 0.32 0.84 Risk seekingness – losses 0.33 0.92 Loss aversion 1.07 2.25 Probability weight – gains 0.37 0.71 Probability weight – losses 0.70 1.00 Hyperbolic bias 0.10 0.50
Exhibit QQ: Summary of Decision Maker Parameter Bounds
0%
20%
40%
60%
80%
100%
0%
10%
20%
30%
40%
50%
60%
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 NoFit
Cum
ulat
ive
Freq
uenc
y
Hyperbolic Bias Parameter usingthe Loewenstein and Prelec [1992] Functional Form
5-17
5.3.2 Decision Processes and Sequencing
Whereas the previous section characterized the population of decision makers
according to their preferences (as represented by the parameters of their objective functions),
in this section I turn to the functional representation of their decision processes and
sequencing. The conventional partition of decision makers has been that normative decision
makers follow both exponential discounting and expected utility theory. Likewise, the
descriptive model assumes that heuristic decision makers follow hyperbolic discounting and
prospect theory.
Exponential Discounting Hyperbolic Discounting
Expected Utility Theory Normative Normative Risk / Descriptive Time
Prospect Theory Descriptive Risk / Normative Time
Descriptive
Exhibit RR: Basic Taxonomy of Decision Processes
The off-diagonal entries in Exhibit RR, however, are also possible. Decision makers
might discount time in normative fashion (a hyperbolic function where → 0), but not
uncertainty. Similarly, decision makers might exhibit hyperbolic bias for time ( ≫ 0), but
respond normatively to uncertain choice tasks (a prospect theory formulation for which →
1, → 1, etc.). Accordingly, I allow for “quasi-normative” decision makers by adding [7]
and [8] to complement [1] and [2], thereby allowing for decision makers’ decision processes
to be affected by task presentation in such a way as to “resolve” one form of bias but not the
other (e.g., levelizing may suppress biases involving uncertainty, but not the perception of
time).
hyp [7]
exp [8]
5-18
Likewise, the sequencing of the decision process itself may be influenced by task
presentation to the extent that it raises the prominence of one dimension over another. Do
decision makers resolve time first or uncertainty? Conventionally, it is assumed that
uncertainty is resolved first, and then time second [9].
Uncertainty1st
Time2nd
[9]
But suppose that task presentation alters the decision maker’s perception of time such that
time is resolved first, and uncertainty is resolved second [10].
Time1stUncertainty2nd
[10]
Under the descriptive decision model (for example, if decision makers’ perception of time
varies [Read, 2001]), changing the decision sequence can result in different choices. In a
risky intertemporal choice framework, Weber and Chapman [2005] found that adding risk
to a choice task eliminated the immediacy effect (i.e., the violation of stationarity that gives
rise to the hyperbolic discounting model), while adding delay (time) to a choice task
eliminated the certainty effect (i.e., the violation of the independence axiom that gives rise
to the Allais paradox and can be explained by the incorporation of a nonlinear probability-
weighting function). These findings, however, were dependent on the presentation of the
various prospects to their subjects. As a result, I allow for the different sequences of decision
making in [9] and [10] to investigate these findings further.
5.4 Analysis and Findings
Having described the population of decision makers and their decision processes, I
now turn to their analysis. First, as a normative basis for comparison, I consider a utility
5-19
function [11] that exhibits constant absolute risk aversion and for which the Arrow-Pratt
coefficient of absolute risk aversion is ′′ ⁄ .
;
11 [11]
The analysis looks at the population of decision makers and illustrates the usefulness
of computational methods in informing experimental design. Based on the experimental
results in the literature, I create a “population” of decision makers. Each decision maker is a
(1 7)-vector of decision model parameters , , , , , , , where each parameter
is distributed according to the experimental results.61 As noted previously, because studies
have typically reported median parameter values, it is a distribution of medians, although I
assume (at minimal cost) that the distributions are asymptotically normal according to the
limiting distribution for the median order statistic.
The population, therefore, is multivariate normal, and each decision maker
~ , , with the vector of mean parameters and its variance-covariance matrix, both
parameterized and subject to truncation as specified in Section 5.3. If prospect theory and
hyperbolic discounting are accurate descriptive models of choice, then this specification
characterizes the “decision demographics” of the population and Monte Carlo simulation
can be used to conduct “virtual” experiments.
For each such virtual experiment, I generate 100,000 samples of x using a seven-
dimensional Sobol’ sequence of uniform [0,1] random variates transformed into a ,
distribution. A Sobol’ sequence [Sobol’, 1967] is a low-discrepancy or “quasirandom”
sequence that trades off statistical randomness for minimal statistical discrepancy (or equi-
distribution) in order to increase rates of convergence in Monte Carlo simulation.
61 The risk aversion coefficient is assumed to be truncated by 0.0001 and 0.02. The lower bound of 0.0001 is selected because it is approximately risk neutral. The upper bound of the risk aversion coefficient was calculated such that losses carried 7 times the weight of gains, reflecting extreme risk aversion. Given [11], and letting be the maximum level of payoff (or terminal wealth), I solved ; ; 7⁄ for , which is ln 7 ≅ 2 . When $100, ≅ 0.02.
5-20
To be sure, individual decision makers are likely to have a unique set of parameters.
The researcher undertaking an experiment with human subjects is drawing a sample of
from an “unknown” distribution. Thus, they must contend with not knowing whether the
effect they seek does not exist or whether the individuals subject to such an effect are simply
infrequently found within the population. By modeling decision makers at the population
level, this difficulty can be avoided.
I see two primary benefits of these virtual experiments. First, the incorporation of
probabilistic information (i.e., the decision demographics) can highlight areas where results
are “significant, but insignificant.” In other words, indicating where an experimental result
can be confirmed, but is only likely to affect a small portion of the population of decision
makers and is therefore of limited practical significance. Second, the virtual experiments
help to identify what should be expected from applying existing theory to novel problems.
In particular, the ability to evaluate the role of variable interactions on choice problem
outcomes is useful.
5.4.1 Incorporating Probabilistic Information into Decision Problems
First, consider the probabilistic information contained in the simulated population
results. Prior to examining the distribution of the population of decision makers, there was
no means by which to determine whether positive (or negative) findings were of practical
significance. That is, it is less interesting whether or not an experiment can be constructed
to elicit some outcome; rather, the result of interest should be how large the percentage of
the population (or what sub-population) produces a given result. By incorporating the
distribution of parameters throughout the population, one can observe both the distribution
of the prospect theory components themselves (Exhibits SS, TT, and UU), but also the
resulting distribution of population behavior for any given choice problem. For example,
consider the choice problem eliciting a preference for either a fixed immediate payment F
or a lottery at time t (the per period discount rate is 10%) of a 90% chance of receiving $200
and a 10% chance of owing (or losing) $1,000. The expected value of the lottery is $72.4 at
5-21
t = 1. The percentage of the population preferring the lottery measured relative to both F and
t is illustrated in Exhibit VV.
Exhibit SS: Probabilistic Representation of the Probability Weighting Function for Gains for the Simulated Population of Decision Makers
Exhibit TT: Probabilistic Representation of the Probability Weighting Function for Losses for the Simulated Population of Decision Makers
5-22
Exhibit UU: Probabilistic Representation of the Value Function for the Simulated Population of Decision Makers
Exhibit VV: Population Decision Demographics for a Choice Problem
5-23
Among the findings from this simulation are: (i) that the decision weighting function
exhibits considerably more dispersion for gains than for losses, (ii) that most of the activity62
in the decision weighting function for gains involves small probabilities, while a sizable
portion of the activity in the decision weighting function for losses involves larger
probabilities, and (iii) that only a relatively small portion of the population exhibits
meaningful loss aversion, although for those that exhibit loss aversion it tends to be severe.
Applying the population of parameters to a decision problem reveals (in this case) that the
percentage of population electing the lottery varies considerably according to the fixed
alternate payment F, but varies only slightly with time delay.
5.4.2 Evaluating Novel and Complex Problems
The second benefit is the ability to generate and test hypotheses with regard to the
outcome of novel choice tasks. I illustrate this ability with the results of two of the main
anomalies noted in the risky intertemporal choice literature. The first is the Certainty-Risk
Asymmetry, which proposes that discount rates are higher for certain outcomes than for risky
outcomes. In other words, a risky gain shifted into the future will become relatively more
attractive than a certain gain (and the opposite occurs for losses). If supported, this anomaly
would turn finance on its head, as discount rates are (normatively) a function of risk and
therefore should increase in risk and decrease in certainty. The experimental evidence for
Certainty-Risk Asymmetry is mixed, however, with Keren and Roelofsma [1995] finding
support for it, but Anderson and Stafford [2009] finding no support for it.
Consider the following two choices adapted from Anderson and Stafford [2009]63:
62 By “activity,” I mean that a plurality of the population is located in the referenced area. For example, in Exhibit SS, the plurality of decision makers is concentrated at overestimating small probabilities, while a smaller number exhibit far more diverse behavior for larger probabilities. 63 I begin the comparisons five periods into the future to avoid issues with the “immediacy effect” found by Karen and Roelofsma [1995].
5-24
Choice #1: 50 at or 50 at Δ Choice #2: 50 at or p chance of at Δ
1 chance of 0 at Δ
The lottery option is constructed to have the same expected present value as the fixed
options, except for the variable , which is used to manipulate the state 1 lottery payoff to
less than, the same, or greater than the expected present value. Certainty-Risk Asymmetry
is present if the discount rate implied by the risky option (Choice #2) is lower than the
discount rate implied by the certainty option (Choice #1), causing a preference reversal. For
example, in Choice #1, a preference for the deferred payoff would imply a discount rate of
less than 10%, while in Choice #2 a preference for the fixed payoff would imply a discount
rate of greater than 10% (at 0), the results together implying an asymmetry.
I performed simulations varying the state 1 probability p (5%, 50%, and 95%), the
adjustment to the state 1 lottery payoff value (-5, 0, +5), and the delay period Δ (1, 2, 3,
4, 5). As output, I captured the percentage of the simulated population switching their
preferences from Choice #1 to Choice #2. Exhibit WW illustrates the percentages of the
population showing preference reversals under each tested condition. An examination of
these results indicates how obtaining mixed results is feasible in a heterogeneous population
of decision makers. Looking primarily at the central column in Exhibit WW (the column
indicating no adjustment to the state 1 lottery payoff), several general findings emerge:
1. Decision makers following the normative rule (EDEU) do not exhibit Certainty-Risk Asymmetry (as expected).
2. Among heuristic decision makers following the traditional “uncertainty first” decision sequence, only between roughly 50% and 90% of decision makers exhibit Certainty-Risk Asymmetry.
3. Among mixed-rule decision makers (i.e., those that exhibit normative choice for either risk or time, but non-normative choice for the other), such as HDEU, the proportion exhibiting Certainty-Risk Asymmetry falls into the single digits.
4. Resolving time-related biases (shifting HD to ED) plays a bigger role in reducing the percentage of the population exhibiting Certainty-Risk Asymmetry than resolving uncertainty-related biases (shifting PT to EU).
5-25
Exhibit WW: Percent of Population Exhibiting Certainty-Risk Asymmetry The State 1 Adjustment reflects the amount added to either the fixed deferred payment or the deferred lottery payment. At zero, the expected values are equal, thus the adjustment amount can reveal risk preference. Simulations were performed based on whether the decision makers were assumed to resolve uncertainty first or time first as described in Section 5.3.2. The abbreviations in the second column refer to the decision models employed: EDEU (exponential discount, expected utility), HDEU (hyperbolic discounting, expected utility), EDPT (exponential discounting, prospect theory), and HDPT (hyperbolic discounting, prospect theory).
Probability of State 1 (p )
State 1 Adjustment (S 1)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 100 64.9 42.5 29.2 20.2HDEU 0.1 65.1 79.9 84.5 86.3 91.7 91.3 90.1 90.5 90.1 99.8 98.0 95.8 94.0 92.7EDPT 16.6 14.5 12.6 10.9 9.5 16.8 14.7 12.7 11.0 9.6 29.1 10.0 12.4 11.1 9.7HDPT 20.5 32.6 43.6 47.2 48.5 55.6 55.4 55.0 54.2 53.0 76.0 67.0 62.3 59.1 56.5EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 100 100 100 100 100HDEU 0.3 79.1 89.9 92.7 93.7 99.7 98.8 98.0 97.2 96.5 99.6 98.8 97.9 97.2 96.5EDPT 19.0 19.0 19.0 19.1 19.1 19.2 19.2 19.2 19.2 19.2 80.6 80.6 80.6 80.1 80.6HDPT 21.6 64.0 67.7 66.6 64.4 78.2 75.3 72.3 69.1 65.9 77.9 75.2 72.1 68.9 65.7
Probability of State 1 (p )
State 1 Adjustment (S 1)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 92.0 59.0 38.1 25.9 17.7HDEU 3.1 52.2 57.1 55.0 52.0 80.1 70.4 63.0 57.6 53.3 79.9 71.0 63.8 58.0 53.7EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 46.1 5.1 0.4 0.0 0.0HDPT 0.0 37.9 58.8 67.5 72.4 71.5 73.8 75.7 77.4 78.8 96.9 89.2 85.6 84.1 83.5EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 83.3 84.9 86.3 87.6 88.8HDEU 4.4 59.8 54.7 45.5 37.4 82.8 67.0 54.5 44.2 36.0 70.0 58.0 47.8 39.4 32.5EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 100 100 100 100 100HDPT 0.0 80.1 91.9 95.4 97.0 100 100 100 100 99.8 100 100 100 99.9 99.7
Probability of State 1 (p )
State 1 Adjustment (S 1)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 18.9 13.0 7.4 5.2 3.2HDEU 0.0 8.3 4.0 2.7 2.0 9.3 4.1 2.8 2.1 1.8 0.7 1.9 2.0 1.8 1.7EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 46.1 5.1 0.4 0.0 0.0HDPT 0.0 37.9 58.7 67.0 70.3 71.5 73.8 75.4 75.7 74.2 96.9 88.8 84.0 79.9 75.9EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 18.9 13.0 7.4 5.2 3.2HDEU 0.0 3.0 0.9 0.4 0.3 2.6 1.1 0.6 0.5 0.3 0.0 0.0 0.0 0.0 0.0EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 100 100 100 100 100HDPT 0.0 80.1 91.8 94.1 91.2 100 100 99.1 95.5 89.0 100 99.1 96.0 90.4 83.1
Lottery Delay ()
50%
95%
Uncertainty First
T ime First
-5 0 5
-5
Lottery Delay ()
Lottery Delay ()
Uncertainty First
T ime First
Uncertainty First
T ime First
0 5
-5 0 5
5%
5-26
The question, then, is less about whether Certainty-Risk Asymmetry exists, but
rather: (i) for whom does it exist, (ii) under what conditions does it exist, and (iii) how should
debiasing efforts be organized (if at all)? Exhibit WW starts to answer these questions. First,
Certainty-Risk Asymmetry is primarily a function of biased decision making with respect to
time. Succumbing to prospect theory biases alone does relatively little to induce widespread
Certainty-Risk Asymmetry behavior. The behavior exhibits relatively little sensitivity to
either the state 1 adjustment or the delay in receiving the lottery outcome. There is a greater
degree of sensitivity to the probability attached to the lottery payoff, and that sensitivity itself
is a function of the decision process used (see Exhibit XX). This is interesting because, as I
have previously stated, the prospect theory variables play a relatively small role in the
problem, but yet the probabilities (which are transformed by the prospect theory decision-
weighting function) vary significantly with other variables. It appears that the probabilities
become relevant even outside of the decision weighting function, because they control the
amounts subject to hyperbolic discounting.
Exhibit XX: Percent of Population Exhibiting Certainty-Risk Asymmetry with State 1 Adjustment at Zero, Using “Uncertainty First” Decision Sequence, and Varying
Decision Process
5%
50%
95%
5%
50% 95%
5%
50% 95%
0
20
40
60
80
100
Per
cent
age
of P
opul
atio
n E
xhib
iting
C
erta
inty
-Ris
k A
sym
met
ry
Lottery Delay (periods)1 2 3 4 5
EDEU
HDEU
EDPT
HDPT
Probability of State 1 Payoff
5-27
These interrelationships can be confirmed through further analysis of the simulation
data. I performed a logistic regression on the data in order to assess the influence on the
“switching” behavior of each member of the population as a function of their decision
demographics (Exhibit YY). The results of the analysis indicate that, in addition to the
prospect theory risk and loss aversion coefficients, the hyperbolic bias coefficient is also
strongly significant. More importantly, even though the decision-weighting function and
coefficient of risk aversion parameters are not individually significant, they are in
combination with other variables. These results suggest that while resolving hyperbolic bias
appears (as in Exhibit WW) to be most influential in affecting exhibition of Certainty-Risk
Asymmetry, the interaction of that bias with elements of the uncertainty-related parameters
is also influential.
z | | | | Constant -59.458 10.626 -5.595 ****
94.869 13.697 6.926 **** 13.283 1.430 9.290 **** -2.444 0.545 -4.481 **** -7.785 9.925 -0.784 -0.756 10.650 -0.071
-3.745 31.003 -0.121 69.206 22.908 3.021 ***
-29.286 1.992 -14.703 **** 1.002 0.648 1.545 64.985 34.887 1.863 * -17.861 18.059 -0.989 12.502 1.679 7.446 **** 19.753 12.812 1.542 -16.457 21.312 -0.772 -13.938 13.730 -1.015 54.893 22.967 2.390 **
AIC 8,594.4 Cragg-Uhler Pseudo-R2 95.6% Likelihood ratio statistic ( ⋯ 0) 102,347 , 0.0001 No. of observations 100,000
*, **, ***, and **** indicate significance at the 90%, 95%, 99%, and 99.9% levels, respectively.
Exhibit YY: Logistic Regression Results for Certainty-Risk Asymmetry. Case used is p = 0.5, = 3, and = 0 for HDPT, Uncertainty First.
5-28
The second choice task I examine involves Short-Long Asymmetry. In this anomaly,
the discount rate is observed to be higher in the short-term than in the long-term, as shifting
choices into the future makes them more attractive. This behavior is a clear violation of the
stationarity axiom, as shifting the choices across time should not alter subjects’ preferences.
Here, again, the experimental results in the literature are mixed. Albrecht and Weber [1997]
find (some) support for Short-Long Asymmetry. Öncüler and Onay [2009] find support for
the effect, but find the direction of the results mixed. Weber and Chapman [2005] note that
there is disagreement about the effect when subjects resolve risk or time first.
The basic structure of the anomaly is as follows. Subjects choose between two
lotteries in two choices:
Choice #1 50% chance of $100 now or 50% chance of $100+premium at t 50% chance of $0 now 50% chance of $0 at t
Choice #2 50% chance of $100 at Δ or 50% chance of $100+premium at Δ 50% chance of $0 at Δ 50% chance of $0 at Δ
Under normative choice rules, shifting the entire choice into the future by an amount Δ
should have no impact on the selection of lottery within each choice. The choice should,
normatively, remain consistent across time. I performed simulations varying the state 1
probability (5%, 50%, and 95%), the premium attached to the state 1 payoff for deferral of
receipt (-10, 0, 10), and the period of time the lotteries are moved into the future (Δ) and the
time between lotteries within each choice (t):
Short time shifts
Short vs long
Moderate time shifts
Long vs short
Long time shifts
Δ 1 1 5 10 10 t 1 10 5 1 10
5-29
Exhibit ZZ: Percent of Population Exhibiting Short-Long Asymmetry The State 1 payoff adjustment reflects the amount added to the state 1 payoffs in the deferred lotteries. Simulations were performed based on whether the decision makers were assumed to resolve uncertainty first or time first as described in Section 5.3.2. The abbreviations in the second column refer to the decision models employed: EDEU (exponential discount, expected utility), HDEU (hyperbolic discounting, expected utility), EDPT (exponential discounting, prospect theory), and HDPT (hyperbolic discounting, prospect theory).
Probability of State 1
State 1 Payoff Adjustment
11
110
55
101
1010
11
110
55
101
1010
11
110
55
101
1010
EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 25.2 0.0 3.1 76.0 1.8EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 15.6 0.0 0.3 83.8 0.1EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 10.8 0.0 16.9 15.0 10.9EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 10.8 0.0 16.9 15.0 10.9
Probability of State 1
State 1 Payoff Adjustment
11
110
55
101
1010
11
110
55
101
1010
11
110
55
101
1010
EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 25.2 0.0 3.1 76.0 1.8EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 15.6 0.0 0.3 83.8 0.1EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 10.8 0.0 16.9 15.0 10.9EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 10.8 0.0 16.9 15.0 10.9
Probability of State 1
State 1 Payoff Adjustment
11
110
55
101
1010
11
110
55
101
1010
11
110
55
101
1010
EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 25.2 0.0 3.1 76.0 1.8EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 15.6 0.0 0.3 83.8 0.1EDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDEU 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 10.8 0.0 16.9 15.0 10.9EDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0HDPT 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 10.8 0.0 16.9 15.0 10.9
5%
-10 0 10
Uncertainty First
T ime First
Uncertainty First
T ime First
50%
-10 0 10
Uncertainty First
T ime First
Shift into future ()T ime between lotteries (t )
Shift into future ()T ime between lotteries (t )
Shift into future ()T ime between lotteries (t )
95%
-10 0 10
5-30
As before, the output captured the percentage of the simulated population switching
their preferences from Choice #1 to Choice #2. Exhibit ZZ illustrates the percentages of the
population showing preference reversals under each tested condition. The vast majority of
variations showed no evidence of Short-Long Asymmetry. However, some elements of the
population do exhibit the effect. I consider this particular anomaly to investigate the question
as to whether the order of resolution of the decisions affects the likelihood of observing the
bias.
Keren and Roelofsma [1995], for example, have asserted that uncertainty and (time)
delay share a common dimension, but that uncertainty is the “more fundamental” of the two.
That is, that delay makes an outcome more uncertain, either due to the possibility that the
subject may not be alive (or otherwise available) to receive a deferred payoff, or that the
value of the deferred payoff may be reduced once it is actually received due to some then-
unforeseen event. Similarly, Weber and Chapman [2005] consider whether the presence of
one factor (uncertainty or time) can affect response to the other factor. The authors’
experiments supported answering both questions in the affirmative.
The design of the simulation developed here allows these claims to be investigated
directly. I simulated decision rules that implied resolution of uncertainty first and others than
implied resolution of time first. From this construct, I can evaluate the influence of each part
of the decision demographics on the percentage of the population exhibiting Short-Long
Asymmetry given that population members resolve uncertainty or time first. As the
classification tree in Exhibit AAA illustrates, however, use of the standard logistic regression
model is likely to fall prey to quasi-complete separation, as certain predictors work “too
well” in explaining the dependent variable. This is a common problem with logistic models
[Albert and Anderson, 1984]; here, it is the hyperbolic bias parameter that has an outsized
influence on the result.
Resolving the separation issue typically involves employing some form of bias
reduction in the maximum likelihood estimation process, since separation causes improper
inflation of the maximum likelihood estimates. An early bias reduction method was the
5-31
penalized maximum likelihood method proposed by Firth [1993], which employs what is
essentially a Jeffreys prior to remove first-order bias from the maximum likelihood estimate.
But the use of Jeffreys priors can sometimes cause computational issues of its own. Instead,
I use the Bayesian generalized linear model approach of Gelman et al. [2008], which uses a
Cauchy prior to provide a more stable estimation process in fitting the logistic regression.
Exhibit AAA: Classification Tree Analysis Illustrating Quasi-Complete Separation. Case used is p = 0.05, = 10, t = 1, and Premium = 100 for HDPT, Uncertainty First.
The traditional “Uncertainty First” model is presented in Exhibit BBB. When the
population of decision makers is assumed to resolve uncertainty first, all of the prospect
theory and hyperbolic bias parameters are strongly significant in explaining the occurrence
of Short-Long Asymmetry. That is, the population’s biases influence their judgments, and
83.8% of the population therefore exhibits Short-Long Asymmetry. In contrast, when the
population of decision makers resolves time first (in Exhibit CCC), only the time-related
parameter ( ) is significant ( , the prospect theory risk aversion parameter for losses (or
“risk-seekingness”), is weakly significant, but small). In the “Time First” case, the
population’s uncertainty-related biases do not influence their decision making, and only
5-32
15.0% of the population exhibits Short-Long Asymmetry, which appears to occur primarily
because of time, or hyperbolic bias.64
Std. Error z | | | | Constant -29.626 1.786 -16.753 ****
36.185 2.564 14.111 **** 8.315 0.787 10.572 **** 2.092 0.276 4.580 **** -17.977 1.547 -11.625 **** 22.427 1.645 13.632 ****
-2.543 17.579 -0.145 92.941 5.134 18.104 ****
-12.987 1.067 -12.166 **** -0.860 0.351 -2.448 ** -17.185 19.578 -0.878 8.843 10.062 0.879 -15.149 1.016 -14.913 **** 26.177 2.301 11.374 **** 40.073 4.393 9.121 **** -30.938 2.454 -12.605 **** -57.118 4.619 -12.367 ****
Null deviance 88,604 on 99,999 degrees of freedom Residual deviance 25,509 on 99,983 degrees of freedom AIC 25,543 Cragg-Uhler Pseudo R2 79.6%
*, **, ***, and **** indicate significance at the 90%, 95%, 99%, and 99.9% levels, respectively.
Exhibit BBB: Bayesian Logistic Regression Results for Short-Long Asymmetry. Case used is p = 0.05, = 10, t = 1, and Premium = 100 for HDPT, Uncertainty First.
Using the simulation approach provides a useful addition to the results of Keren and
Roelofsma [1995] and Weber and Chapman [2005]. The results presented here suggest that
directing the decision makers to time first may “fix” biases related to uncertainty in the
Short-Long Asymmetry tasks. Obviously, because the simulation is not a human subject,
this is not a result of the decision makers’ biases being corrected, but rather that changing
64 This can also be confirmed by “turning off” hyperbolic bias in the decision demographics (in effect, making all decision makers exponential discounters) and re-running the simulation. In this case, where the population exhibited all of the uncertainty-related parameters, but not the time-related parameter, no cases of Short-Long Asymmetry occurred.
5-33
the structure of the problem to emphasize resolving time-related decision making first can
“immunize” decision makers in this task against the consequences of uncertainty-related
biases because most of the effect is caused by time-related factors.
Std. Error z | | | | Constant 7.215 1.406 5.133 ****
0.045 2.293 0.020 -1.521 0.902 -1.687 * -0.232 0.348 -0.666 -0.180 1.293 -0.139 0.157 1.205 0.131
-24.098 21.426 -1.125 -66.348 4.970 -13.349 ****
1.382 1.203 1.149 0.346 0.431 0.804 30.058 24.105 1.247 1.598 12.549 0.127 -0.721 1.759 -0.410 -1.768 2.109 -0.838 7.623 6.286 1.213 -0.522 2.184 -0.239 0.402 2.351 0.171
Null deviance 84,680 on 99,999 degrees of freedom Residual deviance 17,126 on 99,983 degrees of freedom AIC 17,160 Cragg-Uhler Pseudo R2 86.0%
*, **, ***, and **** indicate significance at the 90%, 95%, 99%, and 99.9% levels, respectively.
Exhibit CCC: Bayesian Logistic Regression Results for Short-Long Asymmetry. Case used is p = 0.05, = 10, t = 1, and Premium = 100 for HDPT, Time First.
Using simulation to investigate novel and complex decision problems in virtual
experiments is useful not only because of the lower cost of conducting the experiments, but
because a greater variety of treatments and manipulations can be explored, and cross effects
more easily examined. Further, while simulating experiments is not a substitute for the use
of human subjects, it is a useful testing ground for designing such experiments and
generating and evaluating hypotheses. Lastly, the parallel use of simulated experiments
provides a means of investigating the degree to which unintended experimental factors can
influence the results. The simulated populations are not sensitive to the actions of the
experimenter, or the design of the instrument, or the instructions given. Thus, it provides an
5-34
objective means of testing whether the results observed in an experiment are substantive or
superficial.
5.5 Applications and Implications
Relative to many other areas in the social sciences, behavioral decision making has
remaining rather underexplored by the computational social science approach. In this
chapter, I developed a computational framework to simulate experiments in risky
intertemporal choice on a population of decision makers whose demographics have been
estimated from a large body of prior experimental work on human subjects. Two primary
classes of applications were identified.
First, incorporating probabilistic information about the population of decision
makers helps to provide context to experimental results about the practical impact of biases
and the likelihood of observing certain patterns of behavior in large populations. This
information can be used to moderate attention on areas that are “significant, but
insignificant,” or areas where an effect is determined to be present, but lacking in a material
impact for a particular problem or problem environment. Such attention can be redirected to
areas where effects may be weak, but widely prevalent in a population or particularly
material for a problem.
Second, simulation can be used to perform virtual experiments on novel or complex
environments. This can be done both as a prelude and guide to experimentation with human
subjects, as well as a substitute in areas where experimentation with human subjects may be
impractical or impossible. The use of computational methods to facilitate the design of
experiments is a novel application, and may allow a more efficient and structured evaluation
of hypotheses when interactions are present, as well as to provide an experimental default
case where effects can be tested in a virtual “clean room,” stripped of the influence of human
experimenters and experimental instruments.
As a result, I believe the computational approach can provide for a symbiotic
relationship with the experiments on human subjects common to behavioral decision
5-35
making, as each approach informs and provides a check on the other. Behavioral decision
making has tended, necessarily perhaps, to focus on the decision making of individuals. A
different level of insight can be obtained from using computational methods to aggregate the
behavior of these diverse individuals and thus provide an atomistic means of examining the
behavior of large, diverse populations.
6-1
6 Presentation Compression: Investment Metrics & Heuristics
6.1 Introduction
Finance commonly considers decision problems that involve both risk and time.
These problems usually involve choice between “projects,” “investments,” or “policies.”65
Corporate managers must choose between undertaking different projects that have different
lifespans and different risks. Investors must choose among securities that have different
payoff profiles. Policymakers must consider alternative policies that have varying and
uncertain cost-benefit streams. More abstractly, these choice problems may be considered
intertemporal lotteries (combinations of payoffs and probabilities), which I shall call
“prospects” after Kahneman and Tversky [1979].
Finance practitioners are commonly taught to calculate and use certain metrics to
facilitate decision making about prospects [Shapiro, 1988; Brealey and Myers, 2003]. These
metrics include Net Present Value (“NPV”) and Internal Rate of Return (“IRR”) for business
applications, Equivalent Uniform Annual Cost66 (“EUAC”) for engineering applications,
and, for the power industry, the Levelized Cost of Energy (“LCOE”). Certain heuristic rules
are associated with these metrics. For example, the NPV rule states that Project A can be
judged preferred to Project B if the NPV of Project A is greater than the NPV of Project B.
These metrics, therefore, “compress” information about a prospect in that they collapse risk
and time into single figures (“expected present values” or “discounted expected values”) that
can be compared directly.
This compression process is intended to simplify decision making (it is “convenient”
[EIA, 2016]), even though classical finance has not generally discussed the use of metrics in
65 Or, implicitly, choice between the maintaining the status quo and another option. 66 There are a variety of iterations of this concept, which is commonly used in engineering and dates to Fish [1923], including Equivalent Uniform Annual Cost, Equivalent Uniform Annual Benefit, Equivalent Uniform Annual Worth, and Equivalent Uniform Annual Cash Flow, among others.
6-2
such decision-making terms. Normatively, however, the compressed presentation form of a
prospect is extensionally equivalent to the full presentation form of a prospect (see
Hirshleifer [1958], for example, regarding NPV). The information “lost” to compression has
no normative relevance to decision making. Heuristic decision makers, however, may not
see the two forms as extensionally equivalent [Bourgeois-Gironde and Giraud, 2009].
The complexity of transforming prospect presentation without altering preference
has long been demonstrated in the decision-making literature [Rugg, 1941; Schelling, 1981;
Tversky and Kahneman, 1981]. The practical challenges that compression imposes have led
some to assert that, in practice, decision makers should be given less information to make
better decisions, contradicting traditional economic theory [Iyengar and Lepper, 2000;
Schwartz, 2004]. However, this difficulty in maintaining consistent preferences across
presentation forms has done little to diminish the demand for such tools, which is ubiquitous:
“In spite of these manifest inadequacies in the available information, the policy-maker must frequently make some sort of decision without delay. The temptation for her/his advisors is to provide her/him with a single number, perhaps even embellished with precise confidence limits of the classic statistical form. When such numbers are brought into the public arena, debates may combine the ferocity of sectarian politics with the hyper-sophistication of scholastic disputations. The scientific inputs then have the paradoxical property of promising objectivity and certainty by their form, but producing only greater contention by their substance.” [Funtowicz and Ravetz, 1987: page 62]
This clash between presentation form and substance is at the core of this essay.
To the extent that presentation form influences preference over prospects for
heuristic decision makers, it becomes important to know how, and to what extent, such
influence extends. In other words, compression and the metrics created for it can be both
helpful and harmful devices depending on the characteristics of the decision maker and the
nature of the problem in question. For what types of problems do the use of metrics by
cognitively-constrained decision makers improve or worsen decisions as measured by
agreement between the descriptive choice rule employed and their choice made by a
normative choice rule? In this chapter, I apply the computational framework developed in
Chapter 5 to a practical application. I evaluate presentation forms for prospects designed to
be representative of many power-related investments (among other applications) and
6-3
identify not only the types of problems for which compression may be either problematic or
beneficial, but also the types of decision makers for whom compression may be problematic
or beneficial.
In Section 6.2, I introduce LCOE as a metric and discuss its use. Section 6.3 develops
a model of the decision problem faced by heuristic decision makers, as well as the measures
used to evaluate decision performance. Section 6.4 reviews the specific analyses performed
and communicates their findings. Finally, Section 6.5 concludes by discussing the
implications of the findings and their practical applications.
6.2 The Levelized Cost of Energy in Theory and Practice
6.2.1 The Levelized Cost of Energy
Although several metrics are commonly calculated for decision making involving
prospects, I focus on one, the Levelized Cost of Energy, that is of particular relevance to
energy policy analysis. In its most basic form, the LCOE is a means of combining both fixed
and variable costs into a unitized cost. As such, it has three components: (i) recovery of fixed
cash flows, (ii) pass-through of variable cash flows, and (iii) unitization. Although LCOE
has traditionally been utilized to discuss competing energy investments or policies, its
calculation is applicable to any series of cash flows and is closely related to EUAC and to
NPV [Brown, 1994].
Assume a series of n costs or capital requirements , , … , . Given a discount
rate r, the NPV of those costs is ∑ 1 . The per-period “levelized” fixed payment
Γ is determined such that:
Γ1 1
[1]
The left side of [1] has a closed form:
6-4
Γ1
Γ Γ 1
Γ 1 1
[2]
Equating the levelized fixed payment to the costs:
Γ ∙1 1
1 [3]
Rearranging, the levelized fixed payment amount [4] can be determined.
Γ
1
TotalCapitalEmployed
∙1 1FixedChargeFactor
[4]
To arrive at the traditional operational definition of LCOE, the summation on the right side
of [4] can be referred to as the Total Capital Employed, which is multiplied by the Fixed
Charge Factor (which is essentially an annuity factor). The product of these figures, Γ,
reflects the per-period recovery of fixed capital costs. To this figure must be added per-
period non-capital fixed costs and variable costs, and the entire figure expressed on a unitized
basis (e.g., divided by per-period generation in megawatt-hours). The LCOE, which I shall
refer to as L, is then given by [5], which is expressed in dollars per MWh or cents per kWh,
or a similar form.
Γ FixedCostsGeneration
VariableCosts [5]
6-5
Although referred to as the Levelized Cost of Energy (or sometimes Electricity),
Reichelstein and Rohlfing-Bastian [2015] note that it is equivalent to long-run marginal
product cost, or what is called “life-cycle full cost” in cost accounting, and has applications
in manufacturing and service industries. Similarly, Grabowski and Vernon [1990] show that
levelized cost is also used in production of pharmaceutical products. However, it is by far
most commonly used in energy policy analysis worldwide for evaluating responses to
environmental regulation [MacDonald et al., 2016], capacity planning [ATSE, 2011; Nitsch
et al., 2012; Alberici et al., 2014; Danish Energy Agency, 2015; IEA, 2015; JANRE, 2015],
and project investment [Lazard, 2016].
In support of this widespread use are staunch defenders such as Sovacool [2008],
who has noted that within LCOE all decision elements are “reflected and quantified as best
as they can be” [ibid., p. 249]. Notwithstanding his advocacy, a cottage industry has emerged
in generating “fixes” for LCOE. These modifications are intended to address what are
asserted to be shortcomings in the use of LCOE that lead to less-than-optimal choices. They
have included its failure to include residual values and financing effects [Velosa and Aboudi,
2016], its handling of outages and mid-life capital additions and replacements [Kakade,
1989], its failure to recognize differential rates for investing and borrowing [Manzhos,
2013], and its failure to include system integration costs (for intermittent renewable energy
sources) [Ueckerdt et al., 2013; Hirth, Ueckerdt, and Edenhofer, 2015]. In virtually each
case, the solution proposed was the inclusion of additional parameters in the formula. A
single value for each option was still produced, but it was computed with more input data—
that is, significant increases in compression were involved.
Taking a slight step backward, the U.S. Department of Energy’s Energy Information
Administration’s proposed “fix” was the creation of a separate measure called Levelized
Avoided Cost of Energy (“LACE”) [EIA, 2016], that was intended to be used in conjunction
with LCOE in order to better reflect “economic value.” This could be considered a degree
of “decompression.” It is telling, however, that the EIA acknowledged that their long-term
projections of capacity additions “use neither LACE nor LCOE concepts” [ibid.].
6-6
6.2.2 Levelized Cost vs. Levelized Cash Flow
What many of the above “fixes” demonstrate is that LCOE has exclusively been a
measure of costs, or capital outflow. This characteristic is by design, as it emerged from the
regulated utility world in which the costs of an investment approved by regulators were its
revenues. But critiques of LCOE on such matters as differential borrowing and lending rates
or exclusion of residual values target this cost-only foundation.
Perhaps the most forceful critique of this nature has been that of Joskow [2010;
2011]. Joskow [2010] noted that LCOE had emerged from the regulated utility world and
was not generally suitable for the merchant power generation world. Joskow [2011] then
demonstrated that LCOE tends to misrepresent the value of intermittent (renewable) sources
of generation relative to dispatchable sources of generation because it neglects the value of
the electricity generated by each source (recall that LCOE is a cost measure). Because
electricity in deregulated markets is not a homogeneous commodity (it is delivered at
different times in different locations), its value varies widely (by, I would note, time and
uncertainty of delivery). Not only does LCOE then create an improper comparison between
intermittent and dispatchable resources, but also between different intermittent resources
(e.g., wind and solar) themselves.
Rather than “fix” LCOE as so many others have proposed, Joskow’s [2011] advice
was to abandon it in favor of looking at actual market values for the electricity by each
source. Restated in the terms of this essay: Joskow argued that the use of LCOE as a metric
and the heuristics that employ it could lead to preference reversals because the compression
process omitted information that did, in fact, have normative relevance. I don’t take as
extreme a stance. The idea of “levelizing” cash flows (not just costs) appears in many areas
and can have useful applications. Annuities and mortgages, for example, are typically
levelized positive and negative cash flows (at least to the recipient), respectively, that can be
attractive means of intertemporal income and consumption smoothing. Even in electric
power markets, expected capacity revenue is often modeled as a “residual fixed cost,” which
is the levelized revenue requirement of a generator unmet by variable revenue sources and
6-7
thus the result of the classic two-part tariff problem in economics [Hopkinson, 1892; Oren,
Smith, and Wilson; 1985]. In most deregulated electric power markets, for example, capacity
prices are determined (in general terms) as the intersection of the Cost of New Entry
(“CONE”) less earnings from energy market operations (energy and ancillary services
revenue) (“Net CONE”) and the demand curve. Net CONE is expressed as a levelized figure
inclusive of both positive and negative cash flows (depending on the perspective assumed).
For the remainder of this essay, I use LCOE in this general sense to refer to the
“levelized” or compressed presentation of cash flows (both positive and negative). For
convenience of exposition, I ignore the characteristics that make it specific to electric power
(i.e., the per-period fixed costs, variable costs, and unitization), as they are not material to
the discussion of compression. Accordingly, my references to LCOE in the remainder of this
essay are to Γ as defined in [4] above—the levelized portion of (capital) cash flows. It is this
portion of the LCOE that deals directly with compression.
6.3 The Decision Problem
Consider two prospects, A and B. Each decision maker follows either a normative
(N) or descriptive (D) choice rule that indicates their preference over the prospects as
presented.67 For example, ≻ indicates that a decision maker using a normative choice
rule would prefer A to B. A rational decision maker is assumed to act in accordance with N;
a heuristic decision maker in accordance with D. A prospect may be presented in one of two
forms: a full form F or a compressed form C. The full form of a presentation is as a lottery
or decision tree or, in finance terms, a cash flow profile (i.e., it conveys detailed information
about uncertainty and time). The compressed form of a presentation is a metric—here, as
LCOE. To continue the example, ≻ indicates that a decision maker using a
normative choice rule and presented with A and B in full form would prefer A to B.
67 I ignore indifference throughout this description in the interest of clarity.
6-8
The usefulness of a presentation and its associated heuristic (which I refer to as a
“presentation-heuristic pair”68) is measured by the agreement their use induces with a
normative choice rule under the full (or compressed) presentation. The full and compressed
presentations are treated equivalently by the normative rule and therefore I do not distinguish
between them further. Exhibit DDD illustrates the various paths of this decision-problem
construction. The outcome space can be characterized by six basic descriptions. The first
two refer to the sorts of heuristics and biases commonly discussed in the behavioral decision-
making literature. I describe a choice rule D that results in the same choice as a choice rule
N as well-calibrated, in the sense that the heuristic choice rule leads to the normative
selection. In contrast, a choice rule D that results in a choice other than the normative choice
rule N is considered biased, in the sense that its use of simplifying heuristics potentially
subjects the decision maker to biases in judgment.
Exhibit DDD: Diagram of Choice Rule and Presentation Relations
The remaining four descriptions noted in Exhibit DDD refer to the actions of
heuristic decision makers presented with both full and compressed presentations. A heuristic
68 To reiterate, for the sake of clarity, the presentation is, for example, NPV, and the heuristic in that case is “pick the project with the greatest NPV.”
≻ ≺
≻ ≺
≻
BiasedWell-calibrated
Simplifying
Counter-productive
IrrelevantUseful
Areas of the problem space where the use of
C should be encouragedAreas of the problem space where
use of C should be carefully monitored or discouraged
6-9
decision maker presented with a decision problem in compressed form that makes the same
decision as a heuristic decision maker using the full form (i.e., a well-calibrated decision
maker) and a normative decision maker indicates that the compressed form is simplifying.
That is, the correct decision is being made with less information and therefore less cognitive
effort. If, in contrast, the compressed form leads such a well-calibrated decision maker to
indicate a preference in opposition to the descriptive and normative rule, the compressed
form is counter-productive. That is, the compression of information causes a bias in
decision making that was otherwise not present with the full form of the problem.
Discordance between the full and compressed forms, however, may not necessarily be a
negative trait. If a heuristic decision maker is led to make a normatively-incorrect decision
when presented with the full form of a decision problem that could be “corrected” by use of
the compressed form, then the compressed form is useful.69 In this case, it is the reduction
in (rather than the presence of) information presented that triggers a biased decision by the
heuristic decision maker. If, however, such a decision maker continues to make the
normatively-incorrect choice using the compressed form, then the compressed form is
irrelevant. In this case, presentation compression is “transparent” in that it has no bearing
on the original biased decision made by the decision maker.
The simplifying and useful areas of the outcome space, then, demarcate areas where
certain aspects of the decision problem and/or certain characteristics of the decision maker
lead to favorable outcomes that should be encouraged. In short: areas where compression is
a beneficial force. In contrast, the counter-productive and irrelevant areas of the decision
space indicate areas to be avoided. In short: areas where compression is an unhelpful force.
The purpose of this analysis, then, is to identify the areas of the problem space and the
segments of the population of decision makers where and for which the use of metrics and
heuristics are simplifying or useful and where they are counter-productive or irrelevant.
69 For example, if the full presentation form of a problem involved decision problems involving heavy losses or small probabilities, etc., one may expect a heuristic decision maker to succumb to the biases commonly-acknowledged in such circumstances. Such a decision maker’s biases may not be triggered if the compressed form obscures those characteristics through levelizing.
6-10
Any practical advice on the proper use of heuristics must be given cognizant of the
boundaries imposed by the “imperfectness” of the compression they employ. In finance, by
way of example, advocacy for the use of the internal rate of return (“IRR”) as a metric is
always tempered by the caveat that multiple sign changes in cash flows between periods
leads to unstable results (because the solution of the polynomial equation then has multiple
roots). So, too, should use of LCOE—or indeed any metric—be guided by the contours of
the space it traverses.
The response to this space can be estimated by using well-established theories of
heuristic decision making under risky intertemporal choice. A considerable body of evidence
(cited extensively in Chapter 5) supports the prospect theory [Kahneman and Tversky, 1979]
and hyperbolic discounting [Thaler, 1981; Ainslie, 1991] models in such circumstances. If
prospect theory and hyperbolic discounting models are accurate representations of human
behavior with regard to uncertainty and time, then one means by which such insight could
be obtained is the computational exploration of that space. This sort of computational
approach is shared with the “Candle-Lighting Analysis” of Kimbrough and colleagues
[Kimbrough et al., 1992; Kimbrough, Oliver, and Pritchett, 1993], as well as Miller’s [1998]
“Active Nonlinear Tests,” although their approaches generally employ heuristic
optimization methods.
I argue that such exploration should occur prior to experimental exploration of that
space because of the relative efficiency of computational methods in exploring vast search
spaces. Additionally, because knowledge of the problem space can facilitate the generation
of structural hypotheses that can then be supplemented and shaded with the substructural
detail more appropriate for exploration with human subjects. Such a computational approach
has two distinct components: (i) a model of the decision makers and (ii) a model of the
decision problems.
Chapter 5 of this dissertation develops a computational model of heuristic decision
makers and develops a simulation framework for evaluating the decision-making behavior
of a population confronted with various decision problems. I use that model and approach
6-11
here and proceed with a simulated population of decision makers, for which each member
of the population is represented by the 7-tuple , , , , , , , which comprises,
respectively, the risk aversion coefficients for gains and losses, the loss aversion coefficient,
the decision weight parameters for gains and losses, the Arrow-Pratt coefficient of risk
aversion, and the hyperbolic bias parameter particular to that decision maker. Each of these
“virtual” decision makers are then exposed to a set of decision problems. The remainder of
this section details the process by which the decision problems are evaluated. Section 6.4
then develops a model of decision problems and conducts the search over the space of
problems.
6.3.1 Decision Problem Structure
Consider a population of decision problems S.70 In order to set out the evaluation
structure of the compression decision problem, I define a decision function
≻ ; , ∈ that takes a choice rule71 ∈ , (as described in Sections 5.3.1
and 5.3.2) and a prospect presentation form ∈ , (described below) and indicates a
preference for a pair of prospects in S, the set of all prospects (described in more detail in
Section 6.4).72 For example, ≻ ; , indicates that a normative choice rule acting
on the full-form presentation of prospects A and B results in the choice A. Having established
this notation, the following steps are performed:
70 In abstract terms, these decision problems are intertemporal lotteries. In the application discussed in this chapter, they are specific power project investments. 71 By “choice rule,” I mean exponential discounting-expected utility for normative (N) decision makers and hyperbolic discounting-prospect theory for heuristic (D) decision makers. 72 The notion of a “decision function” here may be considered a computational substitute for a human subject in a more traditional experimental setting. They are, in a sense, automata, designed to represent the choices of decision makers acting according with certain rules and presented with certain choices.
6-12
1. Randomly select two prospects from S, 2. Randomly select a representative sample of decision makers from the
simulated population of decision makers (Chapter 5), 3. Apply ≻ ; to the prospects, 4. Apply ≻ ; to the prospects, 5. Apply ≻ ; to the prospects, 6. Measure agreement between the choices indicated by the functions in steps
3-4.
With each of the steps evaluated, an outcome classification can be made in keeping with
Exhibit DDD. This outcome classification (Exhibit EEE) draws a correspondence between
the agreement among presentation-heuristic pairings and their usefulness. The evaluation of
a particular presentation-heuristic pair (as Simplifying, Useful, etc.) is a function of the
consistency pattern of its decision functions.
Two presentation forms are considered: Full and Compressed. The Full
presentation form is an intertemporal lottery of multiple periods, with the per-period time
discount rate equal to r. The Full intertemporal lottery form (for three periods) is illustrated
in Exhibit FFF. The Compressed presentation form (also illustrated in Exhibit FFF) is the
levelized equivalent of the lottery. The levelized amount Γ is defined as in [4] such that the
NPV of the lottery (at r) is equal to the NPV of the levelized payment stream (at r).73 I denote
probabilities as p and the payoff amounts as Π , where s represents the state of the world
and t indexes time. Having defined the structure of the problem and the means of evaluation,
I now turn to the analysis of the prospects.
73 That is, ∑ Π 1 Π Γ∑ in this notation.
6-13
≻ ; , ≻ ; , ≻ ; , Simplifying
≻ ; , ≻ ; , ≻ ; , Counterproductive
≻ ; , ≻ ; , ≻ ; , Irrelevant
≻ ; , ≻ ; , ≻ ; , Useful
Exhibit EEE: Presentation-Heuristic Pairing Agreement Classification
Exhibit FFF: Full and Compressed Presentation Forms
=
=
Simplifying
=
≠
Counterproductive
≠
=
Irrelevant
≠
≠
Useful
1
Π
Π1
Π
Π1
Π
Π
Γ Γ Γ
6-14
6.4 Analysis and Findings
6.4.1 Power Investments as Intertemporal Lotteries
The objective of this study is to investigate decision making involving power projects
in a plausibly realistic, yet tractable setting. The task, then, becomes how to represent power
projects as intertemporal lotteries. Such projects can be characterized generally as consisting
of three phases: (i) a construction phase, (ii) an operating phase, and (iii) a decommissioning
or salvage phase. The total duration of the three phases varies according to project type, and
can involve considerable uncertainty related to technological, economic, and regulatory
factors [Rode, Fischbeck, and Páez, 2017]. Furthermore, there is uncertainty involved in
each phase as to the costs, as well as the operating and financial performance, of the projects.
Most of the common financial metrics are based on the cash available for finance (by
either debt or equity) (“CAFF”).74 Alternatively, in certain settings, such as cost-of-service
regulation, a cost-only measure is more appropriate. In either case, the metrics in question
(such as NPV, IRR, or LCOE) make use of a “bottom line” cash flow number per period
(see Exhibit GGG). Accordingly, each period must comprise a separate lottery reflecting the
cash flow or cost experiences particular to that period. The resolution of that uncertainty in
each period provides the cash flow or cost employed by the financial metric for decision-
making purposes.
74 Cash available for finance is defined as the total cash available for distribution to a firm’s capital structure. It is equal to the free cash flow in an Adjusted Present Value model [Brealey and Myers, 2003] and the free cash flow to equity in a 100% equity-financed model. The CAFF is what is available either for the payment of debt service or distribution to equity holders.
6-15
Exhibit GGG: Typical Power Project Cash Flow Derivation
These lottery representations include many of the characteristics known to influence
decision making (e.g., low probability-high consequence outcomes, periods framed as gains
and losses, payoffs stretched across time, and risk). Accordingly, in the context of the
framework described in Section 6.3.1, each period involves a lottery between a base outcome
(the top branch of the lottery) and an alternate outcome (the bottom branch of the lottery),
each of which occurs with a particular probability. I seek to examine the particular
characteristics of such constructions that pose potential problems or opportunities for
heuristic decision makers.
6.4.2 Basic Prospect Structures and Design Elements
To populate S, the set of prospects, I create power project intertemporal lotteries
defined along nine dimensions and then create a “population” of power projects by exploring
all of the combinations possible along these nine dimensions. The nine dimensions are
Revenues Costs+ Energy Revenue + Fixed Operating & Maintenance Costs ("FOM")+ Capacity Revenue + Variable Operating & Maintenance Costs ("VOM")
+ Fuel Costs= Total Revenues = Total Costs
– Initial Capital Cost & Future Capital Expenditures= Total Expenses ← Expense-only presentation
Total Revenues – Total Costs= Gross Margin – Depreciation= Taxable Income – Income Tax= Net Income+ Depreciation+ Tax Incentives – Initial Capital Cost & Future Capital Expenditures= Cash Available for Finance ("CAFF")
6-16
defined as below, and reflected in a series of binary lotteries over time. All of the “payoff”
values are converted to dollars per kilowatt-year ($/kW-year, or $/kW for short, when the
time period is unambiguous) for comparative purposes and presented in constant dollars.
Exhibit HHH: Four Example Cash Flow Profiles
-$10.0
-$5.0
$0.0
$5.0
$10.0
1 6 11 16 21 26 31 36 41 46 51 56Ln(
$/kW
)(c
onst
ant d
olla
rs)
Nuclear
Base
Alternate
-$10.0
-$5.0
$0.0
$5.0
$10.0
1 6 11 16 21 26 31 36 41 46 51 56Ln(
$/kW
)(c
onst
ant d
olla
rs)
Wind
Base
Alternate
-$10.0
-$5.0
$0.0
$5.0
$10.0
1 6 11 16 21 26 31 36 41 46 51 56Ln(
$/kW
)(c
onst
ant d
olla
rs)
Solar
Base
Alternate
-$10.0
-$5.0
$0.0
$5.0
$10.0
1 6 11 16 21 26 31 36 41 46 51 56Ln(
$/kW
)(c
onst
ant d
olla
rs)
Natural Gas Combined Cycle ("NGCC")
Base
Alternate
6-17
First, I model four general types of project: nuclear, wind, photovoltaic solar, and
natural gas combined cycle (“NGCC”). These four main project types reflect a diverse array
of cash flow profiles due to their operational, economic, and incentive-related characteristics
(illustrated in Exhibit HHH).75 The alternate cases within each lottery are created by
worsening the estimated capital cost and decommissioning costs by 50% and the operating
costs by 10%. Especially in the case of capital costs, these estimates may prove conservative,
as there is a great deal of evidence of significant cost overruns among large infrastructure
projects [Flyvbjerg, Bruzelius, and Rothengatter, 2003]. Each of these project types has a set
of inputs unique to their technology, including capital cost, capacity factor (percent of the
year in operation), and operating costs, as described in Exhibit III.76
Exhibit III: Four Power Plant Classes and Class-Specific Information
Second, I model six different project-lifespan levels based on a sixty-year lifespan
limit, with the model variables reflecting 50%, 60%, 70%, 80%, 90%, or 100% of that limit.
I emphasize that not all technologies are equally likely to realize different parts on this range
of lifespans. For example, while a nuclear power plant may be commonly expected to remain
75 The figures in Exhibit HHH are presented in a log scale, with negative values calculated as ln . 76 The items Capital Cost (the initial construction cost of the plant, expressed as an overnight cost), Fixed Operating & Maintenance (“FOM”) (such as labor and routine maintenance that does not vary with the level of production), Variable Operation & Maintenance (“VOM”) (such as consumables and usage-dependent maintenance), and Fuel are calculated based on figures from the U.S. Department of Energy’s Energy Information Administration’s Assumptions to the AEO and Annual Energy Outlook 2017 document, with the exception of the fuel costs for nuclear plants, which is taken from figures provided by the Nuclear Energy Institute. Estimated nuclear decommissioning costs are based on the actual costs of the nuclear plants that have completed the decommissioning process. All dollar figures provided in $/MWh are converted to $/kW-year based on assumed capacity factors, as discussed below. The Capacity Availability Rating reflects the percentage of plant capacity that is permitted to receive capacity revenue, and is based on the forced outage ratings commonly accepted in many wholesale capacity markets (I assume availability = 1 – eFORd).
Capital Cost (2016$)
FOM(2016$/kW-year)
VOM (2016$/MWh)
Fuel (2016$/MWh)
Capacity Availability
Rating
Decommissioning Expense
(2016$/kW)Nuclear $5,880 $99.65 $2.29 $6.76 95% -$1,200Wind $1,686 $46.71 $0.00 $0.00 5% $0Solar $2,277 $21.66 $0.00 $0.00 0% $0NGCC $1,094 $9.94 $1.99 $30.93 90% $0
6-18
in operation for sixty years, such a period of commercial operation would be rather
unprecedented for a solar project.
Third, I model capacity-factor levels uniquely for each technology based on ranges
commonly experienced. The resulting seven levels for each technology are provided in
Exhibit JJJ below. Capacity factor is important not just as a measure of energy generation,
but also as an indicator of the energy prices realized by generators (as prices vary by time of
day, day of week, and month of year). Only a generator that operates 100% of the year could
be expected to realize the average annual energy price. In contrast, although a solar power
plant may only operate 20% of the year, those hours tend to be the highest-priced peak hours,
and therefore they would be expected to receive revenue more than the average annual
energy price (or the average price, conditional on time of delivery). For the sake of simplicity
in this analysis, I ignore transmission constraints and assume each generator operates under
daily operating discretion as a merchant. I constructed a price-duration curve based on 10-
years of historical energy price data for four regions across the country.77 From that data, a
conditional average energy price can be constructed based on the assumption that each
generator’s operation is scheduled to occur during the highest-price hours.78
Exhibit JJJ: Capacity-Factor Levels for Each Type of Power Plant Included
77 The four regions were selected to reflect a broad cross-section of the country’s electricity markets are comprise prices at the Southern (south), Palo Verde (western), PJM (mid-Atlantic), and ISO-NE (northeast) trading hubs. 78 Except for wind, which operates intermittently across all hours.
Level Nuclear Wind Solar NGCC1 70.0% 25.0% 15.0% 30.0%2 73.3% 26.7% 15.8% 36.7%3 76.7% 28.3% 16.7% 43.3%4 80.0% 30.0% 17.5% 50.0%5 83.3% 31.7% 18.3% 56.7%6 86.7% 33.3% 19.2% 63.3%7 90.0% 35.0% 20.0% 70.0%
6-19
Exhibit KKK: Capacity Factor and Energy Price Multiplier Curves
Fourth, an important part of the economic equation for many generators is the level
of government incentives they receive. These incentives can take several forms from
accelerated depreciation, to production tax credits, to investment tax credits. Each
technology is modeled based on the incentives to which they are entitled. Each power plant
type benefits from MACRS depreciation appropriate to its property class (5 years for
renewable projects, 15 years for nuclear, and 20 years for NGCC).79 Further, wind projects
are eligible for production tax credits amounting to $23/MWh for ten years (less 20% for
projects entering commercial operation in 2017). Lastly, solar projects are eligible for an
investment tax credit, paid at commercial operation, equal to 30% of the eligible cost basis.
These incentives play a significant role in altering the cash-flow profiles for renewable
projects.
Fifth, in order to incorporate uncertainty over future energy prices, I incorporated
five different “trajectories” to energy prices: flat, rising, falling, peak, and trough. The levels
are set in each such scenario to not alter the lifetime average price, but simply to shift that
79 MACRS is the Modified Accelerated Cost Recovery System and is used by the Internal Revenue Service to determine the amount of annual depreciation permitted for a given asset.
0%
50%
100%
150%
200%
250%
0.0x
0.5x
1.0x
1.5x
2.0x
2.5x
0% 20% 40% 60% 80% 100%
Ene
rgy
Pric
e R
eali
zed
Con
ditio
nal o
n C
apac
ity
Fac
tor a
nd R
elat
ive
to A
vera
ge
Prc
ie
Ene
rgy
Pri
ce R
elat
ive
to A
nnua
l Ave
rage
Percent of the Year
Price Duration Curve (left axis) Realized Price Curve (right axis)
6-20
price across time. Given that all energy prices are given in constant dollars, the “rising”
trajectory could indicate real increases in the price of energy. The “peak” trajectory could
indicate a “break-in” period early in a powerplant’s life, followed by performance
degradation as it approaches retirement. The remaining trajectories could be characterized
similarly. Exhibit LLL illustrates the five trajectories incorporated.
Exhibit LLL: Energy-Price Trajectories with Multipliers Relative to Average Price
Sixth, evaluations can be made using the expenses only, or the cash available for
finance calculations. In the expense-only option, all revenues are excluded, but incentive
payments are still allowed to offset expenses.
Lastly, the seventh, eighth, and ninth dimensions comprise the probabilities
associated with the capital cost, operational, and decommissioning phases of project life.
Three levels (5%, 50%, and 90%) are used during each phase and applied to the top branch
of each lottery. The probabilities for the bottom branch are set to be one minus the top branch
probability in each case.
1 Flat
2 Rising
3 Falling
4 Peak
5 Trough
1.0x 1.0x
0.75x
1.25x
0.75x
1.25x
0.75x
1.25x
0.75x
0.75x
1.25x 1.25x
6-21
Collectively, the combinations of these elements allow for 4 6 7 2 5 2
3 3 90,720 different possible prospects covering a wide range of cash flow or expense
profiles. From this set, 60,480 scenarios reflecting infeasible combinations are eliminated.
These include scenarios that include the presence of renewable energy incentives for nuclear
and NGCC and the presence of decommissioning expenses for non-nuclear plants. The
remaining 30,240, because of their construction, are also designed to be representative of the
cash flow profiles of actual projects.80 Therefore, these 30,240 power plant project prospects
are defined as the population of prospects, from which comparisons are drawn for evaluation
purposes in the next section.
6.4.3 Compression and Different Prospect Designs
This paper began by proposing to investigate the extent to which one form of problem
presentation—compression—influences choice, here in a specific application: choice
between power projects. As outlined in Exhibit DDD above, there are two parts to this
question: (i) for what types of problems is compression characterized as Simplifying,
Counterproductive, Irrelevant, or Useful, and (ii) for what types of decision makers is
compression similarly characterized? The intent of asking these questions is to receive a
better understanding of the role of compression, and the usefulness (or lack thereof) of LCOE
as a compressed metric and basis of a heuristic rule within a real-world problem
environment.
In addition to describing the general simulation environment used, this section
addresses the first question. Section 6.3.1 described the modeling process in general terms.
As to specifics, 6,000 random draws of two prospects from the population of 30,240
prospects are performed. Each pair of prospects is evaluated by a population of 500
simulated decision makers using each of three rules to produce a “choice triple.” Each choice
80 Similarly, scenarios that compare expense-only cash flow profiles to full cash flow profiles are eliminated in the sampling process as they represent unfair comparisons. Also, because the order in which prospects A and B are selected is not relevant, results from comparing Prospect i to Prospect j are consolidated with those comparison Prospect j to Prospect i.
6-22
triple is then classified as Simplifying, Counterproductive, Irrelevant, or Useful, as defined
in Section 6.3.1. I then further consolidate the outcomes as either Good (Simplifying or
Useful) or Bad (Counterproductive or Irrelevant) and indicating either Change
(Counterproductive or Useful) or No Change (Simplifying or Irrelevant) for purposes of
describing the results. Outcomes classified as Good represent those for which the use of
compression does no harm. Also of interest is whether or not the use of compression prompts
decision makers to change their decisions. Such changes may or may not have positive
outcomes, but are clearly evidence of a lack of procedural invariance.
I examine whether there are different characteristics of the prospects (described in
Section 6.4.2) that result in different outcomes. For example, does the use of compression
(i.e., levelizing) with prospects of a certain type (e.g., long-lived or accelerated cash flows)
result in “Good” or “Bad” outcomes when applied to choices involving power projects?
There are clear a priori reasons behind such questions that allow for the formulation of
hypotheses. Many of these reasons, however, have been based on largely abstract
experiments. Of interest is whether the findings from such experiments carry over into a
more realistic environment.
For example, the Short-Long Asymmetry studied by Albrecht and Weber [1997].
This finding suggests that the discount rates applied by decision makers decline over time,
in support of the hyperbolic discounting model, but in contrast to most normative theories
of discount rates. If such a finding holds, it might be expected to manifest in decision making
for power plant investments by biasing decisions made for investments with different time
horizons or that are subject to incentives that alter the temporal cash flow profile of an
investment. Accelerated depreciation, for example, shifts project values toward the present,
increasing returns from a normative perspective, but also shifting them toward a period
where decision makers tend to apply higher perceived discount rates, perhaps reducing the
perceived return of the investment.81
81 Accelerated depreciation is common to many renewable power investments. By accelerating the realization of depreciation over time, the project’s taxable income in early years is reduced (because depreciation is a tax-deductible expense), lowering cash taxes paid and allowing more cash to flow earlier to investors.
6-23
The question, then, is not simply whether or not two different projects would result
in different choices being made by normative versus heuristic decision makers, but to what
extent does the process of levelizing influence that choice? Just as accelerating depreciation
moves cash flow emphasis toward the present, levelizing “undoes” such an emphasis by
equalizing cash flows over time. Accordingly, it may be a “useful” metric in those instances.
Similarly, comparing projects with different expected lives may also create issues for
decision makers that exhibit Short-Long Asymmetry.
Consider, for the sake of example, two projects, A and B. Project A’s lifespan is 3
years; Project B’s lifespan is 5 years. In comparing the two projects, the discounted cash
flows of each project can be compared (as illustrated in Exhibit MMM). For a decision maker
using exponential discounting, Project A is preferred to Project B. However, for a heuristic
decision maker using hyperbolic discounting, Project B is preferred to Project A. If the
projects’ cash flows are levelized prior to being given to the heuristic decision maker to
compare, however, that decision maker would revert to the normative choice, with Project
A being preferred to Project B. This pattern, as described in Section 6.3, is considered
“Useful,” as the use of compression for this problem allows the heuristic decision maker to
align his choice with the normative decision maker.
The example below illustrates, for a single set of projects and a single decision
maker, the particular characterization (“Useful”) of compression as a heuristic. With the
simulation approach described previously, a far greater diversity of project types and
decision makers can be explored to identify the portions of the population likely to exhibit
each possible outcome and compare those results to hypotheses suggested by the existing
literature on risky intertemporal choice. In each case, a set of hypotheses is presented using
a common format (illustrated in Exhibit NNN) that illustrates the difference in decision-
maker population demographics (i.e., the percentage of the population resulting in the
Simplifying/Counterproductive/Irrelevant/Useful/Bad/Change outcome) under two project
trait conditions (e.g., does project trait “X” or “Y” result in a larger percentage of
Simplifying outcomes?).
6-24
Project A is indicated in blue. Project B is indicated in orange. The columns represent the per-period (undiscounted) cash flows for each project. The lines represent the cumulative present value of the cash flows, as measured by each type of decision maker. The discount rate is 15% and 0.5.
Normative
Exponential Discounting Heuristic with Full Profile
Hyperbolic Discounting Heuristic with Compression
Hyperbolic Discounting
Exhibit MMM: Comparison of Two Stylized Projects with Different Lifespans for which Compression is a Useful Heuristic
Exhibit NNN: Interpretation of Hypothesis Tables
I begin with Short-Long Asymmetry and projects with different lifespans as
described in Exhibit OOO. The results of this analysis are presented in Exhibit PPP (in this
exhibit and those that follow, hypotheses that are not supported are indicated in red text). In
this case, the only hypothesis that was not supported was the “Irrelevant” one. Otherwise, I
$0
$50
$100
$150
1 2 3 4 5
A≻B
$0
$50
$100
$150
1 2 3 4 5
A≺B
$0
$50
$100
$150
1 2 3 4 5
A≻B
6-25
conclude that the larger the lifespan difference between the power plants, the more useful
compression is and the more valuable LCOE is as a metric. In this case, it was achieved by
moving members of the population who had dropped out of the Simplifying class (indicating
that applying heuristic decision making was inducing errors) into the Useful class (indicating
that the application of compression corrected those errors). No meaningful movement was
observed in the Irrelevant or Counterproductive classes.
Next, one of the classic findings of the behavioral decision-making literature is the
asymmetry between gains and losses, with many decision makers demonstrating loss
aversion, suggesting that the framing of choices as gains or losses could potentially bias
decision making. Shelley [1994], among many others, demonstrated the existence of Gain-
Loss Asymmetry in risky intertemporal choice, concluding that discount rates were smaller
for gains than for losses. Recall that the original definition of LCOE was as an expense-only
measure, meaning that all lotteries were framed in terms of costs, or losses. I contrast this
with a more complete cash flow measure (Cash Available for Finance) that includes
revenues, and therefore has outcomes framed as both gains and losses. The set of hypotheses
for this evaluation is provided in Exhibit QQQ and the results are presented in Exhibit RRR.
6-26
Relative Lifespan the prospects are divided into subgroups by the ratio of the longer-lived plant’s lifespan to the shorter-lived plant’s lifespan
Choice Outcome Characterization
Hypotheses about the percentage of the decision-
maker population characterized in each way in each prospect subgroup Hypothesis Rationale
Projects have same
lifespan
Project lifespan
differential is 2x
Simplifying > Fewer time biases are initially present when projects have identical lifespans.
Counterproductive = No reason to believe levelizing would introduce error into an initially correct choice.
Irrelevant > Levelizing is not likely to correct a bias not likely to have originally been caused by time differences.
Useful < If the initial choice was non-normative and due to lifespan differences, levelizing more likely to correct.
Bad No prediction Hypotheses follow from individual components.
Change <
Exhibit OOO: Relative Life Hypotheses and Short-Long Asymmetry
Exhibit PPP: Population Classification According to the Relative Lifespans of the Prospects, Measured as Longer Lifespan Divided by Shorter Lifespan
0% 20% 40% 60% 80% 100%
Simplifying
Counterproductive
Irrelevant
Useful
Bad
Change
2.00 1.75 1.50 1.25 1.00
p < 0.001
No prediction
p < 0.001
p = n.s.
p = n.s.
p < 0.001
6-27
Cash Flow Profile Framing the prospects are divided into subgroups by whether or not they include only the expense-related cash flows or all the project net cash flows
Choice Outcome Characterization
Hypotheses about the percentage of the decision-
maker population characterized in each way in each prospect subgroup Hypothesis Rationale
Expense-only cash
flows
Net cash available for
finance
Simplifying <
Biases are less prominent when cash flows are mixed or predominantly positive. As a result, errors are less likely to be made initially and remain unchanged by levelizing.
Counterproductive <
If the initial choice was made normatively, for levelizing to be counterproductive, it most likely involves a sign change, which is only possible under the cash available for finance framing.
Irrelevant >
Although biases are more likely to be present in the loss framing of expenses-only, levelizing the cash flows does nothing to change their sign, and therefore initial biases are more likely to remain in place.
Useful < If an error was made initially, levelizing the full cash flows is more likely to preserve their (positive) sign.
Bad > Hypotheses follow from individual components.
Change <
Exhibit QQQ: Cash Flow Framing Hypotheses and Gain-Loss Asymmetry
Exhibit RRR: Population Classification According to Cash Flow Profile Framing
0% 20% 40% 60% 80% 100%
Simplifying
Counterproductive
Irrelevant
Useful
Bad
Change
Expenses Only Cash Available for Finance
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.01
p < 0.1
6-28
Inclusion of Incentives the prospects are divided into subgroups according to whether or not their cash flows include government-provided (i.e., project external) tax incentives for renewable electricity generation
Choice Outcome Characterization
Hypotheses about the percentage of the decision-
maker population characterized in each way in each prospect subgroup Hypothesis Rationale
Projects include
incentives
Projects do not include incentives
Simplifying No prediction
Counterproductive > Levelizing reverses the perceived acceleration of benefits from incentives, potentially leading to a preference reversal opposing the normative choice.
Irrelevant No prediction
Useful <
The absence of incentives increases the likelihood of realizing negative cash flows, which would trigger loss aversion, but would be reversed when levelizing restores a positive perceived valence to the cash flows.
Bad No prediction Hypotheses follow from individual components.
Change >
Exhibit SSS: Incentive Hypotheses and Short-Long Asymmetry
Exhibit TTT: Population Classification According to Use of Incentives
0% 20% 40% 60% 80% 100%
Simplifying
Counterproductive
Irrelevant
Useful
Bad
Change
Both Renewable, No Incentives Both Renewable with Incentives
p < 0.001
p < 0.001
p < 0.001
No prediction
No prediction
No prediction
6-29
Project Type the prospects are divided into subgroups based on whether or not they are comparing two power projects of the same type or of different types
Choice Outcome Characterization
Hypotheses about the percentage of the decision-
maker population characterized in each way in each prospect subgroup Hypothesis Rationale
Projects are the same
type
Projects are different
types
Simplifying > Projects of the same type are less likely to succumb to biases initially, and levelizing is not likely to introduce further bias.
Counterproductive > Levelizing should be more helpful on different types of projects, not less helpful.
Irrelevant < Comparing different types of projects is more likely to be problematic to begin with, and levelizing may not correct those problems.
Useful < Levelizing would be expected to help correct the errors associated with comparing different types of projects.
Bad No prediction Hypotheses follow from individual components.
Change <
Exhibit UUU: Project Type Hypothesis and Gain-Loss and Magnitude Asymmetries
Exhibit VVV: Population Classification Across Project Types vs. Within Project Types
0% 20% 40% 60% 80% 100%
Simplifying
Counterproductive
Irrelevant
Useful
Bad
Change
Different Plant Types Same Plant Types
p < 0.001
No prediction
p < 0.001
p = n.s.
p < 0.001
p < 0.001
6-30
Third, both Short-Long Asymmetry and Gain-Loss Asymmetry can impact the
comparison of projects presented with and without tax-related incentives. The inclusion of
incentives in a project’s cash flows tilts the cash flow profile of the project toward more
present-biased returns. Similarly, many renewable projects are likely to show losses at
various periods of time without the incentive payments, raising the influence of loss aversion
on decision makers. The comparisons here must be made carefully, as the inclusion of tax
incentives in a project’s cash flows can only be positive from a normative perspective. Thus,
the comparison of interest is not between projects with and without incentives, but within
each class (i.e., between two projects presented with incentives or between two projects
presented without incentives). For example, when projects are compared without inclusion
of incentives, the resulting cash flows are more likely to be negative. In those cases, a loss
aversion bias may present, which would be mitigated by levelizing those cash flows (and
thus returning them to uniformly positive), making levelizing more Useful for projects
presented without incentives. In contrast, when incentives are present, losses are less likely,
and therefore only the Short-Long Asymmetry effect is likely to be present. In those cases,
levelizing reverses the acceleration of perceived benefits, leading to more weight being
placed on future benefits contrary to the normative choice rule. The set of hypotheses for
this evaluation is provided in Exhibit SSS and the results are presented in Exhibit TTT.
Fourth, levelizing is less likely to be useful for projects of the same type, which are
more likely to exhibit similar cash flow profiles. In comparing projects of different types
(e.g., nuclear vs. wind, solar vs. NGCC), both Gain-Loss Asymmetry and Magnitude
Asymmetry may have an effect. Magnitude Asymmetry results in perceived discount rates
being higher for larger payoffs [Thaler, 1981]. Projects of different types are more likely to
have per-period payoffs of different magnitudes, making decision makers considering them
more susceptible to Magnitude Asymmetry. Projects of the same type are more likely to
have similar cash flow profiles and fewer instances of differential gain and loss occurrence.
In other words, when projects are similar and both succumb to similar biases, the effects
may tend to “cancel each other out.” Therefore, compression is more likely to be a beneficial
force in comparing projects of different types.
6-31
The set of hypotheses for this evaluation is provided in Exhibit UUU and the results
are presented in Exhibit VVV. Interestingly, most of the hypotheses related to plant type
similarity were not supported. Levelizing proved to be more Simplifying for different plant
types and more Useful for the same plant types, both in contradiction to my hypotheses.
These effects reverse, however, if I also control for whether or not the plants are renewable
(wind or solar). If the comparison is limited only to renewable projects, the original
hypotheses become well supported (except for the very small Counterproductive class).
Exhibit WWW illustrates the differences. I would speculate that the renewable project types,
which tend to have the most variable cash flow profiles due to their incentives, are more
likely to emphasize the differences that would create the expected effect from similar or
different plant types.
Exhibit WWW: Refining the Population Classification Across Project Types by Controlling for Renewable Power Types
Lastly, I examine the influence of each of the prospect characteristics on the outcome
classifications. Exhibit XXX illustrates a considerable amount of similarity in which
prospect characteristics are influential on the role of compression. To help clarify the results,
I limited the analysis only to cases where the effects were greater than 75% of the population
or less than 25% of the population for each class. Prospect characteristics that do not seem
to matter include the ages of the plants being compared, the trajectory of their cash flows
(other than through incentives), and the probabilities associated with their capital cost and
operating expenses.82 In contrast, characteristics that do matter for outcome classification
are: (i) whether or not the power plants are renewables, (ii) the capacity factors of the plants,
82 The somewhat surprising finding that probabilities alone do not appear to matter is reminiscent of March and Shapira’s [1987] claims about managerial insensitivity to estimates of outcome probabilities.
Hypothesis
Change < Not supported p < 0.001 Supported p < 0.001
Bad No prediction No prediction No prediction
Useful < Not supported p < 0.001 Supported p < 0.001
Irrelevant < Not supported n.s. Supported p < 0.05
Counterproductive > Supported p < 0.001 Not supported p < 0.05
Simplifying > Not supported p < 0.001 Supported p < 0.001
Original Result Renewable-only Result
6-32
(iii) the presence of incentives for renewable energy, (iv) whether the projects are described
on an expense-only basis, and (v) the risks associated with decommissioning expenses for
nuclear power projects. In each of those five cases, dividing the prospects along that
characteristic produced a significant difference in the outcome classification, suggesting that
the use of levelizing in projects with those characteristics is likely to have an impact on the
choices made.
Exhibit XXX: Influence of Prospect Characteristics on Decision Problems with Compression
To conclude, it should be noted that these impacts are not necessarily negative ones.
In fact, the results suggest that roughly two-thirds of the time, presentation compression
through levelizing is a Simplifying metric when comparing power projects. More
importantly, however, between 20% and 40% of the time, compression through levelizing is
a Useful metric, meaning that it corrects for biases that would otherwise lead decision makers
astray from the normative choices in comparing power projects.
t statistic p value t statistic p value t statistic p value t statistic p value t statistic p value t statistic p value
A - Renewable -4.19 **** -6.49 **** 1.51 3.08 *** -0.16 3.55 ***
A - Age -0.69 0.32 -0.87 1.34 -0.89 1.59
A - Capacity Factor 6.67 **** 6.62 **** -3.45 *** -5.71 **** -1.27 -5.38 ****
A - Incentives -3.88 **** -7.65 **** -2.40 ** 8.19 **** -3.73 **** 6.58 ****
A - Trajectory -0.99 -5.22 **** 1.18 0.61 0.44 0.69
A - Expenses Only -10.25 **** -10.55 **** 14.34 **** 10.51 **** 9.73 **** 6.80 ****
A - Cost Probability 0.53 -3.77 **** 1.11 -1.28 0.66 -1.63
A - Operating Probability -0.29 0.21 0.83 -0.56 0.65 -0.10
A - Decommissioning Probability 7.93 **** 6.44 **** -7.48 **** -4.79 **** -4.74 **** -4.15 ****
B - Renewable -4.59 -12.25 **** 10.54 **** 1.42 * 5.72 **** -1.06 **
B - Age 0.71 -5.77 **** -5.32 1.84 -6.32 2.40
B - Capacity Factor -0.22 -6.20 **** -0.61 1.42 ** -1.14 *** 1.16 *
B - Incentives 0.97 * 9.49 **** 0.85 **** -2.34 **** 3.37 **** -1.81 ****
B - Trajectory -1.97 -8.37 -3.91 7.22 -5.43 5.30
B - Expenses Only 0.53 **** -1.61 **** -0.03 **** -0.23 **** -0.10 **** -0.13 ****
B - Cost Probability -10.25 -10.55 ** 14.34 10.51 9.73 ** 6.80
B - Operating Probability -0.56 2.33 1.63 -1.16 2.08 -0.39 **
B - Decommissioning Probability 1.74 **** -0.29 * -0.45 **** -1.19 -0.50 *** -2.05
**** = significance at the 0.001 level
*** = significance at the 0.01 level
** = significance at the 0.05 level
* = significance at the 0.10 level
Bad ChangeSimplifying Counterproductive Irrelevant Useful
6-33
6.4.4 Compression and a Heterogeneous Population of Decision Makers
Having explored the effect of compression on prospects with different
characteristics, I now ask the question in the other direction: for what types of decision
makers is compression likely to be more or less beneficial? In the previous section, I reported
the percentage of the population exhibiting specific choice outcomes (e.g., Simplifying,
Useful). By extending the analysis backward into the simulated population of decision
makers, however, I can examine the characteristics of decision makers for which
compression is categorized by each outcome classification.
Such an analysis is important because if investors can be “typed,” then information
and problem presentations can be tailored to be of particular use to them (or even against
their interests, if persuasion is the goal). Is compression, for example, more beneficial to a
particular subset of the population of decision makers? If a particular decision maker is
known to be extremely loss averse, for example, should or should not compression be used
as a metric? On the other hand, if the choice outcomes resulting from use of compression
are broadly stable across a population, then one may conclude that it has considerable value
(assuming, of course, that the analysis described in Section 6.4.3 indicated that compression
was also a beneficial force relative to the prospect characteristics).
I hypothesize that compression (here as levelizing as applied to power projects) will
be disproportionately useful for more loss-averse decision makers (large ), and by
extension, for decision makers with elevated levels of risk seeking in the domain of losses
(smaller ), and for decision makers with more nonlinear probability weighting for losses
(smaller ). Because levelizing moderates the appearance of extreme outcomes, which
might otherwise bias decision makers’ choices, decision makers with those characteristics
are likely to be disproportionately benefitted by compression.
At first glance in Exhibit YYY, only three coefficients appear to have meaningful
differences in mean when the population of decision makers is divided by outcome class.
Although the large sample sizes mean that all of the differences are statistically significant,
6-34
the differences only appear meaningful for , , and . The effects on the remaining
coefficients are de minimis. Exhibit ZZZ illustrates, for example, the entire distribution for
the coefficient when the group is further divided by renewable generator and by
presentation as expense only or CAFF. I do not consider the coefficients without meaningful
differences further.
Exhibit YYY: Percentage Variation of Subgroup Mean Coefficient Value from Population Coefficient Means
Exhibit ZZZ: Example of De Minimis Impact on from Further Dividing the Group
-12%
-10%
-8%
-6%
-4%
-2%
0%
2%
4%
6%
8%
Alpha Beta Lambda Gamma+ Gamma- Psi
Dev
iatio
n fr
om P
opul
atio
n M
ean
Simplifying Counterproductive Irrelevant Useful
0%
20%
40%
60%
80%
100%
0.3 0.4 0.5 0.6 0.7 0.8 0.9
Cum
ulat
ive
Pro
babi
lity
AlphaRenewable-Expense Only Not Renewable-Expense OnlyRenewable-CAFF Not Renewable-CAFFTotal Population
6-35
The inclusion of as a meaningful variable, however, is somewhat surprising.
Because levelizing does not change the duration of cash flow profiles, I had not expected the
population of decision makers to subdivide on that coefficient. However, the shifting of cash
flows within terms of the same duration appears to have a meaningful impact on how
different proportions of the population of decision makers responds. I explore each
meaningful coefficient in turn.
The hypothesis for beta was that increasing risk-seekingness (smaller ) would be
associated with compression being more beneficial. The two primary effects found in
Exhibit YYY support this pattern. Decision makers with higher (those that are less risk-
seeking), find compression Simplifying, in that they are less likely to be prone to biases
initially, and therefore levelizing allows for task simplification without additional effect. In
contrast, decision makers with lower (those that are more risk-seeking), find compression
Useful, in that their increased susceptibility to biases allows levelizing to mitigate the effects
of such biases. If one peers further into the data, additional separation is apparent. The
separation of the population into less and more risk-seeking in the Simplifying and Useful
classes appears to be driven entirely by choices made regarding non-renewable projects
using full CAFF profiles. Exhibit AAAA illustrates the entire distribution for each cross
between decision maker characteristic and project characteristic for the Simplifying and
Useful outcome classes. The Simplifying result is obtained primarily by predominantly less
risk-seeking decision makers confronted with non-renewable and full CAFF projects; in
contrast, the Useful result is obtained primarily by predominantly more risk-seeking decision
makers confronted with those same problems.
The hypothesis for lambda was that compression would be more beneficial for
decision makers with higher levels of loss aversion (larger ). As Exhibit YYY illustrates,
the results here are similar. The Simplifying class is associated with increased levels of loss
aversion. In contrast, the Useful) class is associated with lower levels of loss aversion. In
terms of the beneficial or “Good” (Simplifying + Useful) classes, the results are roughly off-
setting. More loss-averse decision makers find compression Simplifying, but less loss-averse
decision makers are the ones finding compression Useful. Exhibit BBBB further
6-36
decomposes these results by relating them to prospect characteristics. The Simplifying and
Irrelevant classes are largely consistent in their response to different prospect types. More
variation between prospect types is apparent in the Counterproductive and Useful classes. In
those cases, the populations divide along the cash flow presentation lines, with decision
makers with lower levels of loss aversion finding compression Counterproductive under the
expense-only presentation and Useful under the full CAFF presentation.
Exhibit AAAA: Connection of Decision Maker Characteristics ( ) to Project Characteristics for the Simplifying and Useful Outcome Classes
0%
20%
40%
60%
80%
100%
0.30 0.50 0.70 0.90Cum
ulat
ive
Pro
babi
lity
Beta - Simplifying Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
0%
20%
40%
60%
80%
100%
0.30 0.50 0.70 0.90Cum
ulat
ive
Pro
babi
lity
Beta - Useful Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
6-37
Exhibit BBBB: Connection of Decision Maker Characteristics ( ) to Project Characteristics for All Outcome Classes
0%
20%
40%
60%
80%
100%
1.00 1.25 1.50 1.75 2.00 2.25Cum
ulat
ive
Pro
babi
lity
Lambda - Simplifying Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
0%
20%
40%
60%
80%
100%
1.00 1.25 1.50 1.75 2.00 2.25Cum
ulat
ive
Pro
babi
lity
Lambda - Counterproductive Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
0%
20%
40%
60%
80%
100%
1.00 1.25 1.50 1.75 2.00 2.25Cum
ulat
ive
Pro
babi
lity
Lambda - Irrelevant Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
0%
20%
40%
60%
80%
100%
1.00 1.25 1.50 1.75 2.00 2.25Cum
ulat
ive
Pro
babi
lity
Lambda - Useful Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
6-38
Exhibit CCCC: Connection of Decision Maker Characteristics ( ) to Project Characteristics for All Outcome Classes
0%
20%
40%
60%
80%
100%
0.00 0.25 0.50 0.75 1.00Cum
ulat
ive
Pro
babi
lity
Psi - Simplifying Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
0%
20%
40%
60%
80%
100%
0.00 0.25 0.50 0.75 1.00Cum
ulat
ive
Pro
babi
lity
Psi - Counterproductive Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
0%
20%
40%
60%
80%
100%
0.00 0.25 0.50 0.75 1.00Cum
ulat
ive
Pro
babi
lity
Psi - Irrelevant Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
0%
20%
40%
60%
80%
100%
0.00 0.25 0.50 0.75 1.00Cum
ulat
ive
Pro
babi
lity
Psi - Useful Only
Renewable-Expense Only
Not Renewable-Expense Only
Renewable-CAFF
Not Renewable-CAFF
6-39
I had not originally hypothesized an effect for psi, but Exhibit YYY illustrated a clear
effect. Compression is likely to lead to a Change (Counterproductive + Useful) outcome for
more present-biased ( → 1) decision makers, while the No Change (Simplifying +
Irrelevant) classes were realized primarily by less present-biased ( → 0) decision makers
(or those closer to exponential discounting). These mixed results are more complicated to
interpret, but also suggest more concern regarding the application of compression, since
particular classes of decision makers appear predisposed to outcomes when faced with
certain prospects that could be either good or bad.
Exhibit CCCC further decomposes the psi results according to prospect
characteristics. Here, it is the Counterproductive decomposition that is of interest. Especially
among moderately present-biased decision makers, but also across the range of the degree
of hyperbolicity exhibited, decision makers faced with non-renewable projects were 20-30
percentage points more likely to end up classified as Counterproductive. Because the
primary distinguishing (cash flow) feature of renewable projects (vis-à-vis non-renewable
projects) is the presence of incentives that accelerate the realization of value across time, it
is understandable that renewable projects may be more impacted by compression, which by
equalizing cash flows over time shifts more of the accelerated gains back into the future
where they may be perceived to have more value. Such results are then less likely to be
classified as Counterproductive than their non-renewable peers.
6.5 Applications and Implications
The purpose of this chapter is to investigate decision making relative to problem
presentation. This is not a new area of research. The particular contribution of my research
comprises multiple distinct elements. First, the novel computational decision-making
simulation approach developed in Chapter 5 is applied to a specific set of corporate finance
questions involving risky intertemporal choice. Second, a heuristic choice framework is
developed to allow for the assessment of compression as a useful form of presentation and
levelized cost as an efficient metric. Third, effort is made to confront the simulated subjects
with prospects designed to exhibit a high degree of realism as power project investments.
6-40
Fourth, the computational approach allows for the examination of a heterogeneous pool of
prospect design elements in combination with a heterogeneous pool of decision makers. By
connecting these approaches, linkages can be established between decision maker
characteristics and prospect characteristics that can assist in allowing compression to be
more effectively used in power plant investment decision making.
In practical terms, my research suggests that users of the compressed LCOE
presentation form should be informed as to areas where reliance on that presentation is likely
to trigger biases and thereby cause them to make non-normative choices. Just as Joskow
[2011] has shown that LCOE is “not a useful way to compare generating technologies [that
have] different production profiles” (e.g., baseload (nuclear) vs. intermittent (wind)), this
work shows that LCOE is also not a universally reliable way for decision makers with certain
cognitive characteristics to compare technologies with certain risk profiles (e.g., high fixed
cost vs high variable cost) or distributions of costs and benefits across time (e.g., heavily-
incentivized renewables vs fossil or nuclear). Specifically, this lack of universal reliability
is because compression resulting from the metric may tend to alter the decision problem’s
characteristics in ways that are known to trigger biases. Compression, for example, obscures
the presence of risk by replacing explicit probabilities with expected values. Certainty-risk
asymmetry results in decision makers applying higher discount rates to certain outcomes.
Likewise, compression explicitly “levelizes” cash flows over time, and short-long
asymmetry results in discount rates declining as events are pushed into the future (which
levelizing, in effect, does with the accelerated incentives provided for renewable power
projects). In many cases, however, it is the removal of these details via compression that
mitigates biases. Although Joskow’s [2011] recommendation was to “abandon” levelized
cost comparisons, I propose a less drastic response for cognitively-constrained decision
makers.
A heuristic is “good” if it allows a simpler decision process to be substituted (in the
sense of Kahneman and Frederick’s [2002] “attribute substitution”) reliably for a more
complicated but normative decision process. This work identified classes of areas where the
6-41
use of LCOE leads heuristics to become unreliable, but also areas where compression as a
metric and its related heuristic could be a beneficial force and useful for debiasing.
At the most general level, the use of compression in power-investment decision
making is beneficial for a majority of projects and decision makers. In most cases, it is
Simplifying. However, this research also identifies particular circumstances for which
compression is not a beneficial force. As I noted at the outset, for heuristics to be used
effectively, their strengths and weaknesses must be well understood. Accordingly, the results
of this study should inform the use of compression (levelizing) as a metric. To summarize
the practical implications of these findings, this research shows that:
1. Compression is more Useful when comparing projects with large lifespan
differences. Because different types of power project tend to have different
estimated lifespans, project type can be a proxy for lifespan in many applications.
2. Compression is more Useful when projects are evaluated with their full cash
flows, rather than only the expenses, as is commonly done with levelized cost.
3. In fact, use of the expense-only framing results in approximately twice as many
Bad (Counterproductive or Irrelevant) outcomes as the full CAFF framing. This
is, in a sense, restating Joskow’s [2011] conclusion in more decision-making
language, even if the underlying sentiment is similar.
4. Levelizing is more Useful in no-incentive comparisons. Levelizing tends to
reverse the acceleration of value provided by many incentives for renewable
power projects. As a result, any attempt that shifts any perceived value into the
future may be overvalued by present-biased decision makers, but contrary to
what a normative decision maker should prefer. As levelized cost tends to be
disproportionately used for evaluating renewable power projects, this result
should prompt some caution.
5. Project type matters only for renewable power projects. It is Useful for comparing
among types of renewable power projects, but not for comparing renewable
projects to non-renewable projects.
6-42
6. Three parameters show meaningful separation by outcome classification: risk-
seekingness under losses, loss aversion, and present bias.
7. The application of compression by less risk-seeking or more loss-averse decision
makers tends to produce Simplifying outcomes, while the application of
compression by more risk-seeking or less loss-averse decision makers tends to
produce Useful outcomes. These results are driven primarily by comparisons
involving non-renewable projects, together with full CAFF framing.
8. The results for present bias (degree of hyperbolicity) were mixed, suggesting that
both beneficial and harmful outcomes could occur in certain circumstances.
9. For moderately present-biased decision makers and non-renewable projects, such
decision makers are 20-30 percentage points more likely to end up with a
Counterproductive outcome.
6.5.1 Compression as a Tool for Debiasing
Although in this chapter I have asked the question from a “top down” or “information
supplier” perspective, inquiring as to whether letting decision makers use compressed
metrics is beneficial, it is also worth contemplating a “bottom up” or “information
consumer” interpretation of this problem. Since many business decision makers see the
compressed metrics (such as LCOE or NPV) first, one might ask whether there is value in
additionally requesting full information about a problem. Such a question is rarely
considered in normative analysis because additional information always has non-negative
value. Numerous researchers, however, have questioned the value of providing more
information to cognitively-constrained decision makers [Camerer, Loewenstein, and Weber,
1989; Kleinmuntz and Schkade, 1993; Iyengar and Lepper, 1999].
In fact, consider the cognitively-constrained information consumer’s most basic
question: “how much information should I request or consider?” A simple objective function
for such a question is the percent of the time the requested information results in a “correct”
outcome (in that the option selected is consistent with a normative choice rule). By
examining the outcomes of choices made by each of the 500 decision makers in the simulated
6-43
population for each of the 6,000 prospect comparisons in the population of problems,
answers to this question emerge. Exhibit DDDD tabulates these results.
Compressed
Correct Incorrect
Full Correct
69% (Simplifying)
2% (Counterproductive)
71%
Incorrect 22%
(Useful) 7%
(Irrelevant) 29%
91% 9%
Exhibit DDDD: Information Consumer Choice Strategy Outcomes
A decision maker using the full presentation can expect to achieve the normatively
correct outcome 71% of the time (ranging from 50% to 90% across all decision makers).
However, that same decision maker using the compressed LCOE presentation instead can
expect to achieve the normatively correct outcome fully 91% of the time (ranging from 88%
to 99% across all decision makers). For these cognitively-constrained decision makers,
requesting less information improves their decision performance, suggesting that the biases
they exhibit in processing the full information can be mitigated mostly by compressing much
of that detail. I also look at the decision maker-by-decision maker and comparison-by-
comparison performance of such a rule. For 100% of decision makers, the frequency of
Useful outcomes exceeds the frequency of Counterproductive outcomes. Likewise, for 96%
of all project comparisons, the frequency of Useful outcomes exceeds the frequency of
Counterproductive outcomes. The performance of the compression heuristic first-order
stochastically dominates decisions made using the full presentation form.
As a result, a decision strategy that comprises “request the compressed information
only and act on it” proves to be a powerful tool for improving the performance of
cognitively-constrained decision makers. LCOE may have normative problems, but
compression is a broadly beneficial heuristic for bias-prone decision makers. The benefits it
provides from allowing such decision makers to avoid the consequences of their biases
appears to exceed the costs of LCOE’s normative failings.
6-44
The compression heuristic is sufficiently powerful that there does not appear to be
incremental value in using more information with it. Consider the case of a decision maker
with the compressed presentation who is offered the full presentation result as well. In this
case, the decision maker observes two outcomes that either agree or disagree. The outcomes
agree 76% of the time (the diagonal results in Exhibit DDDD) and disagree 24% of the time
(the off-diagonal results in Exhibit DDDD). But if there is agreement (i.e., the full
presentation matches the compressed presentation), the decision maker remains with the
outcome precipitated by the compressed presentation. Meanwhile, if there is disagreement
(i.e., the full presentation conflicts with the compressed presentation), the performance of
the compressed presentation remains better. As a result, the decision maker’s behavior is
unchanged and therefore there is no additional value in obtaining both problem
presentations.83 However, because in most circumstances requesting more information does
not impact the choice made, it can nevertheless be entertained if some ancillary benefit
results (e.g., it improves post-decision satisfaction).
These results concerning decision rule use are significantly impacted, however, by
the cognitive profile of the decision maker (as represented by the simulated 7-tuple of
prospect theory and hyperbolic discounting parameters). Decision makers who are more
loss-averse, more risk-seeking, and more present-biased than the population median can shift
the likelihood of a Counterproductive (or Useful) outcome given that a change in choice
occurs by as much as 30%. Exhibit EEEE provides a “heuristic map” of where different
cognitive subsets of the general population are more or less likely to realize
Counterproductive or Useful outcomes as a result of changing their choices upon receipt of
full information. Likewise, as illustrated in Exhibit FFFF, the usefulness of the LCOE
heuristic varies across problem characteristics and decision maker characteristics as well.
The exhibit illustrates the accuracy of the full and compressed heuristics across various
subpopulations of prospects and decision makers (based on a median split of decision maker
83 76% of the time, the two presentations result in the same choice. At that point, the decision maker can nominally elect to remain with the Compressed (C) choice or switch to the Full (F) choice. However, as they are the same in this branch, there is no choice, and the correct outcome results in 69/76 of the cases. In contrast, 24% of the time the two presentations disagree. In that case, remaining with C results in the correct outcome 22/24 of the time. Because 76% × (69/76) + 24% × (22/24) = 69 + 22 = 91%, the additional information has no incremental value over the “request the compressed information and act on it” strategy.
6-45
subpopulations). Although the compressed presentation is more accurate than the full
presentation in every instance, the performance of the compressed presentation is
particularly high in prospect comparisons involving full cash flows (i.e., CAFF) and
different technologies. This conclusion was noted previously, but I add here that it holds
across different decision-maker subpopulations as well.
If, as results such as Albert [1967] suggest, businesspeople differ “significantly”
from the general population of decision makers, it becomes important to be able to evaluate
decisions conditional on the characteristics of that subpopulation. The ability of the
simulation approach developed in Chapter 5 and applied here to specify such subpopulations
against, for example, simple calibration tasks, makes it especially well-suited to evaluate
such decision-making environments.
Although this chapter has focused specifically on decision making involving power
plant investments, there is a broader message in the results. Heuristic decision making can
be a powerful tool for reducing cognitive load. Likewise, the metrics so commonly used in
finance, but rarely contemplated from a decision-making perspective, can be beneficial, but
as with all tools, their proper use is required to make most effective use of their benefits.
This chapter examines an application to a single domain; other problem domains may not
find similar benefits for compression. The simulation approach developed in Chapter 5 and
applied here, however, allows for the in-depth examination of decision problems with
realistic levels of complexity among a diverse and flexible artificial population. The
identified advantages and weaknesses of compression—regardless of domain—should be
used by policy and investment decision makers to improve the effectiveness of information
given to problem stakeholders and optimize the impact of heuristic decision making.
6-46
Exhibit EEEE: Heuristic Map of Change-in-Choice Consequences
0.00
0.20
0.40
0.60
0.80
1.00
0%10
%20
%30
%40
%50
%60
%70
%80
%90
%10
0%
Alpha
Fre
quen
cy o
f O
utco
me
| Cho
ice
is C
hang
ed
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0%10
%20
%30
%40
%50
%60
%70
%80
%90
%10
0%
Gamma+
Fre
quen
cy o
f O
utco
me
| Cho
ice
is C
hang
ed
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0%10
%20
%30
%40
%50
%60
%70
%80
%90
%10
0%
Beta
Fre
quen
cy o
f O
utco
me
| Cho
ice
is C
hang
ed
Incr
easi
ngly
risk
-see
king
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0%10
%20
%30
%40
%50
%60
%70
%80
%90
%10
0%
Gamma–
Fre
quen
cy o
f O
utco
me
| Cho
ice
is C
hang
ed
0.00
0.50
1.00
1.50
2.00
2.50
0%10
%20
%30
%40
%50
%60
%70
%80
%90
%10
0%
Lambda
Fre
quen
cy o
f O
utco
me
| Cho
ice
is C
hang
ed
Cou
nter
prod
ucti
ve |
Cha
nge
Use
ful |
Cha
nge
Lin
ear
(Cou
nter
prod
ucti
ve |
Cha
nge)
Lin
ear
(Use
ful |
Cha
nge)
Incr
easi
ngly
loss
-ave
rse
0.00
0.20
0.40
0.60
0.80
1.00
1.20
0%10
%20
%30
%40
%50
%60
%70
%80
%90
%10
0%Psi
Fre
quen
cy o
f O
utco
me
| Cho
ice
is C
hang
ed
Cou
nter
prod
ucti
ve |
Cha
nge
Use
ful |
Cha
nge
Lin
ear
(Cou
nter
prod
ucti
ve |
Cha
nge)
Lin
ear
(Use
ful |
Cha
nge)
Mor
epr
esen
t-bi
ased
6-47
Exhibit FFFF: Rule Accuracy by Prospect and Decision Maker Characteristics
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
71%
1%72
%C
orre
ct74
%1%
75%
Cor
rect
63%
1%64
%C
orre
ct82
%1%
83%
Cor
rect
76%
1%77
%C
orre
ct69
%1%
70%
Inco
rrec
t27
%1%
28%
Inco
rrec
t23
%1%
25%
Inco
rrec
t34
%2%
36%
Inco
rrec
t16
%1%
17%
Inco
rrec
t21
%2%
23%
Inco
rrec
t29
%1%
30%
98%
2%97
%3%
97%
3%98
%2%
97%
3%98
%2%
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
63%
6%69
%C
orre
ct64
%6%
70%
Cor
rect
63%
6%69
%C
orre
ct64
%6%
70%
Cor
rect
63%
6%69
%C
orre
ct64
%6%
70%
Inco
rrec
t24
%8%
31%
Inco
rrec
t23
%7%
30%
Inco
rrec
t24
%8%
31%
Inco
rrec
t23
%7%
30%
Inco
rrec
t24
%8%
31%
Inco
rrec
t23
%7%
30%
87%
13%
87%
13%
87%
13%
87%
13%
87%
13%
87%
13%
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
71%
1%72
%C
orre
ct71
%1%
72%
Cor
rect
71%
1%72
%C
orre
ct71
%1%
72%
Cor
rect
71%
1%72
%C
orre
ct71
%1%
72%
Inco
rrec
t16
%13
%28
%In
corr
ect
15%
13%
28%
Inco
rrec
t16
%13
%28
%In
corr
ect
15%
13%
28%
Inco
rrec
t16
%13
%28
%In
corr
ect
15%
13%
28%
87%
13%
86%
14%
87%
13%
86%
14%
87%
13%
86%
14%
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
Inco
rrec
tC
orre
ctIn
corr
ect
Cor
rect
61%
4%65
%C
orre
ct62
%4%
66%
Cor
rect
61%
4%65
%C
orre
ct62
%4%
66%
Cor
rect
61%
4%65
%C
orre
ct62
%4%
66%
Inco
rrec
t28
%7%
35%
Inco
rrec
t28
%7%
34%
Inco
rrec
t28
%7%
35%
Inco
rrec
t28
%7%
34%
Inco
rrec
t28
%7%
35%
Inco
rrec
t28
%7%
34%
89%
11%
90%
10%
89%
11%
90%
10%
89%
11%
90%
10%
Num
bers
may
not
add
due
to r
ound
ing
Exp
ense
Onl
yS
ame
Tec
hnol
ogy
Bet
aL
ambd
aP
si
CA
FF
Dif
fere
nt T
echn
olog
y
CA
FF
Sam
e T
echn
olog
y
Exp
ense
Onl
yD
iffe
rent
Tec
hnol
ogy
Bel
ow M
edia
nA
bove
Med
ian
Bel
ow M
edia
nA
bove
Med
ian
Bel
ow M
edia
nA
bove
Med
ian
7-1
7 Final Thoughts & Future Directions
7.1 Securities Design
The virtually limitless bundling and unbundling of risks possible in theoretical
finance faces a very pedestrian challenge in the real world: “consumer” marketing. In the
absence of arbitrage bounds in tradable markets for novel securities, issuers must take into
account the ability of potential buyers to understand the risks being transferred, to price them
rationally, and to believe in the lack of informational asymmetry in the transaction. This is
often a tall order when firms tend toward design complexity as a means of exerting oligopoly
power [Henderson and Pearson, 2011]. This oligopoly power, which manifests in larger-
than-equilibrium bid-ask spreads in novel securities is a two-edged sword though: issuer and
market-maker profits increase right up to the point where buyers withdraw from the market.
The first essay in this dissertation developed eight categories of design features that
can affect the successful marketing of a novel security to investors motivated by heuristic
decision making. Although the application addressed was the securitization of catastrophic
risk, which suffered from an initial reluctance of investors to participate, the resulting design
principles can be extended to other securities.
For example, Shefrin [2000], among others, has noted that the lottery bonds issued
by several countries (the U.K., Sweden, and others) remain popular investments. Lottery
bonds are bonds for which the investor is guaranteed return of principal, but in lieu of a
coupon, receives a “lottery ticket” entitling them to a chance at winning a large prize. These
bonds, in many ways, are strikingly similar to catastrophe bonds if viewed abstractly (they
have uncorrelated returns, they have features that result in small probabilities of large
changes in value, etc.). And yet, they are perceived very differently by investors. Cat bonds
are less attractive than they “should” be; lottery bonds are more attractive than they “should”
be. The implication of my investigation into cat bonds is that, when attempting to appeal to
7-2
investor clienteles in the creation of novel securities, issuers should take a more holistic view
of investors’ preferences and biases.
The areas of financial innovation and securities design have significant practical
relevance, but have received relatively little attention from a behavioral economics
perspective. Future work could explore other novel securities, but perhaps more importantly,
it could serve to guide the creation of novel securities. Finance has dealt extensively with
the notions of clienteles formed by tax policy, by regulatory policy, and by market frictions,
but has generally not sought to cater explicitly (or at least openly) to clienteles formed by
investors’ decision-making practices. This path of research potentially also provides a way
to test whether the involvement of high stakes can be used to incentivize investors to engage
in debiasing behavior. Could behaviorally-prompted securities innovations endure, for
example, if they prove to be suboptimal investments?
For corporations, this research suggests that it should be possible to design a security
to exploit investors’ biases in such a way as to lower the issuer’s cost of capital. My analysis
of cat bonds suggested that they were (initially and unintentionally) designed almost as
counterproductively as possible, triggering a variety of biases and behaviors that made
investors averse to owning them. One might ask whether it would be similarly possible (with
intention) to design securities to take advantage of as many biases and behaviors as possible.
Lottery bonds appear to be a good example, but how much further could structures be
pushed?
For regulators, this research suggests that disclosures and communication about
securities offerings could be evaluated for material that, while not intentionally misleading
or false, may be likely to trigger suboptimal choice among investors. Whether or not this is
a legitimate function of government regulators is a separate question. Should corporations
be allowed to design securities in a way that appeals to the non-normative preferences of
investors? While government may have a legitimate interest in preventing inaccurate or
misleading disclosure to potential investors, it is far from clear that barring corporations
from designing securities to appeal explicitly to investors’ preferences (whether rational or
7-3
not) would be equally legitimate (e.g., gambling is generally legal, but arguably not
“rational”).84
The co-evolution of this problem space, between issuer, investor, and regulator
interests, suggests a rich area for future research. Future research in this area may also benefit
from the structured computational analysis of risky intertemporal choice developed in
Chapter 5. A framework for the design of novel securities could be developed as a
constrained nonlinear optimization problem that seeks to maximize prospect theory values
while holding expected values constant by changing the structure of the cash flows. The
objective of this effort would be to increase the attractiveness of a security to a potential
investor while minimizing its expected cash cost to the issuer.
7.2 Regulatory Intermediation
Electric power regulators are charged with protecting both the public interest and the
economic stability of utilities. Neither of these mandates implies that rates for electric power
or utility profits cannot either increase or decrease over time. Many valid potential reasons
exist for their variation. My research, however, has shown that utilities’ authorized return
premiums (and earned returns such as net income) have generally increased over the past 35
years. Additionally, my research has failed to find any significant economic justification for
this steady increase. Thus, given that underlying riskless rates of return have generally
decreased over that period, profits to utilities have expanded significantly more than can be
explained by traditional theories in finance.
This is a problem that is somewhat unique to the regulated world. In a competitive
market, excess returns would be subject to the self-correcting forces of competition. But in
regulated markets, it is the intermediation of regulators that is intended to provide such
“correction.” Stakeholders, then, are relying on the judgment and decision making of those
84 This is not an abstract philosophical question. Traditionally, the U.S. Securities and Exchange Commission would take the position that so long as disclosure is complete, caveat emptor would prevail with regard to securities design and private contracts. The much newer Consumer Financial Protection Bureau (stated objective: “We protect consumers from unfair, deceptive, or abusive practices”) may have a different posture.
7-4
regulators to serve as an objective proxy for the market. In my work, the hypothesis that
cannot be rejected is that regulators are subject to biased judgments, and that a money
illusion-type effect has impeded allowed returns from dropping below a subjective 10%
threshold, thus increasing the risk premiums implicitly granted to utilities.
Over the last decade, retail prices for electricity have increased, while wholesale
prices for electricity have decreased. At a general level, the spread between retail and
wholesale prices, which I shall refer to as a gross spread, reflects profits to power generators.
This spread has increased significantly in many regions (see Exhibit GGGG). This growing
spread can be partially explained by the growing risk premiums authorized by regulators.
Exhibit GGGG: Gross Margin Trends in California, Pennsylvania, and Texas
If the growing risk premiums authorized cannot be explained as reasonable
additional compensation for additional risks borne by utilities, and if they are instead
explained by regulatory biases, then the consequence of these biases may be, in part,
responsible for billions of dollars in additional costs. These biases may be deemed
consequential enough so as to justify even costly (in a general sense) debiasing. If regulators
cannot be debiased, this research suggests that the public interest may be better served by
adopting a model that standardizes and “automates” the determination of rates of return for
0
2
4
6
8
10
12
2007 2008 2009 2010 2011 2012 2013 2014
Gro
ss M
argi
n (c
ents
/kw
h)
PA TX CA
+112%
–2%
+66%
7-5
utilities (as the Canadian model does). In other words, the public as a whole may be better
off by instantiating this part of the regulatory function in a fixed model and thus removing
it from the discretionary judgment of potentially fallible regulators.
A brief examination of the direct influence of regulatory posture can be seen by
including a “score” for public utilities commissions based on their reputation for energy
regulation. Although the analysis briefly described here and presented in Exhibit HHHH
below is only suggestive, it does seem to indicate that commission posture (or at least the
reputation of that posture) has an influence on authorized rate of return spreads (but virtually
no additional explanatory power as the Akaike (“AIC”) and Bayesian Information Criterion
(“BIC”) results are conflicting). To be sure, the “scoring” used here [Winegarden and Miles,
2014] is merely one subjective assessment for one static point in time and therefore, cannot
account for the trend in spreads over the 35 years studied. Here, the results suggest merely
that higher spreads are associated with “less accommodating” commissions—which is
something of a paradox.85 Further, to be sure, the authorized rate of return is merely one
lever a regulator has to influence earnings (varying the size of the allowed rate base is another
that is not addressed here).
I leave further exploration to future work, where several paths of research are
immediately obvious. Is this behavior unique to electricity regulation, or is it also present in
regulation of natural gas utilities or FERC regulation of transmission, or pipelines? Is this
behavior unique to the U.S. regulatory system, or does it manifest (perhaps in different form)
under other system as well? Looking forward, the behavior of regulators as interest rates
steadily renormalize will be of interest. Will authorized returns become “sticky” in the
opposite direction, as returns trend back over 10% with rising riskless rates?
85 On one hand, state regulatory risk is non-diversifiable for franchise utilities, and therefore, an additional risk premium may be warranted for such utilities. On the other hand, the perception of such commissions as less accommodating would seem to be at odds with their provision of higher risk premiums.
7-6
Baseline With PUC Score ln ln
, Constant -5.638**** -5.913**** (0.151) (0.209)
, Asset beta,ln -0.184**** -0.184**** (0.027) (0.027)
, Capital structure,ln 1 1 -0.487**** -0.449****
(0.116) (0.116)
, Market risk premium,ln -0.957**** -0.959**** (0.039) (0.039)
, State regulator score, ln Score 0.138* (0.072) R-squared 42.6% 42.9% Adjusted R-squared 42.4% 42.7% F statistic 345.3**** 260.1**** No. of observations 1,402 1,389 AIC 1,627.5 1,625.9 BIC 1,653.7 1,657.3
Standard errors are reported in parentheses. *, **, ***, and **** indicate significance at the 90%, 95%, 99%, and 99.9% levels, respectively. PUC scores were not provided for the District of Columbia, resulting in fewer data points.
Exhibit HHHH: Regression Results Examining Commission Reputation on Spreads
Additionally, although I have examined the link between theory-implied rates of
return and authorized rates of return, and then authorized rates of return and earned rates of
return, a third step is obvious: connecting earned rates of return to the stock performance of
regulated utilities. If there are excess returns in fact being provided and earned, who is
earning them? Either stockholders are realizing excess rates of return or utilities are
demonstrating expense-preference behavior and using regulatory munificence to counteract
the drag of expense inefficiency. Williamson [1963] and Edwards [1977] find evidence of
expense preference in regulated industries, but Awh and Primeaux, Jr. [1985] fail to find
similar evidence in a small sample of regulated electric utilities. Each of these studies,
however, is now dated and not reflective of the state of regulation prevailing during the
period covered by the dataset that I explore in Chapter 4.
7-7
Finally, the consequences of correcting any regulatory biases could also be explored.
If regulators are, in fact, providing utilities with excess returns, correction of that behavior
would reduce utility revenues and potentially alter their capital investment behavior. What
impact would that have on grid reliability, compliance with environmental regulations, and
trends toward either re-regulation or further de-regulation of the industry?
7.3 Toward a Computational Approach to Behavioral Decision Theory
Behavioral decision making represents a fertile field for application of the
computational methods developed in Chapter 5. Such approaches have not generally been
employed in this area, but several decades of extensive experimental work have provided a
large set of data on how human behavior can be characterized. Employing computational
methods expands the ability of behavioral decision making to extend beyond individual-by-
individual analysis of behavior of toward atomistic examination of populations. Likewise, it
allows for extending the nature of analysis beyond whether or not an effect exists, and toward
determining for whom it exists, and under what conditions.
Returning to an application discussed in Section 7.1, the design of financial securities
represents an attractive domain in which to apply computational methods. Many securities
design analyses estimate market share potential for a new security on the basis of the optimal
fraction of wealth that investors should allocate to it in equilibrium (based on mean-variance
preferences, etc.). While an obvious calculation for rational investors, it may provide neither
an upper nor lower bound in practice. Instead, investors—especially in financial innovations
designed with retail investors in mind—may over- or under-respond to such securities based
on the value and risks that they perceive such novel securities to have, as filtered through a
suite of potential biases. A computational platform for examining securities (both new and
existing) could be used both to estimate characteristics of an investor base, but also to
measure the impact of changes to design on the population of investors.
An additional path of research for the computational approach is the parallel conduct
of experiments in both virtual and real worlds as a means of separating intrinsic bias from
7-8
presentation (or experimental design) bias. Researchers are (or should be) scrupulously
attentive to the impact that their own behavior or the manner in which tasks are presented to
subjects have on outcomes. However, the basis for comparison in such circumstances is often
judgment based. There has not been a means, previously, by which to “remove” problem
presentation from problems presented to subjects. The application of computational methods
may provide such a means.
An additional path of exploration in this area would involve enriching the virtual
population’s diversity. Although the majority of studies appear to have converged toward
models that contain prospect theory-like and hyperbolic discounting-like elements, those
theoretical models are not universally accepted. I have demonstrated how the decision
demographics of a population can be estimated and employed under the assumption that
cumulative prospect theory and hyperbolic discounting are the models underlying individual
decision-making behavior. This approach could easily be extended, however, to account for
a more diverse population of decision makers that use other non-expected utility and non-
exponential discounting theories. Alternatively, the approach could also be narrowed to
explore subpopulations of decision makers where such analyses were appropriate (e.g.,
different forms of compensation for corporate managers may cause them to act more risk-
seeking than average [Jensen, 1986; MacCrimmon and Wehrung, 1990], corporate
accounting requirements that may induce elevated levels of loss aversion [Burgstahler and
Dichev, 1997]).
7.4 Presentation Compression in Risky Intertemporal Choice
The practice of finance has traditionally given scant attention to the decision-making
aspects of investment and corporate finance. Despite the reality that financial analysts and
managers are routinely called upon to evaluate complex choices and identify optimal
solutions, most (finance) researchers have been content simply to “wave their hands” at
assertions involving estimation of a particular performance or valuation metric and then its
application to a complex problem. Criticism of work of the type addressed in this dissertation
is commonly given along the lines of “people don’t actually behave like that in reality” and
7-9
“no one would look only at a single metric to make an investment decision.” And yet, a
classic article on managerial decision-making quoted finance executives as saying “[n]o one
is interested in getting quantified measures” and “[…] you don’t quantify the risk, but you
have to be able to feel it.” [March and Shapira [1987]
To be sure, corporate decision makers are called upon to be “risk neutral” when
evaluating corporate investments. All positive expected net present value investments should
be accepted. The reality, of course, is far more nuanced. However risk neutral that firms are
expected to be, the individuals that run them are human and generate very human levels of
risk-averse and risk-seeking behavior [Wiseman and Gomez-Mejia, 1998]. Admonitions to
“feel it” open the door to heuristic decision making, which would suggest that understanding
how the heuristics, and metrics on which they are based, are used and perceived should be
important.
Although some work has examined the role of behavior in corporate finance (e.g.,
Statman and Tyebjee [1985], Statman and Caldwell [1987]), it has mostly focused on
forecasting-related biases. This body of work has not gone far enough. However, to be fair,
the identification and solicitation of meaningful groups of human subjects for such analysis
that also have sufficient domain knowledge of corporate or investment finance is small. As
a result, the simulation approach developed in Chapter 5 and applied to a specific problem
in Chapter 6 shows more promise if the decision-making characteristics of the relevant
subpopulation can be modeled.
Levelized cost is not the only metric used in finance. Nor are heuristics limited to
investments in the power sector. The approach taken in Chapter 6 could be extended to a
wide variety of financial decision making, both at the corporate and individual level.
Financial analysts will commonly forecast cash flows only for a handful of years and then
use a “terminal multiple” to represent all value beyond that point in a project. Many
commercial property tax assessments are based on the “direct capitalization” approach,
meaning that an asset’s entire lifetime of operation is reduced to a single “representative”
year for assessment purposes. The role of compression, as defined and evaluated in Chapter
7-10
6, turns out to be fairly ubiquitous in finance. The notion that using a simplified metric that
is objectively identical to (or at least highly correlated with) a full presentation of a problem
has value in a world in which decision makers’ perceptions are not influenced by problem
presentation. But if the “lossy” compression employed in such processes alters the problem
presentation in a way that changes decision-makers’ perceptions, it becomes essential to
understand how, and why, and what to do about it.
R-1
R References
A Abdellaoui, M. Parameter-Free Elicitation of Utilities and Probability Weighting Functions.
Management Science 46 (2000): 1497–1512. Abdellaoui, M., H. Bleichrodt, and O. L’Haridon. A Tractable Method to Measure Utility
and Loss Aversion under Prospect Theory. Journal of Risk and Uncertainty 36 (2008): 245–266.
Abdellaoui, M., H. Bleichrodt, C. Paraschiv. Loss Aversion under Prospect Theory: A
Parameter-Free Measurement. Management Science 50 (2007): 1659–1674. Abdellaoui, M., F. Vossmann, and M. Weber. Choice-Based Elicitation and Decomposition
of Decision Weights for Gains and Losses under Uncertainty. Management Science 51 (2005): 1384–1399.
Abelson, R. Simulation and Social Behavior. In C. Lindzey and E. Aronson, eds., Handbook
of Social Psychology, Vol. 2 (Reading, MA: Addison-Wesley, 1968). Ainslie, G. Derivation of “Rational” Economic Behavior from Hyperbolic Discount Curves.
American Economic Review 81 (1991): 334-340. Akerlof, G. The Market for Lemons: Quality Uncertainty and the Market Mechanism.
Quarterly Journal of Economics 89 (1970). Alberici, S., S. Boeve, P. van Breevoort, Y. Deng, S. Förster, A. Gardiner, V. van Gastel, K.
Grave, H. Groenenberg, D. de Jager, E. Klaassen, W. Pouwels, M. Smith, E. de Visser, T. Winkel, and K. Wouters. Subsidies and Costs of EU Energy. DESNL 14583 (Utrecht, The Netherlands: Ecofys). Available at https://ec.europa.eu/energy/sites/ener/files/documents/ECOFYS%202014%20Subsidies%20and%20costs%20of%20EU%20energy_11_Nov.pdf (accessed April 17, 2017).
Albers, W., and J. Laing. Prominence, Competition, Learning, and the Generation of Offers
in Computer-Aided Experimental Spatial Games. In R. Selten, ed., Game Equilibrium Models III: Strategic Bargaining (Berlin: Springer-Verlag, 1991).
Albert, A., and J. Anderson. On the Existence of Maximum Likelihood Estimates in Logistic
Regression Models. Biometrika 71 (1984): 1-10.
R-2
Albert, B. Non-Businessmen as Surrogates for Businessmen in Behavioral Experiments. Journal of Business 40 (1967): 203-207.
Albrecht, M., and M. Weber. An Empirical Study on Intertemporal Decision Making under
Risk. Management Science 43 (1997): 813-826. Allen, F., and A. Santomero. The Theory of Financial Intermediation. Journal of Banking
and Finance 21 (1997): 1461-1485. Allen, F., and D. Gale. Financial Innovation and Risk Sharing (Cambridge, MA: MIT Press,
1994). Andersen, S., G. Harrison, M. Lau, and E. Rutström. Eliciting Risk and Time Preferences.
Econometrica 76 (2008): 583-618. Andersen, S., G. Harrison, and E. Rutström. (2006). Dynamic Choice Behavior: Asset
Integration and Natural Reference Points. Working paper 06-07, Department of Economics, College of Business Administration, University of Central Florida.
Anderson, L., and S. Stafford. Individual Decision-Making Experiments with Risk and
Intertemporal Choice. Journal of Risk and Uncertainty 38 (2009): 51-72. Anderson, J., P. Fischbeck, H. Zhai, and D. Rode. Clean Power Plan Insights for
Pennsylvania. Presented at the Center for Climate and Energy Decision Making Annual Meeting, Carnegie Mellon University (May 24, 2016).
Andreoni, J., and C. Sprenger. Estimating Time Preferences from Convex Budgets.
American Economic Review 102 (2012): 3333-3356. Arrow, K. Le rôle des valeurs boursières pour la répartition la meilleure des risques.
Colloques Internationaux du Centre National de la Recherche Scientifique 11 (1953): 41-47.
Arrow, K. The Role of Securities in the Optimal Allocation of Risk Bearing. Review of
Economic Studies 31 (1964): 91-96. Arrow, K., and G. Debreu. Existence of an Equilibrium for a Competitive Economy.
Econometrica 22 (1954): 265-290. Arthur, W., J. Holland, B. LeBaron, R. Palmer, and P. Tyler. Asset Pricing Under
Endogenous Expectations in an Artificial Stock Market. In W. Arthur, S. Durlauf, and D. Lane, eds., The Economy as an Evolving Complex System II (Reading, MA: Addison-Wesley, 1997).
R-3
ATSE. (2011). New Power Cost Comparisons: Levelised Cost of Electricity for a Range of New Power Generating Technologies. (Melbourne, Australia: The Australian Academy of Technological Sciences and Engineering). Available at http://www.garnautreview.org.au/update-2011/commissioned-work/new-power-cost-comparisons.pdf (accessed April 17, 2017).
Averch, H., and L. Johnson. Behavior of the Firm under Regulatory Constraint. American
Economic Review 52 (1962): 1052-1069. Awh, R., and W. Primeaux, Jr. Managerial Discretion and Expense Preference Behavior.
Review of Economics and Statistics 67 (1985): 224-231. Axtell, R. Theory of Model Aggregation for Dynamical Systems with Application to
Problems of Global Change. Unpublished doctoral dissertation, Carnegie Mellon University (September 1992).
B Babcock, L., and G. Loewenstein. Explaining Bargaining Impasse: The Role of Self-Serving
Biases. Journal of Economic Perspectives 11 (1997). Bantwal, V., and H. Kunreuther. A Cat Bond Premium Puzzle? Journal of Psychology and
Financial Markets 1 (2000). Bawa, V., and E. Lindenberg. Mean-Lower Partial Moments and Asset Prices. Journal of
Financial Economics 5 (1977). Becker, G. Nobel Lecture: The Economic Way of Looking at Things. Journal of Political
Economy 101 (1993): 385-409. Benhabib, J., A. Bisin, and A. Schotter. Present-Bias, Quasi-Hyperbolic Discounting, and
Fixed Costs. Games and Economic Behavior 69 (2010): 205-223. Black, F., and M. Scholes. The Pricing of Options and Corporate Liabilities. Journal of
Political Economy 81 (1973): 637-659. Blackburn, M., and W. El-Deredy. The Future is Risky: Discounting of Delayed and
Uncertain Outcomes. Behavioural Processes 94 (2013): 9-18. Bleichrodt, H., and J. Pinto. A Parameter-Free Elicitation of the Probability Weighting
Function in Medical Decision Analysis. Management Science 46 (2000): 1485–1496. Bleichrodt, H., J. Pinto, and P. Wakker. Using Descriptive Findings of Prospect Theory to
Improve the Prescriptive Use of Expected Utility. Management Science 47 (2001): 1498-1514.
R-4
Bodie, Z., A. Kane, and A. Marcus. Investments, Second Edition (Burr Ridge, IL: Irwin,
1993). Booij, A. and G. van de Kuilen. A Parameter-Free Analysis of the Utility of Money for the
General Population under Prospect Theory. Journal of Economic Psychology 30 (2009): 651-666.
Booij, A., B. van Praag, and G. van de Kuilen. A Parametric Analysis of Prospect Theory’s
Functionals for the General Population. Theory and Decision 68 (2010): 115-148. Borch, K. Economics of Insurance (New York, NY: North-Holland, 1990). Borch, K. The Objectives of an Insurance Company. Skandinavisk Aktuarietidskrift (1962). Borch, K. The Optimal Reinsurance Treaty. ASTIN Bulletin 5 (1969). Bourgeois-Gironde, S., and R. Giraud. Framing Effects as Violations of Extensionality.
Theory and Decision 67 (2009): 385-404. Boyes, W., and J. McDowell. The Selection of Public Utility Commissioners: A
Reexamination of the Importance of Institutional Setting. Public Choice 61 (1989): 1-13.
Brealey, R., and S. Myers. Principles of Corporate Finance, Seventh Edition (Boston, MA:
McGraw-Hill, 2003). Brown, D. Levelized Production Cost: An Alternative Form of Discounted Cash Flow
Analysis. Cost Engineering 36 (1994): 13-16. Bruckner T., I. Bashmakov, Y. Mulugetta, H. Chum, A. de la Vega Navarro, J. Edmonds,
A. Faaij, B. Fungtammasan, A. Garg, E. Hertwich, D. Honnery, D. Infield, M. Kainuma, S. Khennas, S. Kim, H. Nimir, K. Riahi, N. Strachan, R. Wiser, and X. Zhang. 2014: Energy Systems. In O. Edenhofer, R. Pichs-Madruga, Y. Sokona, E. Farahani, S. Kadner, K. Seyboth, A. Adler, I. Baum, S. Brunner, P. Eickemeier, B. Kriemann, J. Savolainen, S. Schlömer, C. von Stechow, T. Zwickel, and J. Minx, eds. Climate Change 2014: Mitigation of Climate Change. Contribution of Working Group III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (Cambridge: Cambridge University Press, 2014).
Buffett, W. (1997). Chairman’s Letter. Berkshire Hathaway 1997 Annual Report. Burgstahler, D., and I. Dichev. Earnings Management to Avoid Earnings Decreases and
Losses. Journal of Accounting and Economics 24 (1997): 99-126.
R-5
Butts, C. A Bayesian Model of Panic in Belief. Computational and Mathematical Organization Theory 4 (1998): 373-404.
C Cairns, J., and M. van der Pol. Valuing Future Private and Social Benefits: The Discounted
Utility Model versus Hyperbolic Discounting Models. Journal of Economic Psychology 21 (2000): 191-205.
California Legislative Analyst’s Office. California’s New Earthquake Program. California
Update (January 1997). Camerer, C., and T. Ho. Nonlinear Weighting of Probabilities and Violations of the
Betweenness Axiom. Journal of Risk and Uncertainty 8 (1994): 167–196. Camerer, C., G. Loewenstein, and M. Weber. The Curse of Knowledge in Economic
Settings: An Experimental Analysis. Journal of Political Economy 97 (1989): 1232-1254.
Carayannopoulos, P., and M. Perez. Diversification through Catastrophe Bonds: Lessons
from the Subprime Financial Crisis. Geneva Papers on Risk and Insurance 40 (2015): 1-28.
Carley, K., and M. Prietula. Computational Organization Theory (Hillsdale, NJ: Lawrence
Erlbaum, 1994). Carlin, B. Strategic Price Complexity in Retail Financial Markets. Journal of Financial
Economics 91 (2009): 278-287. Chetty, R., J. Friedman, S. Leth-Petersen, T. Nielsen, and T. Olsen. Active vs. Passive
Decisions and Crowd-Out in Retirement Savings Accounts: Evidence from Denmark. Quarterly Journal of Economics 129 (2014): 1141-1219.
Chicago Board of Trade, PCS Catastrophe Insurance Options Contract Specifications.
Available at http://www.cbot.com/ (accessed May 13, 1998). Chichilnisky, G., and G. Heal. Managing Unknown Risks. Journal of Portfolio Management
(Summer 1998). Chow, J., G. Lee, and I. Yang. Genetic Algorithm to Estimate Cumulative Prospect Theory
Parameters for Selection of High-Occupancy-Vehicle Lane. Transportation Research Record 2157 (2010): 71-77.
Cohen, R. Incorporating Default Risk into Hamada’s Equation for Application to Capital
Structure. Wilmott Magazine (March/April 2008): 62-68.
R-6
Combs, B., and P. Slovic. Newspaper Coverage of Causes of Death. Journalism Quarterly
56 (1979). Conine, Jr., T. Corporate Debt and Corporate Taxes: An Extension. Journal of Finance 35
(1980): 1033-1037. Coombs, C., and P. Lehner. Evaluation of Two Alternative Models for a Theory of Risk I:
Are Moments of Distributions Useful in Assessing Risk? Journal of Experimental Psychology: Human Perception and Performance 7 (1981).
Courville, L. Regulation and Efficiency in the Electric Utility Industry. The Bell Journal of
Economics and Management Science 5:1 (1974): 53-74. Cramer, D., and D. Howitt. The SAGE Dictionary of Statistics (Thousand Oaks: SAGE
Publications, 2004): 159. Cross, J. A Theory of Adaptive Economic Behavior (Cambridge: Cambridge University
Press, 1983). Cummins, J. CAT Bonds and Other Risk-Linked Securities: State of the Market and Recent
Developments. Risk Management and Insurance Review 11 (2008): 23-47. Cummins, J., and N. Doherty. (1997) Can Insurers Pay for the “Big One?” Measuring the
Capacity of an Insurance Market to Respond to Catastrophic Losses. Working Paper, The Wharton School, University of Pennsylvania.
D Danish Energy Agency. (2015). Levelized Cost of Energy Calculator. (Copenhagen,
Denmark: Danish Energy Agency). Available at https://ens.dk/sites/ens.dk/files/contents/material/file/introduction_lcoe_calculator.pdf (accessed April 17, 2017).
David, H., and H. Nagaraja. Order Statistics, Third Edition (Hoboken, NJ: John Wiley,
2003). Dawes, R. A Message from Psychologists to Economists: Mere Predictability Doesn’t
Matter Like it Should (Without a Good Story Appended to it). Journal of Economic Behavior and Organization 39 (1999).
Dawes, R. Rational Choice in an Uncertain World (San Diego, CA: Harcourt Brace, 1988). Dawes, R. The Robust Beauty of Improper Linear Models in Decision Making. American
Psychologist 34 (1979): 571-582.
R-7
Dawes, R., D. Faust, and P. Meehl. Clinical Versus Actuarial Judgment. Science 243 (1989). DeMarzo, P., and D. Duffie. Corporate Incentives for Hedging and Hedge Accounting.
Review of Financial Studies 8 (1995): 743-771. Dewing, A. A Study of Corporation Securities (New York: Ronald Press, 1934). Diamond, D. Financial Intermediation and Delegated Monitoring. Review of Economic
Studies 51 (1984): 393-414. Dong, W., H. Shah, and F. Wong. A Rational Approach to Pricing Catastrophe Insurance.
Journal of Risk and Uncertainty 12 (1996). Donkers, A., B. Melenberg, and A. van Soest. Estimating Risk Attitudes using Lotteries: A
Large Sample Approach. Journal of Risk and Uncertainty 22 (2001): 165–195. Duffie, D., and K. Singleton. Credit Risk: Pricing, Measurement, and Management
(Princeton: Princeton University Press, 2003). Duffie, D., and R. Rahi. Financial Market Innovation and Security Design: An Introduction.
Journal of Economic Theory 65 (1995).
E Economides, N. (1992). Liquidity and Markets. In the New Palgrave Dictionary of Finance. Economides, N. How to Enhance Market Liquidity. In R. Schwartz, ed., Global Equity
Markets (New York, NY: Irwin Professional, 1995). Economides, N. Network Economics with Application to Finance. Financial Markets,
Institutions and Instruments 2 (1993). Edwards, F. Managerial Objectives in Regulated Industries: Expense-Preference Behavior
in Banking. Journal of Political Economy 85 (1977): 147-162. EIA (2016). Levelized Cost and Levelized Avoided Cost of New Generation Resources in
the Annual Energy Outlook 2016. Energy Information Administration, U.S. Department of Energy (August 2016). Available at https://www.eia.gov/outlooks/aeo/pdf/electricity_generation.pdf (accessed April 17, 2017).
Embrechts, P. (1996). Actuarial Versus Financial Pricing of Insurance. Working Paper,
Wharton Financial Institutions Center, University of Pennsylvania.
R-8
Epstein, J., and R. Axtell. Growing Artificial Societies (Washington, DC: Brookings Institution Press, 1996).
Epstein, L., and S. Zin. Substitution, Risk Aversion, and the Temporal Behavior of
Consumption and Asset Returns: A Theoretical Framework. Econometrica 57 (1989): 937-969.
Epstein, R. Catastrophic Responses to Catastrophic Risks. Journal of Risk and Uncertainty
12 (1996). EQE, Inc. The 1994 Northridge Earthquake. EQE Summary Report (March, 1994). EQE, Inc. The January 17, 1995 Kobe Earthquake. EQE Summary Report (April, 1995). EQECAT. (1998) EQECAT’s USWINDTM - The Only Hurricane Model Certified by
Florida Commission for the Use of Cladding Type in Ratemaking. Company Press Release.
Etchart-Vincent, N. Is Probability Weighting Sensitive to the Magnitude of Consequences?
An Experimental Investigation on Losses. Journal of Risk and Uncertainty 28 (2004): 217–235.
F Fabozzi, F., and T. Fabozzi. The Handbook of Fixed Income Securities, Fourth Edition (Burr
Ridge, IL: Irwin, 1995). Fabozzi, F., and T. Fabozzi. Treasury and Agency Securities. In The Handbook of Fixed
Income Securities, 4th Ed., F. Fabozzi and T. Fabozzi, eds., (New York: Irwin, 1995). Fama, E. Efficient Capital Markets: A Review of Theory and Empirical Work. Journal of
Finance 25 (1970): 383-417. Fang, H., and Y. Wang. (2014). Estimating Dynamic Discrete Choice Models with
Hyperbolic Discounting, with an Application to Mammography Decisions. Working Paper, Department of Economics, University of Pennsylvania.
Fehr, E., and J. Tyran. Does Money Illusion Matter? American Economic Review 91 (2001):
1239-1262. Fehr-Duda, H., M. de Gennaro, and R. Schubert. Gender, Financial Risk, and Probability
Weights. Theory and Decision 60 (2006): 283–313. Fennema, H., and M. van Assen. Measuring the Utility of Losses by Means of the Tradeoff
Method. Journal of Risk and Uncertainty 17 (1998): 277–295.
R-9
Fernandez, P., A. Ortiz, and I. Acín. Discount Rate (Risk-Free Rate and Market Risk
Premium) Used for 41 Countries in 2015: A Survey. (November 19, 2015). Available at http://ssrn.com/abstract=2598104 (accessed June 8, 2016).
Financial Times. Music Bonds Are Failing to Prove Chart Toppers. March 2, 1999. Page 27. Firth, D. Bias Reduction of Maximum Likelihood Estimates. Biometrika 80 (1993): 27-38. Fischhoff, B. Debiasing. In D. Kahneman, P. Slovic, and A. Tversky, eds., Judgment Under
Uncertainty: Heuristics and Biases. (New York: Cambridge University Press, 1982): 422-444.
Fischhoff, B. Managing Risk Perceptions. Issues in Science and Technology 2 (1985). Fischhoff, B. Value Elicitation: Is There Anything in There? American Psychologist 46
(1991). Fischhoff, B. What Forecasts (Seem to) Mean. International Journal of Forecasting 10
(1994). Fischhoff, B., P. Slovic, and S. Lichtenstein. Fault Trees: Sensitivity of Estimated Failure
Probabilities to Problem Representation. Journal of Experimental Psychology: Human Perceptions and Performance 4 (1978).
Fischhoff, B., P. Slovic, S. Lichtenstein, S. Read, and B. Combs. How Safe if Safe Enough?
A Psychometric Study of Attitudes Towards Technological Risks and Benefits. Policy Sciences 9 (1978).
Fish, J. Engineering Economics, Second Edition (New York, NY: McGraw-Hill, 1923). Fishburn, P. Mean-Risk Analysis with Risk Associated with Below-Target Returns.
American Economic Review 67 (1977). Fishburn, P., and G. Kochenberger. Two-Piece von Neumann-Morgenstern Utility
Functions. Decision Sciences 10 (1979): 503-518. Fisher, I. The Money Illusion (New York, NY: Adelphi, 1928). Fisher, K., and M. Statman. The Mean-Variance-Optimization Puzzle: Security Portfolios
and Food Portfolios. Financial Analysts Journal (July/August, 1997). Florida Department of Insurance. Nelson Files Legal Challenge to Head Off Rate Hikes
Based on Computer Projections. Florida Department of Insurance Media Advisory, June 4, 1997.
R-10
Flyvbjerg, B., N. Bruzelius, and W. Rothengatter. Megaprojects and Risk (Cambridge: Cambridge University Press, 2003).
Frederick, S., G. Loewenstein, and T. O’Donoghue. Time Discounting and Time Preference:
A Critical Review. Journal of Economic Literature 40 (2002): 351-401. Froot, K. (1997). The Limited Financing of Catastrophe Risk: An Overview. NBER
Working Paper 6025. Funtowicz, S., and J. Ravetz. Qualified Quantities: Towards an Arithmetic of Real
Experience. In J. Forge, ed., Measurement, Realism and Objectivity (Dordrecht: D. Reidel, 1987): 59-88.
G Gaylord, R., and L. D’Andria. Simulating Society (New York, NY: Springer-Verlag, 1998). Gelman, A., A. Jakulin, M. Pittau, and Y. Su. A Weakly Informative Default Prior
Distribution for Logistic and Other Regression Models. Annals of Applied Statistics 2 (2008): 1360-1383.
Genton, M., Y. Ma, and E. Parzen. Discussion of “Sur une limitation très générale de la
dispersión de la médiane” by M. Fréchet, 1940. Journal de la Société Française de Statistique, 147 (2006): 51–60.
Gigerenzer, G., and P. Todd, eds. Simple Heuristics That Make Us Smart (New York:
Oxford University Press, 1999). Glöckner, A., and T. Pachur. Cognitive Models of Risky Choice: Parameter Stability and
Predictive Accuracy of Prospect Theory. Cognition 123 (2012): 21-32. Goodman, J., C. Cryder, and A. Cheema. Data Collection in a Flat World: The Strengths and
Weaknesses of Mechanical Turk Samples. Journal of Behavioral Decision Making 26 (2013): 213-224.
Gonzalez, R., and G. Wu. On the Form of the Probability Weighting Function. Cognitive
Psychology 38 (1999): 129–166. Grabowski, H., and J. Vernon. A New Look at the Returns and Risks to Pharmaceutical
R&D. Management Science 36 (1990): 804-821. Gray, W., C. Landsea, P. Mielke, Jr., and K. Berry. (1998) Extended Range Forecast of
Atlantic Seasonal Hurricane Activity and US Landfall Strike Probability for 1999. Department of Atmospheric Science, Colorado State University.
R-11
Grossman, S., and J. Stiglitz. On the Impossibility of Informationally Efficiency Markets. American Economic Review70 (1980): 393-408.
Gurley, J. and E. Shaw. Money in a Theory of Finance. (Washington, D.C.: The Brookings
Institution, 1960). Guy Carpenter & Company. (1997). The Emerging Asset Class: Insurance Risk.
Unpublished manuscript. New York, NY.
H Haag, H. Comments at the seminar ‘Funding Catastrophe Protection’. New York, NY.
Presented November, 1995. Hagerman, R., and B. Ratchford. Some Determinants of Allowed Rates of Return on Equity
to Electric Utilities. Bell Journal of Economics 9 (1978): 46-55. Hamada, R. Portfolio Analysis, Market Equilibrium and Corporation Finance. Journal of
Finance 24 (1969): 13-31. Hamada, R. The Effect of the Firm’s Capital Structure on the Systematic Risk of Common
Stocks. Journal of Finance 27 (1972): 435-452. Hardisty, D., and J. Pfeffer. Intertemporal Uncertainty Avoidance: When the Future is
Uncertain, People Prefer the Present, and When the Present is Uncertain, People Prefer the Future. Management Science 63 (2017): 519-527.
Harrison, G., and E. Rutström. Expected Utility Theory and Prospect Theory: One Wedding
and a Decent Funeral. Experimental Economics 12 (2009): 133-158. Harrison, G., and J. Swarthout. (2016). Cumulative Prospect Theory in the Laboratory: A
Reconsideration. Center for the Economic Analysis of Risk Working Paper 2, Robinson College of Business, Georgia State University.
Hayashi, P., and J. Trapani. Rate of Return Regulation and the Regulated Firm’s Choice of
Capital-Labor Ratio: Further Empirical Evidence on the Averch-Johnson Model. Southern Economic Journal 42:3 (1979): 384-398.
Helbling, C., G. Fallegger, and D. Hill. (1996). Rethinking Risk Financing. Swiss Re
manuscript. Henderson, B., and N. Pearson. The Dark Side of Financial Innovation: A Case Study of the
Pricing of a Retail Financial Product. Journal of Financial Economics 100 (2011): 227-247.
R-12
Hirshleifer, J. On the Theory of Optimal Investment Decision. Journal of Political Economy 66 (1958): 329-352.
Hirth, L., Ueckerdt, F., and O. Edenhofer. Integration Costs Revisited—An Economic
Framework for Wind and Solar Variability. Renewable Energy 74 (2015): 925-939. Hogan, W., and J. Warren. Toward the Development of an Equilibrium Capital-Market
Model Based on Semivariance. Journal of Financial and Quantitative Analysis 9 (1974).
Holthausen, D. A Risk-Return Model with Risk and Return Measured as Deviations from a
Target Return. American Economic Review 71 (1984). Hommel, U., and M. Ritter. New Approaches to Managing Catastrophic Insurance Risk. In
M. Frankel, U. Hommel, and M. Rudolf, eds., Risk Management: Challenge and Opportunity, 2nd Edition (Berlin: Springer, 2005).
Hopkinson, J. The Cost of Electricity Supply. Transactions of the Junior Engineering
Society 3 (1892). Hull, J. Options, Futures, and Other Derivative Securities, Second Edition (Englewood
Cliffs, NJ: Prentice-Hall, 1993).
I IEA. (2015). Projected Costs of Generating Electricity 2015 (Paris, France: International
Energy Agency). IndexCo, LLC. (1997). The GCCI Manual. Hartford, CT. Iyengar, S., and M. Lepper. Rethinking the Role of Choice: A Cultural Perspective on
Intrinsic Motivation. Journal of Personality and Social Psychology 76 (1999): 349-366.
Iyengar, S., and M. Lepper. When Choice is Demotivating: Can One Desire Too Much of a
Good Thing? Journal of Personality and Social Psychology 79 (2000): 995-1006.
J Jaffee, D., and T. Russell. Catastrophe Insurance, Capital Markets, and Uninsurable Risks.
Journal of Risk and Insurance 64 (1997).
R-13
JANRE. (2015). ⻑期エネルギー需給⾒通し⼩委員会に対する 発電コスト等の検証に関する報告. (Tokyo, Japan: Agency for Natural Resources and Energy, Electricity Generation Costs Verification Working Group). Available at http://www.enecho.meti.go.jp/committee/council/basic_policy_subcommittee/mitoshi/cost_wg/pdf/cost_wg_01.pdf (accessed April 17, 2017).
Jenkins, G. (1998). Modeling Climate on a Regional Scale. Presentation at the 1998 Annual
Meeting of the American Association for the Advancement of Science. Philadelphia, Pennsylvania.
Jenni, K. (1997). Attributes of Risk Decisions. Unpublished doctoral dissertation.
Department of Engineering and Public Policy, Carnegie Mellon University. Jensen, M. Agency Costs of Free Cash Flow, Corporate Finance and Takeovers. American
Economic Review 76 (1986): 323-329. Jewett, W. Converging Roles Within the Insurance and Finance Marketplace. Risk
Management (August 1997). Johnson. M., and W. Bickel. Within-Subject Comparison of Real and Hypothetical Money
Rewards in Delay Discounting. Journal of the Experimental Analysis of Behavior 77 (2002): 129-146.
Joskow, P. (2010). Comparing the Costs of Intermittent and Dispatchable Electricity
Generating Technologies. Working Paper 2010-013, Center for Energy and Environmental Policy Research, MIT.
Joskow, P. Comparing the Costs of Intermittent and Dispatchable Electricity Generating
Technologies. American Economic Review 101 (2011): 238-241.
K Kahn, A. The Economics of Regulation (New York: John Wiley, 1970). Kahneman, D., and S. Frederick. Representativeness Revisited: Attribute Substitution in
Intuitive Judgment. In T. Gilovich, D. Griffin, and D. Kahneman, eds., Heuristics of Intuitive Judgment: Extensions and Applications (New York, NY: Cambridge University Press, 2002).
Kahneman, D., and A. Tversky. Prospect Theory: An Analysis of Decision Under Risk.
Econometrica 47 (1979). Kahneman, D., J. Knetsch, and R. Thaler. Fairness as a Constraint on Profit Seeking
Entitlements in the Market. American Economic Review 76 (1986): 728-741.
R-14
Kakade, M. Economic Analysis of Power Plant Cost and Operation. Cost Engineering 31 (1989): 8-15.
Kammen, D., and D. Hassenzahl. Environment at Risk (Princeton, NJ: Princeton University
Press, 1999). Katten, R. The Chicken Little Syndrome or Catastrophe Modeling - An Agent’s Perspective.
CPCU Journal - Reinsurance Supplement (Summer, 1997). Keren, G., and P. Roelofsma. Immediacy and Certainty in Intertemporal Choice.
Organizational Behavior and Human Decision Processes 63 (1995): 287-297. Kibbee, L. The Public Policy of CAT Models: The ‘Black Box’ Syndrome and Beyond.
CPCU Journal - Reinsurance Supplement (Summer, 1997). Kimbrough, S., S. Moore, C. Pritchett, and C. Sherman. (1992). On DSS Support for Candle
Lighting Analysis. Transactions of DSS 1992: 118-135. Kimbrough, S., J. Oliver, and C. Pritchett. On Post-Evaluation Analysis: Candle-Lighting
and Surrogate Models. Interfaces 23 (1993): 17-28. Kirby, K., and N. Maraković. Modeling Myopic Decisions: Evidence for Hyperbolic Delay-
Discounting within Subjects and Amounts. Organizational Behavior and Human Decision Processes 64 (1995): 22-30.
Kirby, K., and N. Maracović. Delay-Discounting Probabilistic Rewards: Rates Decrease as
Amounts Increase. Psychonomic Bulletin & Review 3 (1996): 100-104. Klein, R. (1997). The Regulation of Catastrophe Insurance: An Initial Overview. Working
Paper, Wharton Catastrophic Risk Project. The Wharton School, University of Pennsylvania.
Kleinmuntz D, and D. Schkade. Information Displays and Decision Processes.
Psychological Science 4 (1993): 221-227. Kollman, K., J. Miller, and S. Page. Adaptive Parties in Spatial Elections. American Political
Science Review 86 (1992): 929-937. Koonce, L., M. McAnally, and M. Mercer. How do Investors Judge the Risk of Financial
Items? The Accounting Review 80 (2005): 221-241. Koopmans, T. Stationary Ordinal Utility and Impatience. Econometrica 28 (1960): 287-309. Kraus, A., and R. Litzenberger. Skewness Preference and the Valuation of Risk Assets.
Journal of Finance 31 (1976).
R-15
Kreps, D., and E. Porteus. Temporal Resolution of Uncertainty and Dynamic Choice Theory. Econometrica 46 (1978): 185-200.
Kroll, Y., H. Levy, and A. Rapoport. Experimental Tests of the Mean-Variance Model for
Portfolio Selection. Organizational Behavior and Human Decision Processes 42 (1988).
Kunreuther, H. Mitigating Disaster Losses Through Insurance. Journal of Risk and
Uncertainty 12 (1996). Kwoka, J. Governance Alternatives and Pricing in the U.S. Electric Power Industry. Journal
of Law, Economics and Organization 18 (2002): 278-294.
L Laibson, D., P. Maxted, A. Repetto, and J. Tobacman. (2017). Estimating Discount
Functions with Consumption Choices over the Lifecycle. Working paper, Harvard University.
Langer, E. The Illusion of Control. Journal of Personality and Social Psychology 32 (1975). Lattimore, P., J. Baker, and A. Witte. The Influence of Probability on Risky Choice: A
Parametric Examination. Journal of Economic Behavior and Organization 17 (1992): 377-400.
Lazard. (2016). Levelized Cost of Energy Analysis—Version 10.0. Available at
https://www.lazard.com/perspective/levelized-cost-of-energy-analysis-100/ (accessed April 17, 2017).
Lewis, C., and P. Davis. (1998). Capital Market Instruments for Financing Catastrophe Risk:
New Directions. Working Paper, Wharton Research Project on Catastrophic Risk. University of Pennsylvania.
Lichtenstein, S., and P. Slovic. Response-Induced Reversals of Preference in Gambling: An
Extended Replication in Las Vegas. Journal of Experimental Psychology 101 (1973): 16-20.
Lichtenstein, S., P. Slovic, B. Fischhoff, M. Layman, and B. Combs. Judged Frequency of
Lethal Events. Journal of Experimental Psychology: Human Learning and Memory 4 (1978).
Lintner, J. Security Prices, Risk, and Maximal Gains from Diversification. Journal of
Finance 20 (1965): 79-96.
R-16
Lo, A., H. Mamaysky, and J. Wang. Foundations of Technical Analysis: Computational Algorithms, Statistical Inference, and Empirical Implementation. Journal of Finance 55 (2000).
Loewenstein, G. Frames of Mind in Intertemporal Choice. Management Science 34 (1988):
200-214. Loewenstein, G., and D. Prelec. Anomalies in Intertemporal Choice: Evidence and an
Interpretation. Quarterly Journal of Economics 107 (1992): 573-597. Luenberger, D. Investment Science (New York, NY: Oxford University Press, 1998). Lutz, W. S. Scherbov, A. Prskawetz, M. Dworak, and G. Feichtinger. Population, Natural
Resources, and Food Security: Lessons from Comparing Full and Reduced-Form Models. Population and Development Review 28 suppl. (2002): 199-224.
M MacCrimmon, K., and D. Wehrung. Characteristics of Risk Taking Executives.
Management Science 36 (1990): 422-435. MacDonald, A., C. Clack, A. Alexander, A. Dunbar, J. Wilczak, and Y. Xie. Future Cost-
Competitive Electricity Systems and their Impact on CO2 Emissions. Nature Climate Change 6 (2016): 526-531.
MacGregor, D. Inferences about Product Risks: A Mental Modeling Approach to Evaluating
Warnings. Journal of Products Liability 12 (1989). Madden, G., A. Begotka, B. Raiff, and L. Kastern. Delay Discounting of Real and
Hypothetical Rewards. Experimental and Clinical Psychopharmacology 11 (2003): 139-145.
Malkiel, B. A Random Walk Down Wall Street (New York: W. W. Norton, 1973). Manzhos, S. On the Choice of the Discount Rate and Role of Financial Variables and
Physical Parameters in Estimating the Levelized Cost of Energy. International Journal of Financial Studies 1 (2013): 54-61.
March, J., and Z. Shapira. Managerial Perspectives on Risk and Risk Taking. Management
Science 33 (1987): 1404-1418. Markowitz, H. Portfolio Selection. Journal of Finance 7 (1952): 77-91.
R-17
Matthews, P., M. Sheffield, J. Andre, J. Lafayette, J. Roethel, and E. Dobkin. Insolvency: Will Historic Trends Return? Best’s Review - Property/Casualty Industry (March, 1999).
McKercher, T., L. Green, J. Myerson, T. Pickford, J. Hill, and S. Stout. A Comparison of
Four Models of Delay Discounting in Humans. Behavioural Processes 81 (2009): 256-259.
Mehra, R., and E. Prescott. The Equity Premium in Retrospect. In G. Constantinides, M.
Harris, and R. Stulz, eds., Handbook of the Economics of Finance (Amsterdam: Elsevier, 2003): 889-938.
Mehra, R., and E. Prescott. The Equity Premium: A Puzzle. Journal of Monetary Economics
15 (1985): 145-161. Mendelsohn, R., W. Nordhaus, and D. Shaw. The Impact of Global Warming on Agriculture:
A Ricardian Analysis. American Economic Review 84 (1994): 753-771. Merton, R. On the Application of the Continuous-time Theory of Finance to Financial
Intermediation and Insurance. Geneva Papers on Risk and Insurance Theory 14 (1989): 225-261.
Michael, A., P. Reasenberg, P. Stauffer, and J. Hendley II. Quake Forecasting - An Emerging
Capability. U.S. Geological Survey Fact Sheet (1996). Miller, J. Active Nonlinear Tests (ANTs) of Complex Simulation Models. Management
Science 44 (1998): 820-830. Miller, J., and S. Page. Complex Adaptive Systems: An Introduction to Computational
Models of Social Life (Princeton, NJ: Princeton University Press, 2007). Moody’s Industry Review: Property and Casualty Insurers. March 21, 1997. Morin, R. New Regulatory Finance (Vienna, VA: Public Utilities Reports, 2006). Mossin, J. Equilibrium in a Capital Asset Market. Econometrica 34 (1966): 768-783. Mossin, J. Optimal Multiperiod Portfolio Policies. Journal of Business 41 (1968). Mowen, J. Judgment Calls (New York, NY: Simon & Schuster, 1994). Murphy, R., and R. ten Brincke. (2017). Hierarchical Maximum Likelihood Parameter
Estimation for Cumulative Prospect Theory: Improving the Reliability of Individual Risk Parameter Estimates. Management Science, Forthcoming.
R-18
Myerson, J., and L. Green. Discounting of Delayed Rewards: Models of Individual Choice. Journal of the Experimental Analysis of Behavior 64 (1995): 263-276.
N Nachman, D., and T. Noe. Optimal Design of Securities Under Asymmetric Information.
Review of Financial Studies 7 (1994). Neftci, S. Naive Trading Rules in Financial Markets and Wiener-Kolmogorov Prediction
Theory: A Study of “Technical Analysis.” Journal of Business 64 (1991). Newell, A., and H. Simon. Human Problem Solving (Englewood Cliffs, NJ: Prentice-Hall,
1972). Nilsson, H., J. Rieskamp, and E. Wagenmakers. Hierarchical Bayesian Parameter Estimation
for Cumulative Prospect Theory. Journal of Mathematical Psychology 55 (2011): 84-93.
Nitsch, J., T. Pregger, T. Naegler, D. Heide, D. Luca de Tena, F. Trieb, Y. Scholz, K.
Nienhaus, N. Gerhardt, M. Sterner, T. Trost, A. von Oehsen, R. Schwinn, C. Pape, H. Hahn, M. Wickert, B. Wenzel. (2012). Langfristszenarien und Strategien für den Ausbau der Erneuerbaren Energien in Deutschland bei Berücksichtigung der Entwicklung in Europa und Global. (Stuttgart, Germany: Federal Ministry for the Environment, Nature Conservation and Nuclear Safety). Available at http://www.dlr.de/dlr/Portaldata/1/Resources/bilder/portal/portal_2012_1/leitstudie2011_bf.pdf (accessed April 17, 2017).
O Odean, T. Do Investors Trade Too Much? American Economic Review 89 (1999): 1279-
1298. Olsen, R. Behavioral Finance and Its Implications for Stock-Price Volatility. Financial
Analysts Journal (March/April, 1998). Olson, M. The Logic of Collective Action (Cambridge, MA: Harvard University Press,
1965). Öncüler, A. (1999). Intertemporal Choice under Uncertainty: A Behavioral Perspective.
Unpublished doctoral dissertation, Graduate Group in Managerial Science and Applied Economics, University of Pennsylvania.
R-19
Öncüler, A., and S. Onay. How Do We Evaluate Future Gambles? Experimental Evidence on Path Dependency in Risky Intertemporal Choice. Journal of Behavioral Decision Making 22 (2009): 280-300.
Oren, S., S. Smith, and R. Wilson. Capacity Pricing. Econometrica 53 (1985): 545-566. Osterland, A. The CATs Are Out of the Bag. Business Week (January 26, 1998). Ostrom, T. Computer Simulation: The Third Symbol System. Journal of Experimental
Social Psychology 24 (1988): 381-392.
P Palm, R. Earthquake Insurance: A Longitudinal Study of California Homeowners (Boulder,
CO: Westview Press, 1995). Palm, R., and M. Hodgson. After a California Earthquake: Attitude and Behavior Change.
University of Chicago Geography Research Paper #223. (Chicago, IL: University of Chicago Press, 1992).
Park, C., and G. Sharp-Bette. Advanced Engineering Economics (New York, NY: John
Wiley, 1990). Paserman, M. Job Search and Hyperbolic Discounting: Structural Estimation and Policy
Evaluation. The Economic Journal 118 (2008): 1418-1452. Payne, J. Alternative Approaches to Decision Making Under Risk: Moments Versus Risk
Dimensions. Psychological Bulletin 80 (1973). Penalva-Zuasti, J. (1997). The Theory of Financial Insurance with an Application to
Earthquakes and Catastrophe Bonds. Unpublished doctoral dissertation, University of California, Los Angeles.
Pennings, J., and A. Smidts. The Shape of Utility Functions and Organizational Behavior.
Management Science 49 (2003): 1251-1263. Pennsylvania Public Utility Commission. Bureau of Technical Utility Services Report on
the Quarterly Earnings of Jurisdictional Utilities for the Year Ended June 30, 2016. Docket M-2016-2570997. Available at http://www.puc.state.pa.us/pcdocs/ 1483248.docx (accessed February 21, 2017).
Peterson, R. On the Use of College Students in Social Science Research. Insights from a
Second-Order Meta-analysis. Journal of Consumer Research 28 (2001): 450-461.
R-20
Pettway, R., and B. Jordan. APT vs. CAPM Estimates of the Return-Generating Function Parameters for Regulated Public Utilities. Journal of Financial Research 10:3 (1987): 227-238.
Phillips, C., Jr. The Regulation of Public Utilities (Arlington: Public Utilities Reports, 1993):
394-400. Pitcher, C. (2008). Investigation of a Behavioral Model for Financial Decision Making.
Thesis in Mathematical and Computational Finance, Magdalen College, University of Oxford.
PJM Interconnection. Resource Investment in Competitive Markets. (May 5, 2016).
Available at http://pjm.com/~/media/documents/reports/20160505-resource-investment-in-competitive-markets-paper.ashx (accessed June 8, 2016).
Polya, G. How to Solve It, 2nd Ed. (Princeton: Princeton University Press, 1957). Poulton, E. Bias in Quantifying Judgment (Hillsdale, NJ: Lawrence Erlbaum, 1989). Pratt, J. Risk Aversion in the Small and in the Large. Econometrica 32 (1964): 122-136. Prelec, D. (1989). Decreasing Impatience: Definition and Consequences. Harvard Business
School Working Paper 90-015. Prelec, D., and G. Loewenstein. Decision Making over Time and under Uncertainty: A
Common Approach. Management Science 37 (1991): 770-786. Property and Casualty Online. (1998) E. W. Blanch’s Hurricane Model Certified by Florida
Commission. News for Risk Management. Available at https://www.propertyandcasualty.com/doc/e-w-blanch39s-hurricane-model-certified-by-fl-0001 (accessed April 17, 2017).
Q Quiggin, J. A Theory of Anticipated Utility. Journal of Economic Behavior and
Organization 3 (1982): 323-343.
R Ray, R. Catastrophe Derivatives. Risk Management (October 1993). Read, D. Is Time-Discounting Hyperbolic or Subadditive? Journal of Risk and Uncertainty
23 (2001): 5-32.
R-21
Read, D. (2003). Intertemporal Choice. London School of Economics Working Paper 03.58. Redelmeier, D., E. Shafir, and P. Aujla. The Beguiling Pursuit of More Information. Medical
Decision Making 21 (2001): 374-379. Reichelstein, S., and A. Rohlfing-Bastian. Levelized Product Cost: Concept and Decision
Relevance. Accounting Review 90 (2015): 1653-1682. Richards, C. Investing in Catastrophe Bonds. Money Management (March 2016). Rieger, M., M. Wang, and T. Hens. Estimating Cumulative Prospect Theory Parameters
from an International Survey. Theory and Decision 82 (2017): 567-596. Rode, D., P. Fischbeck, and A. Páez. The Retirement Cliff: Power Plant Lives and their
Policy Implications. Energy Policy 106 (2017): 222-232. Roll, R., and S. Ross. Regulation, the Capital Asset Pricing Model, and the Arbitrage Pricing
Model. Public Utilities Fortnightly (May 26, 1983): 22-28. Ross, S. The Arbitrage Theory of Capital Asset Pricing. Journal of Economic Theory 13
(1976): 341-360. Ross, S. The Determination of Financial Structure: The Incentive-signaling Approach. Bell
Journal of Economics 8 (1977): 23-40. Rugg, D. Experiments in Wording Questions: II. Public Opinion Quarterly 5 (1941): 91-92.
S Salmon, F. Recipe for Disaster: The Formula that Killed Wall Street. Wired (February 2009). Samuelson, P. A Note on the Measurement of Utility. Review of Economic Studies 4 (1937):
155-161. Savage, L. The Foundations of Statistics (New York, NY: Wiley, 1954). Schelling, T. Micromotives and Macrobehavior (New York, NY: W. W. Norton, 1978). Schelling, T. Economic Reasoning and the Ethics of Policy. The Public Interest 63 (1981):
37-61. Schmidt, U., and S. Traub. An Experimental Test of Loss Aversion. Journal of Risk and
Uncertainty 25 (2002): 233–249.
R-22
Scholes, M. Global Financial Risks, Derivative Securities, and Systematic Risks. Journal of Risk and Uncertainty 12 (1996).
Schumpeter, J. Business Cycles (New York: McGraw-Hill, 1939): p. 116. Schwager, J. The New Market Wizards (New York, NY: Harper Business, 1992). Schwartz, B. The Paradox of Choice (New York, NY: Harper Collins, 2004). Schwarz, N. Self-Reports: How the Questions Shape the Answers. American Psychologist
54 (1999). Scism, L., and A. Das. The Insurance Industry has been Turned Upside Down by Catastrophe
Bonds. Wall Street Journal (August 7, 2016). Scitovsky, T. Ignorance as a Source of Oligopoly Power. American Economic Review 40
(1950): 48-53. Sears, R., and G. Trennepohl. Investment Management (Fort Worth, TX: Dryden Press,
1993). Shafir, E., P. Diamond, and A. Tversky. Money Illusion. Quarterly Journal of Economics
112 (1997): 341-374. Shanteau, J. Psychological Characteristics and Strategies of Expert Decision Makers. Acta
Psychologica 68 (1988): 203-215. Shapiro, A. International Corporate Finance, Second Edition (Cambridge, MA: Ballinger,
1988). Sharpe, W. Capital Asset Prices: A Theory of Market Equilibrium Under Conditions of Risk.
Journal of Finance 19 (1964): 425-442. Shefrin, H. Beyond Greed and Fear (Boston: Harvard Business School Press, 2000): 127. Shelley, M. Gain/Loss Asymmetry in Risky Intertemporal Choice. Organizational Behavior
and Human Decision Processes 59 (1994): 124-159. Shuptrine, F. On the Validity of Using Students as Subjects in Consumer Behavior
Investigations. Journal of Business 48 (1975): 383-390. Silverman, G., D. Sparks, and A. Osterland. A $2.5 Trillion Market You Hardly Know.
Business Week (October 26, 1998). Simon, H. Administrative Behavior (New York: Macmillan, 1947).
R-23
Simon, H. A Behavioral Model of Rational Choice. Quarterly Journal of Economics 59 (1955): 99-118.
Simon, H. Rational Choice and the Structure of the Environment. Psychological Review 63
(1956): 129-138. Simon, H. Rational Decision Making in Business Organizations. American Economic
Review 69 (1979): 493-513. Simon, H., G. Dantzig, R. Hogarth, C. Plott, H. Raiffa, T. Schelling, K. Shepsle, R. Thaler,
A. Tversky, and S. Winter. Decision Making and Problem Solving. Interfaces 17 (1987): 11-31.
Slovic, P. Assessment of Risk Taking Behavior. Psychological Bulletin 61 (1964). Slovic, P. Perception of Risk. Science 236 (1987). Slovic, P. Psychological Study of Human Judgment: Implications for Investment Decision
Making. Journal of Finance 27 (1972). Slovic, P., S. Lichtenstein, and B. Fischhoff. Modeling the Societal Impact of Fatal
Accidents. Management Science 30 (1984). Smith, A. The Wealth of Nations (Edinburgh, Great Britain: 1776). Sobol’, I. On the Distribution of Points in a Cube and the Approximate Evaluation of
Integrals. U.S.S.R. Computational Mathematics and Mathematical Physics 7 (1967): 86-112.
Sortino, F., and R. van der Meer. Downside Risk. Journal of Portfolio Management.
(Summer 1991). Sovacool, B. The Problem with the “Portfolio Approach” in American Energy Policy. Policy
Sciences 41 (2008): 245-261. Spann, R. Rate of Return Regulation and Efficiency in Production: An Empirical Test of the
Averch-Johnson Thesis. The Bell Journal of Economics and Management Science 5:1 (1974): 38-52.
Standard & Poor’s. Property and Casualty Insurers. Industry Surveys, (June 5, 1997). Starmer, C. Developments in Non-Expected Utility Theory: The Hunt for a Descriptive
Theory of Choice under Risk. Journal of Economic Literature 38 (2000): 332-382. Stasser, G. Computer Simulation as a Research Tool: The DISCUSS Model of Group
Decision Making. Journal of Experimental Social Psychology 24 (1988): 393-422.
R-24
Statman, M., and T. Tyebjee. Optimistic Capital Budgeting Forecasts: An Experiment.
Financial Management 14 (Autumn 1985): 27:33. Statman, M., and D. Caldwell. Applying Behavioral Finance to Capital Budgeting: Project
Termination. Financial Management 16 (Winter 1987): 7-15. Stigler, G. The Theory of Economic Regulation. Bell Journal of Economics and
Management Science 2 (1971): 3-21. Stulz, R. Optimal Hedging Policies. Journal of Financial and Quantitative Analysis 19
(1984): 127- 140. Svedsäter, H., A. Gamble, and T. Gärling. Money Illusion in Intuitive Financial Judgments:
Influences of Nominal Representation of Share Prices. Journal of Socio-Economics 36 (2007): 698-712.
Swiss Re. Climatic Change and the Insurance Industry: Still a Critical Issue. Seminar on
Climatic Change and Natural Catastrophes. National Research Programme 31 (1996).
Swiss Re. sigma, Issue 2 (1996); Issue 3 (1997); Issue 3 (1998); Issue 5 (1996); Issue 7
(1997).
T Takeuchi, K. Non-Parametric Test of Time Consistency: Present Bias and Future Bias.
Games and Economic Behavior 71 (2011): 456-478. Tanaka, T., C. Camerer, and Q. Nguyen. Risk and Time Preferences: Linking Experimental
and Household Survey Data from Vietnam. American Economic Review 100 (2010): 557-571.
Tapon, F., and J. van der Weide. Effectiveness of Regulation in the Electric Utility Industry.
Journal of Economics and Business 31:3 (1979): 180-189. Thaler R. Some Empirical Evidence on Dynamic Inconsistency. Economics Letters 8 (1981):
201-207. Thaler, R. The Psychology and Economics Conference Handbook. In R. Thaler, ed.,
Quasirational Economics (New York, NY: Russell Sage, 1994). Thaler, R., and A. Tversky. Myopic Loss Aversion in Financial Investment: An
Experimental Study. Working paper, University of Chicago, 1996.
R-25
Thomas, B. (1997). Homogenizing Catastrophic Risk: An Overview of Catastrophe Indices. Unpublished manuscript.
Tobin, J. Liquidity Preference as Behavior Towards Risk. Review of Economic Studies 25
(1958): 65-86. Tu, Q. (2005). Empirical Analysis of Time Preferences and Risk Aversion. CentER PhD
Thesis 142, Tilburg University. Tucson Electric Power. (March 1, 2016). 2016 Preliminary Integrated Resource Plan.
Available at https://www.tep.com/doc/planning/2016-TEP-IRP.pdf (accessed September 26, 2016).
Turner, C., and E. Martin, eds. Surveying Subjective Phenomena (New York, NY: Russell
Sage Foundation, 1984). Tversky, A., and D. Kahneman. Availability: A Heuristic for Judging Frequency and
Probability. Cognitive Psychology 5 (1973). Tversky, A., and D. Kahneman. The Framing of Decisions and the Psychology of Choice.
Science 211 (1981). Tversky, A., and D. Kahneman. Advances in Prospect Theory: Cumulative Representation
of Uncertainty. Journal of Risk and Uncertainty 5 (1992): 297-323.
U Uecherdt, F., L. Hirth, G. Luderer, and O. Edenhofer. System LCOE: What are the Costs of
Variable Renewables? Energy 63 (2013): 61-75. U.S. Geological Survey. (1996). USGS Response to an Urban Earthquake: Northridge ‘94.
USGS Open-File Report #96-263. USAA. (1998). Technical Description of Proposal to Allow Property Casualty Insurers to
Establish Tax-Deferred Pre-Event Catastrophe Reserves. Draft monograph. USAID. (2009). Best Practices Guide: Integrated Resource Planning for Electricity. United
States Agency for International Development. Available at http://pdf.usaid.gov/pdf_docs/ PNACQ960.pdf (accessed September 26, 2016).
V Valery, N. Fear of Trembling. The Economist (April 22, 1995).
R-26
van der Pol, M., and J. Cairns. A Comparison of the Discounted Utility Model and Hyperbolic Discounting Models in the Case of Social and Private Intertemporal Preferences for Health. Journal of Economic Behavior and Organization 49 (2002): 79-96.
Velosa, A., III, and M. Aboudi. (2016). What is Inside Your LCOE Assumptions? Mercom
Capital Group. Available at http://mercomcapital.com/what-is-inside-your-lcoe-assumptions (accessed April 17, 2017).
Villadsen, B., and T. Brown. International Innovations in Rate of Return Determination.
Presentation at the 2012 Society of Utility and Regulatory Financial Analysts Forum. April 26, 2012.
von Neumann, J., and O. Morgenstern. Theory of Games and Economic Behavior, 2nd Ed.
(Princeton, NJ: Princeton University Press, 1947).
W Wall Street Journal. Hurricane Poses Risks for Bonds. September 15, 1999. Page C1. Wall Street Journal. Investment Bankers Are Moving Fast to Offer Securities Backed by
Insurance Policies. June 16, 1998. Page C2. Wall Street Journal. Investors’ Fear of Risk Translates into a Curb on Credit for Business.
October 7, 1998. Page A1. Wang, M., M. Rieger, and T. Hens. How Time Preferences Differ: Evidence from 53
Countries. Journal of Economic Psychology 52 (2016): 115-135. Warner, J. Bankruptcy Costs: Some Evidence. Journal of Finance 32 (1977): 337-347. Weber, B., and G. Chapman. The Combined Effects of Risk and Time on Choice: Does
Uncertainty Eliminate the Immediacy Effect? Does Delay Eliminate the Certainty Effect? Organizational Behavior and Human Decision Processes 96 (2005): 104-118.
Webster, M., M. Babiker, M. Mayer, J. Reilly, J. Harnisch, R. Hyman, M. Sarofim, and C.
Wang. Uncertainty in Emissions Projections for Climate Models. Atmospheric Environment 36 (2002): 3659-3670.
Wehrly, D., and A. Friedheim. (1998) The Property/Casualty Insurance Industry. Wheat
First Union Equity Research Department. Wharton Alumni Magazine. Catastrophe and the Capital Markets. (Summer, 1998).
R-27
Williamson, O. Managerial Discretion and Business Behavior. American Economic Review 53 (1963): 1032-1057.
Winegarden, W., and M. Miles. (July 2014). The 50 State Index of Energy Regulation.
Pacific Research Institute. Available at https://www.pacificresearch.org/fileadmin/documents/Studies/PDFs/2013-
2015/50StatesEnergy_FirstFinalWeb.pdf (accessed April 11, 2017). Wiseman, R., and L. Gomez-Mejia. A Behavioral Agency Model of Managerial Risk
Taking. Academy of Management Review 23 (1998): 133-153. Wu, G., and R. Gonzalez. Curvature of the Probability Weighting Function. Management
Science 42 (1996): 1676–1690.
X
Y Yates, J., and E. Stone. Risk Appraisal. In J. Yates, ed., Risk-Taking Behavior (New York,
NY: Wiley, 1992). Yoon, A., and L. Scism. Yield Hunt’s Latest Twist: ‘Catastrophe Bonds’. Wall Street
Journal (April 24, 2014).
Z Zebiak, S., and M. Cane. (1998). IRI Climate Information Digest. Lamont-Doherty Earth
Observatory, Columbia University. Zeckhauser, R. The Economics of Catastrophes. Journal of Risk and Uncertainty 12 (1996).