HETEROGENEITYIN RISK PREFERENCES:
EVIDENCEFROMA REAL-WORLD BETTING MARKET∗
Angie Andrikogiannopoulou Filippos Papakonstantinou King’s College London King’s College London
Abstract
We develop a structural model of behavior that accounts for rich heterogeneity in individuals’ risk pref- erences, and we estimate it using a panel dataset of individual activity in a sports wagering market. The number and variety of lottery choices we observe enables us to distinguish the key features of prospect theory — value-function curvature, loss aversion, and probability weighting. Our Bayesian hierarchical mixture model enables us to estimate the proportion of distinct utility ‘types’ characterized by the different prospect theory features, and thus to evaluate their relative prevalence in the population. We find that utility curvature alone does not explain the choices we observe and that, while loss aversion is important, probability weighting is the most prevalent behavioral feature of risk attitudes: Two-thirds of individuals exhibit loss aversion, but all exhibit probability weighting.
Keywords: Risk Preferences, Prospect Theory, Loss Aversion, Probability Weighting, Discrete Choice, Mixture Model, Hierarchical Model, Bayesian Estimation
∗For helpful comments, we thank participants at the 2014 European Economic Association Annual Congress and the 2018 Center for Economic Analysis of Research Workshop, and seminar participants at Cambridge University, Einaudi Institute for Economics and Finance, Imperial College London Business School, McGill University, Oxford University, Princeton University, the Stockholm School of Economics, the University of Amsterdam, the University of Geneva, and the Wharton School. Send correspondence to Filippos Papakonstantinou, King’s Business School, Bush House, 30 Aldwych, London WC2B 4BG, UK; telephone: +44 (0) 20 7848 3770. E-mail: [email protected]. Since Kahneman and Tversky (1979) proposed prospect theory, a plethora of experimental and field studies have shown that individual behavior deviates from the predictions of expected utility theory and exhibits patterns consistent with loss aversion and probability weighting. But even though prospect theory has become the most widely accepted behavioral theory of choice, the limited availability of rich data, particularly in the field, and modeling limitations have made it difficult to distinguish between the different prospect theory features and assess their relative importance and prevalence in the population. Hence, an important question remains unresolved: which of prospect theory’s features are most useful in explaining behavior? In this paper, we aim to address this question in the field by combining (1) a panel dataset that tracks individuals’ choices in a sports wagering market with (2) a hierarchical structural model that allows for an extensive representation of individual heterogeneity within and across utility types characterized by the different prospect theory features, and (3) a Bayesian algorithm designed to estimate this model.
The sports wagering market is an advantageous natural laboratory for risk preference estimation. First, it features choices whose outcomes (i) are determined exogenously by match outcomes, and (ii) are associated with probabilities that are predicted quite accurately by the odds quoted in the market. These features allow for a lottery representation of choices and hence facilitate preference estimation. Second, this market features a variety of lotteries consisting of a wide range of probabilities and prizes, including gains and losses as well as small and large stakes, which enables us to identify and comprehensively study all features of prospect theory. Third, in this setting, individuals typically make over time a large number of choices, which enables us to estimate preference heterogeneity across individuals with improved accuracy.
To estimate preference heterogeneity across individuals, extant studies have followed two main ap- proaches. Most commonly, each individual’s preferences are estimated in isolation; but this may lead to inaccurate estimates, especially when only a few choices per individual are observed. As a result, some recent studies have combined individual information with information from the population distribution, which is estimated under specific distributional assumptions (e.g., discrete or normal). This approach allows for information sharing across individuals, but in a restricted manner. For example, it is inappropriate to assume a normal population distribution if individuals cluster into distinct sub-populations. In this paper, we propose a hierarchical mixture model that generalizes previous methodologies. First, we model the population as a flexible mixture of four sub-populations: one type that exhibits neither probability weighting nor loss aversion, two types that each turn on one of these features, and an unrestricted type that exhibits both features.
This approach introduces the proportions of each type as model parameters, which are jointly estimated
1 with the entire distribution of preference parameters, and allows us to assess the relative importance of the
prospect-theory features. Second, we allow for preference heterogeneity within each sub-population. Thus,
we forgo the restrictive assumption that there is a representative agent for each type, which would not only
understate the heterogeneity but would also bias the preference estimates (and the type allocations). Overall,
our methodology enables us to share information across individuals in a flexible manner dictated by the data
and improves inference at both the individual and population level, as what we learn about each individual’s
preferences and type allocations feeds into the shape of the population distribution, and vice versa. To
estimate our model, we design and implement an appropriate Bayesian Markov chain Monte Carlo algorithm.1
Our model of behavior is based on the cumulative prospect theory (CPT) of Tversky and Kahneman
(1992), according to which individuals use a value function defined over gains and losses relative to a reference
point, have different sensitivity to losses and gains, and systematically distort event probabilities by weighting
the gain/loss cumulative distribution. Our average preference parameter estimates are within the range of
those reported in the experimental literature (see Wakker, 2010): the value function curvature is moderate,
loss aversion is mild, and individuals significantly overweight extreme positive and negative outcomes.
Furthermore, we find that utility curvature alone does not adequately explain individual choices, and that
the distinctive features of prospect theory — loss aversion and probability weighting — offer a significant
improvement in explaining observed behavior. But, importantly, our study reveals that these two features
are not equally prevalent. We find that individuals can essentially be partitioned into two groups: about
two-thirds exhibit both behavioral features, and one-third exhibit probability weighting but not loss aversion.
These findings indicate that prospect theory should not be viewed as a monolithic theory with a set of features
— loss aversion and probability weighting — that all individuals exhibit, but rather reality is more nuanced.
This more nuanced view could facilitate and guide the use of prospect theory in applied models. For example,
our finding that probability weighting plays a more important role than loss aversion could be useful in
finance, where loss aversion has sometimes led to counterfactual predictions for individual investor behavior
and for the aggregate stock market (see Barberis, 2013 for a review). Indeed, in an application we show
that our preference estimates make realistic predictions when applied to a standard portfolio choice problem,
as they yield optimal portfolios similar to those observed in surveys on household portfolio allocations.
We conduct a number of robustness checks, necessitated by the fact that we study a field setting hence we
1This estimation solves for an equilibrium density — the joint of the model parameters — which is roughly a stochastic analogue of a fixed point. Our estimation uses a Gibbs sampler, which is the stochastic analogue of an algorithm that solves for a fixed point.
2 do not have the benefit of experimental control. Regarding individuals’ beliefs, our baseline is the standard assumption in the literature that they are rational which, in our setting, can be approximated quite well by the probabilities implied by the market prices. In a sensitivity analysis, we show that our results are robust to al- ternative approximations of the rational beliefs, as well as to small but significant systematic deviations of the subjective from the rational beliefs. We also show that observed bets are unlikely to be driven by (real or per- ceived) superior information or affinity toward specific teams (e.g., home-area teams), since almost all individ- uals place wagers on a large variety of leagues and teams rather than any specific one, and virtually no individ- uals generate significantly positive returns from betting. Regarding the choice sets individuals face, our base- line assumption is that individuals consider all lotteries available in the sportsbook, except for those that have rarely been chosen by any individual in our sample. In a robustness check, we consider an alternative choice set containing only lotteries that respect a tight, but reasonable, budget constraint of a few hundred euros.
The majority of evidence on estimating risk preferences comes from lab experiments, surveys, and semi-experimental data from TV game shows.2 A small number of studies have estimated preferences in the
field using real decisions. Most of these have used aggregate prices to infer representative-agent preferences: for example, Golec and Tamarkin (1998), Jullien and Salanie (2000), and Snowberg and Wolfers (2010) use horse-racetrack betting prices; Chetty (2006) uses labor supply data; and Kliger and Levy (2009) and
Polkovnichenko and Zhao (2013) use financial market prices.3 Only a handful of studies have used individual- level data, hence are able to relax the representative-agent assumption and estimate heterogeneous preferences:
Cicchetti and Dubin (1994), Cohen and Einav (2007), and Barseghyan et al. (2013) study deductible choices in insurance markets, and Paravisini, Rappoport and Ravina (2017) study portfolio choices in a peer-to-peer lending market. Due to the nature of the data they analyze, these studies estimate only a subset of the CPT features. An exception is Andrikogiannopoulou and Papakonstantinou (2019) who, using the same data as we do, estimate the full CPT specification, but focus on history dependence rather than heterogeneity. Our paper is different in that it accommodates rich heterogeneity both within and across utility types which allows us not only to estimate all CPT features but also to assess for the first time their relative importance and prevalence.
Our findings are more directly related to the small number of field studies that emphasize the importance
2For lab experiments, see, e.g., Tversky and Kahneman (1992), Hey and Orme (1994), Holt and Laury (2002), Choi et al. (2007), and von Gaudecker, van Soest and Wengstrom (2011). For surveys, see, e.g., Barsky et al. (1997) and Donkers, Melenberg and van Soest (2001). For TV game show studies, see, e.g., Gertner (1993), Metrick (1995), Beetsma and Schotman (2001), Post et al. (2008), and Bombardini and Trebbi (2012). 3An exception is the recent study by Chiappori et al. (2019), who use aggregate betting data to back out heterogeneity in a generic one-dimensional utility.
3 of probability weighting. Snowberg and Wolfers (2010) compare a model with only probability weighting to a model with only risk aversion and find that the former better explains aggregate prices in the horse-racetrack betting market. Barseghyan et al. (2013) find that, to explain individual insurance choices, it is necessary to use a model with a form of probability distortions that could be interpreted as probability weighting and/or loss aversion. These studies do not juxtapose probability weighting with loss aversion because, with the data available to them, the two cannot be distinguished. Thus, an important contribution of our study is that we exploit two distinct characteristics of our data — the large number of observations per individual and the great variety of possible choices — to separately identify loss aversion and probability weighting and we develop a rich model of heterogeneity to assess their relative prevalence in the population.
Methodologically, most of the literature has followed a parametric estimation approach, as we do. Some studies have taken a nonparametric approach, but usually at the cost of ignoring individual heterogeneity which is one of the main focuses of our analysis (e.g., Abdellaoui, 2000; Bleichrodt and Pinto, 2000).4
Furthermore, our model builds on and extends earlier work in the experimental literature that characterizes risk preference heterogeneity using mixture or hierarchical models, and generalizes it to incorporate both features simultaneously. Using mixtures, Harrison and Rutstrom (2009) propose that each subject probabilis- tically behaves according to expected utility theory for some choices and according to prospect theory for others; instead, we classify individuals rather than choices to utility types. Considering a mixture of utility types, Conte, Hey and Moffatt (2011) estimate population-level parameters assuming all individuals belong to both types, and then allocate individuals to types. Instead, we estimate parameters at the population and the individual level in a consistent manner, as whatever we learn about the population feeds back into the estimation of individuals’ preferences and type allocations, and vice versa. Finally, Bruhin, Fehr-Duda and Epper (2010) consider a mixture of CPT types, each with different parameters but with no variation across individuals within types; we accommodate a broader variation in preferences by allowing for both across- and within-type heterogeneity. On the other hand, Nilsson, Rieskamp and Wagenmakers (2011),
Murphy and ten Brincke (2017), and Baillon, Bleichrodt and Spinu (2019) use hierarchical models to estimate prospect theory preferences by assuming that parameters are drawn from a normal distribution.
We generalize this approach by allowing parameters to be drawn from a mixture distribution, with each component corresponding to a different utility type. Thus, we are able to estimate the proportions of different
4See Stott (2006) and Barseghyan et al. (2018) for a discussion of the trade-offs between parametric and nonparametric preference estimation. Succinctly, while a parametric approach risks being misspecified, a nonparametric one is less efficient and more demanding of the data in terms of identification.
4 utility types simultaneously with the entire distribution of risk preferences, and to share information across
individuals in a manner that is not restricted by the normal parametric assumption.
Our paper, like other field studies on risk preference estimation, should be read with the following
caveat in mind. It focuses on a subset of the population — people participating in a sports wagering market
— which may not be representative of the whole population. But, even though the group of people who
bet on sports may be self-selected, it is in fact quite large: Betting is a very common and economically
important activity, with “half of America’s population and over two-thirds of Britain’s [placing a] bet on
something” per year (The Economist, 8 July 2010) and with $1 trillion wagered on sports, globally, per year
(H2 Gambling Capital report of 2013). Furthermore, the fact that the risk preferences we estimate (i) are
very heterogeneous, (ii) are within the range of extant estimates from lab experiments, and (iii) yield sensible
predictions when applied to different choice settings as demonstrated in our portfolio choice application,
further indicates that a potential sample selection bias is likely to be limited. Regardless, keeping in mind that
extrapolating preference parameters from any group or context to another necessitates additional assumptions,
our estimates from this setting are complementary to those from other settings (e.g., insurance choice), and
may even be particularly relevant for sub-populations or contexts that share similar characteristics with
those in wagering markets (e.g., traditional financial markets).
The rest of the paper is organized as follows. In Section 1, we describe the data. In Section 2 we present
our model of individual behavior, in Section 3 we describe our estimation methodology, and in Section 4
we discuss our results. In Section 5 we use our estimates in a portfolio choice application, and in Section
6 we present robustness checks. In Section 7, we conclude. An Appendix contains additional details on the
estimation procedure, and an Internet Appendix contains additional tables and figures supporting our analysis.
1 Data
Here, we explain how the online sports-betting market works, and we describe our data.
1.1 Background information and data description
A sportsbook is a quote-driven dealer market that offers betting opportunities on a variety of sports (e.g.,
soccer, tennis, and auto races), a variety of fixtures within each sport (e.g., soccer matches from the English
Premier League and the European Champions League), and a variety of events associated with each fixture
(e.g., for a given soccer match, one can wager on the winning team, the exact score, the total number of
5 corners, etc.). The bookmaker acts as a dealer and, at any point in time, he sets the odds for each of the
possible outcomes of an event, i.e., the inverse of the price for receiving a unit monetary payoff in case the
outcome is realized. Customers can place wagers at these odds and the bookmaker takes the opposite side
of each wager, with the intention of making a profit through the bid-ask spread, often called the commission.
For example, an individual betting on the winner of a soccer match can select one of three possible outcomes
— home win, draw, or away win — each of which has a price determined by the bookmaker. If the outcomes’
prices are, respectively, 0.55, 0.30, and 0.20 (quoted as 1 1.82, 1 3.33, and 1 5.00 in 0.55 ≈ 0.30 ≈ 0.20 = decimal odds5) then an individual backing the home team to win will make AC82 for each AC100 staked if
the home team wins; similarly, he will make AC233 or AC400 for each AC100 staked, if he backs the draw
or away outcome and his bet wins. The sum of the prices on all outcomes of the event is the price of a
portfolio which pays AC1 for sure at the end of the event; the amount by which this sum exceeds 1 (in this
example, 0.55 0.30 0.20 1 0.05) represents the bookmaker’s commission for each euro wagered. + + − = The quoted prices are associated with the outcomes’ implied probabilities, calculated in this example as
0.55 0.30 0.20 0.55 0.30 0.20 0.52, 0.55 0.30 0.20 0.29, and 0.55 0.30 0.20 0.19, respectively. + + ≈ + + ≈ + + ≈ According to the traditional sportsbook pricing model, the bookmaker aims to keep a ‘balanced’ book;
for example, if the home team is bet heavily, the bookmaker could lower its odds to attract money on the
other team. This way, the amount of money paid out to winners, and therefore his commission, is the same
regardless of which team wins. But, even though the bookmaker can change the odds over time, each bet’s
payoff is determined by the odds at which it was placed.6 This means that the book is rarely completely
balanced, and that balancing the book can be a risky strategy that could be exploited by arbitrageurs if
prices deviate substantially from the efficient ones. As a result, it has been suggested that bookmakers
may find it optimal to set the prices close to the efficient ones, even if this implies that the book becomes
slightly unbalanced sometimes (see Paul and Weinbach, 2008, 2009 for empirical evidence supporting this
view). This approach reduces the costs and risks of balancing the book, while the bookmaker still earns
the commission in expectation.7 For example, if the home team is expected to win with probability 0.52
0.52 and the bookmaker wants to make a commission of 5%, he can set the home team’s price at 1 0.05 0.55 − ≈ 5Odds are quoted to two decimal places, with a tick size that depends on the odds, becoming smaller as the odds approach 1, i.e., as an outcome becomes a heavy favorite. 6This applies to fixed-odds betting markets, as the one we study. In parimutuel betting, commonly used in horse-racetrack betting, bettors place wagers in a common pool, and the payout of winning bets is not known in advance but rather is determined at the end of the event by splitting the pool between the winners. 7Levitt (2004) suggests that bookmakers may allow unbalanced betting to exploit bettor biases, but the bookmaker under study has indicated that he does not employ this strategy.
6 (quoted as 1 1.82 in decimal odds). So, in expectation, individuals backing this outcome will make 0.55 ≈ 0.52 ( AC82) 0.48 ( AC100) AC5 and the bookmaker will make AC5 for every AC100 wagered. Our · + + · − ≈ − market efficiency tests below and the conversations we had with our data provider suggest that the efficient pricing model better describes the behavior of the bookmaker under study. Regardless, in Section 6.2 we show that our preference estimates are quite robust to both price-setting behaviors.
An individual can place a single wager as a single bet, or he can combine multiple wagers into various bet types to produce a wide array of payoff profiles and risk levels. Specifically, N wagers can be combined in a type-N accumulator bet, which yields a gross return equal to the product of the gross returns of all N wagers if all of them win, and zero otherwise. Or they can be combined into permutation or full-cover bets, which themselves combine various accumulators. Specifically, one can form a permutation bet that includes