Disappointment Aversion and Expectiles

arXiv:2508.05541 · econ.TH · Submitted 2025-08-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Disappointment Aversion and Expectiles".

Warren: Here's the gist of the paper. This paper recasts Gul’s (1991) theory of disappointment aversion in a Savage framework with general outcomes, new explicit axioms and representations that permit broader applications and a better understanding of its decision-theoretic foundations Our results exploit an unexpected connection between Gul’s model and the econometric framework of Newey and Powell (1987) of asymmetric least squares estimation The main axiomatization result shows that a preference relation over Savage acts is probabilistically sophisticated, invariant biseparable, and disappointment hedging if and only if it admits a representation a la Gul, and hence all explicit equivalent representations that we present in the paper We also derive a neurocomputational foundation of the theory based on recent neuroscience findings and a novel reinforcement learning result.

Ruth: Extended Domain and General Outcomes The original theory of disappointment aversion was formulated for monetary lotteries, yet the psychological phenomenon it seeks to capture appears at least as forcefully in non-monetary domains The paper extends Gul’s model to general outcomes in a Savage framework.

Warren: This extension allows the model to capture disappointment and elation beyond monetary rewards, such as results of personal endeavors or medical treatments The decision maker evaluates each act relative to its own endogenous certainty equivalent except now without any restriction to monetary lotteries.

Ruth: Explicit Representations and Econometric Links The paper provides three explicit representations for Gul’s model.

Title and authors: Iris: What the paper actually does in terms of structure is show that if a preference relation over Savage acts has these certain properties, it means it admits Gul's representation and all its explicit equivalents.

Ruth: That’s a strong result because it connects the behavioral concept of disappointment aversion to these specific mathematical structures like probabilistic sophistication and invariance.

Warren: They also derived a neurocomputational foundation for this theory based on some recent neuroscience findings and a new result in reinforcement learning.

Iris: So, beyond the preference relation itself, they are building up a whole set of supporting evidence from both decision theory and actual brain activity models.

Ruth: It suggests that the way we make choices isn't just about maximizing expected value but is shaped by how we process elation versus disappointment in the brain.

The paper's summary: Warren: The core idea they are presenting is that this framework can be used to capture disappointment and elation beyond monetary rewards, which was a limitation of the original Gul model.

Iris: They provide three explicit ways to represent Gul's model, including one using maxmin expected utility in terms of a set Q over probability measures and another using asymmetric least squares estimation based on the payoff function.

Ruth: And they tie this econometric side together by showing that the expectation E beta of the payoff u evaluated under this framework is equivalent to solutions found by asymmetric least squares estimators called expectiles.

Warren: This means they’re linking decision theory directly to these statistical methods, which is a pretty powerful way to test if these models hold up in real-world data.

Iris: The main result of the axiomatization shows that this preference relation is probabilistically sophisticated, invariant biseparable, and it's disappointment hedging if and only if it has this representation based on a function u and a number beta greater than or equal to zero.

The paper's improvements: Ruth: One major improvement they highlight is extending the domain to general outcomes, which means we can now evaluate risks in non-monetary situations like personal projects or medical decisions, not just money.

Warren: That broadening of the application is significant because it makes Gul's model much more useful for understanding real human behavior beyond simple financial choices.

Iris: They also introduce a new behavioral foundation based on explicit axioms of disappointment aversion and elation seeking that are quite easy to interpret, allowing you to determine the sign of beta just by looking at binary acts alone.

Ruth: And they have a formula for that, which is beta equals two P of F minus one over one minus P of F, where P of F is some probability related to the acts.

Warren: That formula gives us a direct way to identify the parameter beta from actual binary preferences observed in people, which could be useful for calibration.

Conclusion: Ruth: So, to wrap up, this paper shows that Gul's theory of disappointment aversion can be recast into a Savage framework with general outcomes and new representations.

Iris: It has an unexpected neurocomputational foundation based on dopamine-based reinforcement learning where the coefficient beta measures the relative strength of the negative prediction error channel compared with the positive one.

Warren: And they show that neural agents whose reward estimates are updated by this asymmetric reward prediction error learning converge to evaluations of expectiled rewards, which is a convergence guaranteed by Theorem seven.

Ruth: Essentially, this means disappointment aversion isn't just a theoretical construct; it’s read as a limiting outcome of asymmetric reinforcement learning in the brain.

Iris: They characterize expectiled utilities as the worst-case expected utility over a Hilbert projective ball around the reference probability in distributionally robust optimization.

Warren: It connects this theory to modern finance and statistics through this concept of expectiles, which are essentially solutions to problems like distributionally robust optimization.

Ruth: So, the big picture here is taking a classic model and giving it a much broader scope, linking it to econometrics, and grounding it in neurocomputational learning mechanisms.

Iris: I just want to emphasize that the paper notes its limitation is that the original formulation was restricted to monetary outcomes, even though disappointment over non-monetary things seems equally significant.

Warren: That’s a fair point; while they broaden the domain, they are still operating within a Savage framework which might impose some structural assumptions on how those general outcomes interact.

F. Bellini, F. Maccheroni, T. Mao, R. Wang, Q. Wu

Bicocca Bocconi University · USTC Institute of Science and Technology Center of China · University of Waterloo Institute for Applied Mathematics and Statistics Division of Yale University

econ.TH

Submitted: 2025-08-07

Updated: 2026-11-09

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: This paper recasts Gul’s (1991) theory of disappointment aversion in a Savage framework with general outcomes…

Key concepts

Expectiles
Expectiles are solutions derived from asymmetric least squares estimation of random payoffs. They provide a mathematical way to represent the decision-maker's expected utility when facing uncertain outcomes, connecting preference relations to econometric models.
Disappointment Hedging
This concept describes a preference relation where the decision-maker seeks protection against negative outcomes (disappointment) while potentially exploiting positive ones. It is characterized mathematically by the existence of a function u and a coefficient beta that governs this hedging behavior.
Asymmetric Reinforcement Learning
The theory connects disappointment aversion to how agents learn from rewards using asymmetric reward prediction errors (RPE). The coefficient beta quantifies the relative strength of the negative prediction-error channel versus the positive one, showing how this learning mechanism limits expected utility evaluations.

Terminology

Summary

This paper recasts Gul’s (1991) theory of disappointment aversion in a Savage framework with general outcomes, new explicit axioms and representations that permit broader applications and a better understanding of its decision-theoretic foundations Our results exploit an unexpected connection between Gul’s model and the econometric framework of Newey and Powell (1987) of asymmetric least squares estimation The main axiomatization result shows that a preference relation over Savage acts is probabilistically sophisticated, invariant biseparable, and disappointment hedging if and only if it admits a representation a la Gul, and hence all explicit equivalent representations that we present in the paper We also derive a neurocomputational foundation of the theory based on recent neuroscience findings and a novel reinforcement learning result. p. 0

Extended Domain and General Outcomes

The original theory of disappointment aversion was formulated for monetary lotteries, yet the psychological phenomenon it seeks to capture appears at least as forcefully in non-monetary domains The paper extends Gul’s model to general outcomes in a Savage framework. This extension allows the model to capture disappointment and elation beyond monetary rewards, such as results of personal endeavors or medical treatments The decision maker evaluates each act relative to its own endogenous certainty equivalent except now without any restriction to monetary lotteries. p. 3

Explicit Representations and Econometric Links

The paper provides three explicit representations for Gul’s model. These include an explicit representation in terms of maxmin expected utility a la Gilboa and Schmeidler (1989), another one in terms of asymmetric least squares estimation of the random payoff u (X) that the DM is facing, and a representation derived from equation (3) involving expected elation and disappointment. p. 1 The econometric connection is established through the expectation Eβ[u(X)] being equivalent to the solutions of asymmetric least squares estimators denoted by expectiles. p. 4

Axiomatization and Behavioral Foundations

The paper introduces a new behavioral foundation based on explicit axioms of disappointment aversion/elation seeking, which are easy to interpret and determine the sign of β. The key result is Theorem 4, which characterizes expectiled utility as a binary relation that is probabilistically sophisticated, invariant biseparable, and disappointment hedging if and only if there exists a continuous nonconstant function u: X → R and a number β ≥ 0 such that X ≿ Y ⇐⇒ Eβ [u (X)] ≥ Eβ [u (Y)]. p. 6 The behavioral identification of parameters shows that the coefficient β can be recovered from binary acts alone using the formula β = 2P(F) − 1 / 1 − P(F). p. 6

Neurocomputational and Robust Foundations

The paper establishes a neurocomputational foundation by showing that when a value estimate is updated through asymmetric reward prediction errors, the long-run learned value is precisely the expectiled reward. This learning rule relates to dopamine-based reinforcement learning The final theorem shows that this learning dynamic leads to expectiled reward evaluation, connecting dopamine-based learning and disappointment aversion through the econometric theory of expectiles. p. 6 Furthermore, expectiled utilities are characterized as the worst-case expected utility over a Hilbert projective ball around the reference probability in distributionally robust optimization. p. 7

Axiomatic Characterizations of Expectiles

The paper develops several axiomatic characterizations of expectiled utility preferences. Theorem 6 shows that the equivalence between monotonicity, probabilistic sophistication, continuity, and disappointment aversion (resp. p. 0 elation seeking) is equivalent to the strong monotonicity, L∞-norm continuity, and decrease in concordant sums (resp. p. 3 mixtures) of the induced functional I. p. 0 The final result shows that expectiled utility is equivalent to maxmin expected utility under certain conditions. p. 8

Properties of Expectiles

The paper provides several properties of expectiles that will be used in the proofs of our results. Lemma 19 proves that for bounded random variables, the conditional expectation E[G(V, U) H] = g(V) a. p. 5s., where G is jointly continuous and U is independent of H. p. 5 Theorem 7 shows that the stochastic recursion converges to this unique zero as Vt → v⋆ = Eβ[U] a. p. 6s. Lemma 18 provides an inequality for the function f(v) defined by (u − v) + − (1 + β)(v − u) + > - min (1, 1 + β) (v − v')2. Theorem 9 shows that expectiled utility is the minimum over a Hilbert projective neighborhood in distributionally robust optimization. p. 7 The proof of Theorem 3 relies on the dual representation of Eβ in Lemma 11. p. 6 The proof of Theorem 4 shows that disappointment hedging is equivalent to the existence of a nonconstant function u and a number β ≥ 0 such that X ≿ Y ⇐⇒ Eβ [u (X)] ≥ Eβ [u (Y)]. p. 6 The proof of Theorem 8 shows that betweenness is equivalent to the same characterization, leading to expectiled utility. p. 8 The proof of Theorem 13 establishes that I = Eβ for some β > −1 under certain conditions. p. 4

Summary of Results

The results show that the theory of disappointment aversion of Gul (1991) can be “remastered” in a Savage framework with general outcomes, new explicit representations, and an axiomatization that directly appeals to the idea of robustifying decisions to take protection against disappointment or exploit elation opportunities. The theory has an unexpected neurocomputational foundation based on the differential processing of elation and disappointment in the brain. p. 7 The paper concludes by showing that neural agents whose reward estimates are updated by asymmetric RPE learning converge to expectiled reward evaluations. p. 2

Pruning redundant reasoning tokens

The paper presents the following key findings regarding the theory of disappointment aversion of Gul (1991). The core contribution is an extension to general outcomes in a Savage framework and a new axiomatization based on explicit axioms of disappointment aversion/elation seeking. p. 1 This new framework is characterized by the existence of a function u and a coefficient β such that preferences are equivalent to Eβ [u (X)]. p. 1 The theory is linked to econometric frameworks via expectiles, which are solutions to asymmetric least squares estimation problems. p. 0

How the training signal is built

The learning mechanism for value estimation is modeled as temporal-difference learning The neural value estimate Vt = Vt(X) of the random outcome X after trial t is understood as being represented by activity in valuation circuits. This estimate is updated after each trial in proportion to the reward prediction error (RPE). p. 2 The update rule involves coefficients w+ and w- that summarize the net efficacy with which dopaminergic fluctuations induce plasticity in downstream circuits.

Where the gains come from

The gains in this framework stem from the connection between disappointment aversion and asymmetric reinforcement learning. The coefficient β measures the relative strength of the negative prediction-error channel compared with the positive one This means disappointment aversion is read as a limiting outcome of asymmetric reinforcement learning. p. 2 The convergence to expectiled reward evaluation is guaranteed by Theorem 7, which connects dopamine-based learning and disappointment aversion through the econometric theory of expectiles. p. 0 p. 0

The results show that the theory of disappointment aversion of Gul (1991) can be “remastered” in a Savage framework with general outcomes, new explicit representations, and an axiomatization that directly appeals to the idea of robustifying decisions to take protection against disappointment or exploit elation opportunities. The theory has an unexpected neurocomputational foundation based on the differential processing of elation and disappointment in the brain The paper concludes by showing that neural agents whose reward estimates are updated by asymmetric RPE learning converge to expectiled reward evaluations. p. 7

The gains in this framework stem from the connection between disappointment aversion and asymmetric reinforcement learning The coefficient β measures the relative strength of the negative prediction-error channel compared with the positive one This means disappointment aversion is read as a limiting outcome of asymmetric reinforcement learning. The convergence to expectiled reward evaluation is guaranteed by Theorem 7, which connects dopamine-based learning and disappointment aversion through the econometric theory of expectiles.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. General Outcome Evaluation: The improved system can evaluate risky actions beyond monetary rewards by treating outcomes as general measurable spaces, allowing it to model disappointment and elation in non-monetary domains like personal endeavors or medical treatments.

  2. Axiomatic Foundation: The system will incorporate new explicit axioms of disappointment aversion and elation seeking, enabling the determination of the sign of the coefficient beta through behavioral elicitation methods.

  3. Explicit Representations: The AI can utilize novel explicit representations, such as maxmin expected utility in terms of a set Q over probability measures or asymmetric least squares estimation using the loss function l beta(s), to provide more interpretable decision-theoretic foundations.

  4. Neurocomputational Learning Model: The system can model value estimation using a reinforcement learning result where the coefficient beta represents the relative strength of the negative prediction-error channel compared with the positive one, connecting it to dopamine-based learning mechanisms.

  5. Elicitation Procedures: The improved system can perform direct elicitation of parameters by examining indifference judgments between binary acts, yielding beta via the formula β = 2P(F) - 1 / (1 - P(F)).

  6. Robust Decision Making: The system can select a Hilbert projective neighborhood for expected utility, which connects expectiles to distributionally robust optimization solutions in modern finance and statistics.

  7. Preference Characterization: The system can characterize preferences as invariant biseparable if they satisfy probabilistic sophistication, disappointment hedging, and betweenness, providing a new lens for understanding risk aversion beyond standard models.

Sources

Related papers