{
  "id": "expected-utility-theory",
  "title": "Expected Utility Theory",
  "lang": "en",
  "source_lang": "en",
  "summary": "Expected utility theory represents choices under risk or uncertainty as the maximization of probability-weighted utility across possible outcomes.",
  "category": "economics",
  "subcategory": "microeconomics",
  "type": "concept",
  "aliases": [
    "Expected utility hypothesis",
    "Expected utility model"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/expected-utility-theory/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/expected-utility-theory/",
      "json": "https://aiwiki.page/zh/expected-utility-theory.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Expected utility theory is a mathematical framework for representing and evaluating choices whose consequences are uncertain. It assigns a numerical utility to each possible outcome and evaluates an option by the probability-weighted average of those utilities. An expected utility maximizer chooses an option with the greatest such average. The theory distinguishes the desirability of consequences from their likelihood: a decision can be unattractive despite a high expected monetary return if its possible losses are sufficiently undesirable.\n\nThe framework is foundational in microeconomics, finance, insurance, and statistical decision theory. Its central contribution is not simply a calculation rule but a set of representation theorems: under specified consistency conditions, a person's preferences can be represented as if that person maximized expected utility. Versions of the theory distinguish decisions with given probabilities from decisions in which probabilities represent personal beliefs.\n\nExpected utility theory has both normative and descriptive interpretations. It can specify standards for coherent choice, provide a tractable model of economic behavior, or serve as a benchmark against which observed decisions are compared. Its mathematical foundations, treatment of risk, applications, and documented empirical limitations therefore form distinct but closely connected parts of the theory."
    },
    {
      "heading": "Historical origins and axiomatization",
      "text": "The immediate precursor of expected utility theory was the use of expected monetary value to evaluate gambles. For a gamble paying amounts \\(x_i\\) with probabilities \\(p_i\\), this criterion assigns the value \\(\\sum_i p_i x_i\\). It provides a natural accounting measure, but it does not explain why people may reject a favorable gamble or pay to avoid a loss. The distinction between the average monetary payoff and the attractiveness of a risky prospect became central to the development of utility-based evaluation.\n\nThe St. Petersburg paradox, proposed by Nicolaus Bernoulli in 1713, exposed a striking difficulty. In a standard version, a fair coin is tossed until the first head appears. If the first head occurs on toss \\(n\\), the prize is \\(2^n\\) monetary units. Since that event has probability \\(2^{-n}\\), every possible stopping time contributes one unit to expected monetary value, making the total infinite. Nevertheless, ordinary willingness to pay for entry is finite. The paradox challenges the identification of a gamble's value with its expected monetary payoff.\n\nDaniel Bernoulli's 1738 paper proposed evaluating outcomes through their utility rather than through money alone. His logarithmic specification made the utility of an additional monetary unit decline with wealth. Under this approach, the expected utility of the St. Petersburg gamble can remain finite even though its expected monetary payoff is infinite. Gabriel Cramer had independently proposed a related utility-based explanation in correspondence in 1728. These contributions anticipated the modern distinction between money and subjective value, although logarithmic utility is only one possible specification and does not resolve every version of an unbounded-payoff paradox.\n\nThe modern axiomatic formulation emerged in John von Neumann and Oskar Morgenstern's Theory of Games and Economic Behavior, first published in 1944. They showed how preferences over lotteries could support an expected utility representation. The significance was methodological: utility no longer had to be introduced as a directly measurable psychological sensation. Instead, its numerical structure could be derived from restrictions on choices among uncertain prospects.\n\nLeonard J. Savage's The Foundations of Statistics in 1954 extended the approach to situations without externally supplied probabilities. His framework derived both subjective probabilities and utility from preferences over state-contingent acts. Francis J. Anscombe and Robert J. Aumann supplied another influential formulation in 1963, combining subjective uncertainty with objective randomization. These developments established separate but connected foundations for expected utility under known risk and under subjective uncertainty."
    },
    {
      "heading": "Lotteries, axioms, and the representation theorem",
      "text": "In the standard finite-outcome formulation, a lottery is a probability distribution over consequences. Write a lottery as \\(L=(p_1,x_1;\\ldots;p_n,x_n)\\), where \\(p_i\\geq 0\\) and \\(\\sum_i p_i=1\\). An expected utility representation takes the form\n\n\\[U(L)=\\sum_{i=1}^{n}p_i u(x_i).\\]\n\nHere \\(u\\) is a utility function on consequences, while \\(U\\) represents preferences over lotteries. The distinction matters: \\(U\\) is linear in probabilities even when \\(u\\) is nonlinear in money. Expected utility therefore does not assume that doubling wealth doubles welfare.\n\nThe representation rests on several preference conditions. Completeness requires that any two lotteries can be ranked, allowing indifference. Transitivity requires internally consistent rankings: if \\(L\\) is at least as good as \\(M\\), and \\(M\\) at least as good as \\(N\\), then \\(L\\) must be at least as good as \\(N\\). A continuity condition excludes certain discontinuities in preference. For example, when \\(L\\) is preferred to \\(M\\), and \\(M\\) to \\(N\\), suitable mixtures of \\(L\\) and \\(N\\) can match or bracket \\(M\\). Lexicographic preferences, in which one consideration always overrides another regardless of probability, generally fail this condition.\n\nThe distinctive restriction is the independence axiom. For any lotteries \\(L,M,N\\) and \\(0<\\alpha<1\\),\n\n\\[L\\succeq M \\quad\\Longleftrightarrow\\quad \\alpha L+(1-\\alpha)N \\succeq \\alpha M+(1-\\alpha)N.\\]\n\nAdding the same probabilistic component to two alternatives must not reverse their ranking. The argument for independence is that the common component cannot distinguish the options: whenever it occurs, both deliver the same prospect. Its empirical vulnerability arises because people may nevertheless react differently to mixtures involving certainty, very small probabilities, or especially attractive outcomes.\n\nUnder the standard assumptions, expected utility represents preferences, and \\(u\\) is unique up to a positive affine transformation. Thus \\(v(x)=a u(x)+b\\), with \\(a>0\\), represents the same lottery preferences. Arbitrary increasing transformations do not generally preserve expected utility rankings. This is a stronger numerical structure than the ordinal utility representation used for deterministic consumer choice, but it does not establish comparisons of utility levels between different people.\n\nThe usual lottery framework also reduces compound lotteries to their final outcome probabilities. A 50 percent chance of entering a second lottery that pays a prize with probability 40 percent becomes a 20 percent chance of that prize. This reduction assumes that only final consequences matter, not the particular randomization procedure. If suspense, timing of information, or the procedure itself affects preferences, those features must be included among consequences or modeled separately."
    },
    {
      "heading": "Risk aversion, certainty equivalents, and utility curvature",
      "text": "For monetary outcomes, expected utility connects attitudes toward risk to the shape of utility over wealth. An individual is risk averse if receiving the expected wealth for certain is at least as desirable as receiving the risky wealth itself. For random wealth \\(W\\),\n\n\\[u(\\mathbb{E}[W])\\geq \\mathbb{E}[u(W)].\\]\n\nFor an expected utility maximizer, this property corresponds to concavity of \\(u\\), on the relevant domain and for the class of lotteries considered. Linear utility describes risk neutrality; convex utility describes risk seeking. Strict concavity gives a strict preference for the mean over any nondegenerate lottery within the domain.\n\nThe certainty equivalent \\(c\\) of a lottery is a guaranteed amount satisfying\n\n\\[u(c)=\\mathbb{E}[u(W)].\\]\n\nIts risk premium is \\(\\pi=\\mathbb{E}[W]-c\\). Consider an individual with logarithmic utility facing equal chances of final wealth 50 and 150. Expected wealth is 100, but the certainty equivalent is\n\n\\[c=\\exp\\!\\left(\\tfrac12\\ln 50+\\tfrac12\\ln 150\\right) =\\sqrt{7500}\\approx 86.60.\\]\n\nThe corresponding risk premium is approximately 13.40. These values depend on the assumed utility function and on the outcomes being final wealth, rather than gains added to some unspecified background wealth.\n\nKenneth Arrow and John W. Pratt developed influential local measures of risk aversion in the 1960s. Absolute risk aversion and relative risk aversion are respectively\n\n\\[A(w)=-\\frac{u''(w)}{u'(w)}, \\qquad R(w)=-\\frac{w u''(w)}{u'(w)}.\\]\n\nAbsolute risk aversion is useful for comparing attitudes toward monetary risks of a fixed size. Relative risk aversion is useful for risks proportional to wealth. For a small, zero-mean risk \\(\\varepsilon\\), a second-order approximation yields\n\n\\[\\pi\\approx \\frac12 A(w)\\operatorname{Var}(\\varepsilon).\\]\n\nThis connects the compensation required for bearing a small risk to utility curvature and the risk's variance. It is a local approximation, not a general formula for large or highly asymmetric risks.\n\nCommon specifications embody different wealth effects. Exponential utility \\(u(w)=-e^{-aw}\\), with \\(a>0\\), has constant absolute risk aversion. Power utility,\n\n\\[u(w)=\\frac{w^{1-\\gamma}}{1-\\gamma},\\]\n\nfor positive wealth, \\(\\gamma>0\\), and \\(\\gamma\\neq1\\), has constant relative risk aversion \\(\\gamma\\). Its logarithmic limit occurs at \\(\\gamma=1\\). These functional forms simplify modeling but impose substantive restrictions on behavior.\n\nConcavity also explains diversification and demand for protection against losses, but it does not imply that every risky investment should be rejected. A risky option may have sufficiently favorable returns to outweigh its risk. Moreover, an individual's evaluation depends on how the new risk interacts with existing wealth, earnings, liabilities, and other exposures."
    },
    {
      "heading": "Subjective probability and decisions under uncertainty",
      "text": "Objective-risk models treat probabilities as given, as with an idealized roulette wheel or a specified randomization device. Many consequential decisions have no such probabilities: a firm considers a new technology, a household evaluates future employment, or an investor assesses a political event. Subjective expected utility extends the framework by representing uncertainty through personal probabilities over possible states of the world.\n\nLet \\(S\\) be a set of states and let an act \\(f\\) assign consequence \\(f(s)\\) to each state \\(s\\). In a finite-state formulation,\n\n\\[V(f)=\\sum_{s\\in S}P(s)u(f(s)).\\]\n\nThe probabilities \\(P(s)\\) express beliefs, while \\(u\\) expresses the desirability of consequences. Two people can choose differently because they disagree about likelihoods, value consequences differently, or both. This separation helps distinguish disagreements about evidence from disagreements about objectives, although observed choices alone do not always identify the two components.\n\nSavage's representation depends on restrictions on preferences over acts and on a sufficiently rich state structure. A central condition is the sure-thing principle: if one act is preferred to another conditional on an event, and also conditional on its complement, it should be preferred without knowing whether the event occurs. More generally, a ranking between acts that differ only on an event should not depend on the common consequences outside that event. This resembles lottery independence, but it is formulated for uncertain states rather than objectively randomized mixtures.\n\nThese foundations also reveal limitations. Standard subjective expected utility normally uses a state-independent utility function: a given consequence has the same utility whichever state produces it. That assumption becomes problematic if consequences are incompletely described. Money received during illness may matter differently from money received in good health because health changes opportunities and needs. Expanding consequences to include health can clarify the model, but state-dependent preferences make the separate recovery of probabilities and utilities more difficult.\n\nSubjective probabilities can be updated through Bayesian inference. For an event \\(E\\) with positive prior probability,\n\n\\[P(s\\mid E)=\\frac{P(E\\mid s)P(s)}{P(E)}.\\]\n\nExpected utility then combines updated beliefs with consequences when selecting an action. Bayesian updating is not itself an attitude toward risk: people with identical beliefs can still choose different actions because their utilities differ.\n\nThe distinction between risk and ambiguity remains important. Risk concerns variation under a specified probability distribution; ambiguity concerns the absence, imprecision, or contested reliability of such a distribution. Subjective expected utility represents uncertainty through a single probability measure. That is a substantive modeling commitment, not a demonstration that every uncertain situation has an objectively correct numerical probability."
    },
    {
      "heading": "Sequential decisions, information, and intertemporal choice",
      "text": "Many decisions unfold through time. An initial action changes what can be observed, later choices depend on that observation, and final outcomes depend on the entire sequence. A decision tree represents such a problem through choice nodes, chance nodes, information, and terminal consequences. Expected utility can evaluate complete contingent plans: instructions specifying what to do at every possible future information set.\n\nWith a fixed utility over terminal consequences and Bayesian conditional probabilities, backward induction generally produces the same plan as an initial evaluation of all complete strategies. This property is a form of dynamic consistency. At each future information set, the selected continuation remains optimal given the information then available. The conclusion requires an unchanged preference structure and appropriately specified information. Changes in tastes, unmodeled learning, or a different treatment of compound uncertainty can disrupt it.\n\nInformation has value because it allows actions to respond to events. Suppose signal \\(Y\\) arrives before action \\(a\\) is chosen. Its expected value, measured in utility units before any acquisition cost, is\n\n\\[\\mathbb{E}_Y\\!\\left[ \\max_a \\mathbb{E}[u(x(a,S))\\mid Y] \\right] - \\max_a \\mathbb{E}[u(x(a,S))].\\]\n\nUnder the standard assumptions this expression is nonnegative, because the decision maker can ignore the signal and retain the original action. It does not imply that information is worth any price. Information-acquisition costs, processing burdens, or direct preferences concerning knowledge must be incorporated separately. Converting a utility gain into a monetary willingness to pay also requires the relevant wealth-dependent utility calculation.\n\nIntertemporal models frequently assume expected discounted utility,\n\n\\[\\mathbb{E}\\left[\\sum_{t=0}^{T}\\beta^t u(c_t)\\right],\\]\n\nwhere \\(c_t\\) is consumption and \\(\\beta\\) is a discount factor. This combines several assumptions that should be distinguished. Expected utility governs uncertainty; additive separability governs how consumption at different dates is combined; exponential discounting governs the treatment of delay. None of the latter two properties follows from the basic lottery representation theorem. In particular, ordinary expected utility does not rule out present bias unless additional temporal restrictions are imposed.\n\nIn the familiar time-additive power-utility model, a single curvature parameter governs both risk aversion and the inverse elasticity of intertemporal substitution. This can be restrictive because attitudes toward uncertain consumption need not coincide with willingness to shift certain consumption between dates. Recursive utility, including Epstein–Zin preferences developed in 1989, allows these dimensions to be separated. Such models distinguish evaluation of current consumption from evaluation of risky future continuation utility and can capture preferences over when uncertainty is resolved."
    },
    {
      "heading": "Economic and financial applications",
      "text": "Insurance provides a direct application. Suppose initial wealth is \\(w\\), a loss \\(L\\) occurs with probability \\(p\\), and full insurance costs \\(pL\\). Without insurance, expected utility is\n\n\\[(1-p)u(w)+pu(w-L).\\]\n\nWith actuarially fair full insurance, wealth is certainly \\(w-pL\\). Concavity implies that the insured position is weakly preferred, and strictly preferred under strict concavity when the loss is genuinely uncertain. Administrative expenses and other loadings can make full insurance unattractive. A deductible leaves some risk with the policyholder and may also address moral hazard, which arises when insurance changes incentives to prevent losses.\n\nPortfolio theory applies the same logic to combinations of assets. An investor selects holdings to maximize expected utility of terminal wealth or future consumption subject to budget and trading constraints. A risky asset's attractiveness depends not only on its own expected payoff but also on its relationship with the investor's other resources. An asset paying well in states where wealth would otherwise be low can provide insurance value. Expected utility thus explains why separate evaluation of each investment's risk can misrepresent the risk of the portfolio as a whole.\n\nMean–variance portfolio models are connected to expected utility under particular assumptions or approximations. With normally distributed wealth and exponential utility, maximizing expected utility is equivalent to maximizing expected wealth minus a term proportional to its variance. Quadratic utility also produces a mean–variance objective, although it can imply decreasing utility at sufficiently high wealth. For general payoff distributions and utility functions, means and variances alone are insufficient: skewness, tail losses, and other distributional features can affect choices.\n\nConsumption-based asset pricing relates prices to the marginal utility of payoffs. In a standard representative-agent formulation,\n\n\\[P_t=\\mathbb{E}_t[M_{t+1}X_{t+1}], \\qquad M_{t+1}=\\beta\\frac{u'(c_{t+1})}{u'(c_t)}.\\]\n\nHere \\(P_t\\) is an asset's price, \\(X_{t+1}\\) its next-period payoff, and \\(M_{t+1}\\) the stochastic discount factor. A payoff is especially valuable when it arrives in a state with high marginal utility, often a state of low consumption. This connects risk compensation to covariance with economically important bad states rather than to volatility alone.\n\nExpected utility also appears in principal–agent models, where compensation contracts balance incentives against the costs of imposing risk on workers or managers. In statistical decision theory, an action can instead be chosen to minimize expected loss, which is formally equivalent to maximizing negative expected loss. In welfare economics, however, using individual expected utility does not by itself determine how different people's interests should be aggregated. Social weights, distributional judgments, and interpersonal comparisons require additional assumptions."
    },
    {
      "heading": "Experimental evidence and systematic violations",
      "text": "Expected utility yields testable restrictions on choices across related lotteries. The best-known challenge is the Allais paradox, introduced by Maurice Allais in 1953. In a familiar presentation, one choice is between a guaranteed 1 million and a lottery offering 5 million with probability 0.10, 1 million with probability 0.89, and nothing with probability 0.01. A second choice is between a lottery offering 1 million with probability 0.11 and a lottery offering 5 million with probability 0.10, each paying nothing otherwise.\n\nA commonly reported pattern is to prefer the guaranteed amount in the first comparison but the chance of 5 million in the second. This is incompatible with expected utility using the same utility function across both comparisons. The first preference implies\n\n\\[0.11u(1\\text{ million}) > 0.10u(5\\text{ million})+0.01u(0),\\]\n\nwhile the second implies the opposite inequality. The common 0.89 probability of receiving 1 million in the first comparison has been replaced by a common 0.89 probability of receiving zero in the second. Independence requires that this common change not reverse the ranking.\n\nDaniel Ellsberg's 1961 examples challenge subjective expected utility through choices involving unspecified probabilities. In a standard urn, 30 balls are red and 60 are black or yellow in unknown proportions. Many participants prefer a bet on red to a bet on black, but prefer a bet on black or yellow to a bet on red or yellow. With otherwise identical prizes, the first preference suggests a subjective probability of black below that of red; the second suggests the reverse. The pattern reflects a preference for known probabilities that cannot be represented by a single additive probability measure under the standard assumptions.\n\nOther findings concern the description and organization of choices. In preference-reversal experiments, people may choose one gamble but assign a higher selling price to another. Framing effects occur when different descriptions of equivalent consequences alter decisions. Loss aversion and mental accounting can make choices depend on a reference point or on whether a gamble is evaluated separately from other wealth. These observations challenge a model in which only the probability distribution of fully specified final consequences determines preference.\n\nInterpretation requires attention to experimental design. Comprehension, stakes, repetition, payment procedures, and random decision error affect results. Noisy choices do not automatically establish a systematic axiom violation, while consistent violations across well-controlled tasks cannot be explained merely by selecting a different concave utility function. The evidence rejects expected utility as a universally accurate description of behavior; it does not show that every individual violates it, that every application fails, or that the theory has no normative role."
    },
    {
      "heading": "Alternative models of risky and ambiguous choice",
      "text": "Alternative theories relax different components of expected utility rather than replacing it with a single universally accepted model. Some retain probabilities but change their decision weights. Others preserve expected utility within each possible probability model while changing how uncertainty about that model is evaluated. Still others introduce reference dependence, regret, incomplete preferences, or procedural constraints. The distinction matters because a theory designed to explain ambiguity need not also explain framing or attitudes toward known risks.\n\nProspect theory, introduced by Daniel Kahneman and Amos Tversky in 1979, evaluates gains and losses relative to a reference point rather than necessarily evaluating final wealth. Its value function is typically concave for gains, convex for losses, and steeper for losses than for comparable gains. It also uses decision weights that need not equal probabilities. Cumulative prospect theory, developed in 1992, applies weighting to cumulative probabilities. This allows sensitivity to small probabilities and near-certainty while, under its standard conditions, respecting stochastic dominance.\n\nRank-dependent utility, developed by John Quiggin in 1982, retains utility over outcomes but transforms cumulative probabilities according to outcomes' ranks. Unlike separately transforming each outcome probability, cumulative weighting can preserve appropriate ranking properties when outcomes are combined or reordered. Standard expected utility is recovered when the probability-weighting function is the identity. Rank dependence therefore provides a structured relaxation of independence without requiring reference-dependent gains and losses.\n\nFor ambiguous choices, maxmin expected utility, axiomatized by Itzhak Gilboa and David Schmeidler in 1989, represents beliefs by a set \\(\\mathcal{P}\\) of probability distributions:\n\n\\[V(f)=\\min_{P\\in\\mathcal{P}}\\mathbb{E}_P[u(f)].\\]\n\nAn act is evaluated by its least favorable expected utility within that set. If the set contains only one distribution, the model reduces to subjective expected utility. Its conservative treatment of uncertainty can explain some Ellsberg-type choices, although the set of priors must itself be specified or inferred. Smooth ambiguity models instead aggregate expected utilities across possible probability models through an additional function, separating risk preferences from attitudes toward ambiguity.\n\nRegret theory, developed independently by Graham Loomes and Robert Sugden and by David Bell in 1982, allows evaluation to depend on comparisons between the chosen outcome and what another available action would have produced. It therefore makes the menu of alternatives relevant in a way absent from standard expected utility over final consequences. More broadly, behavioral economics and bounded rationality examine attention, computational limits, and choice procedures. These approaches address different empirical mechanisms and must be judged against the particular phenomena they are intended to explain."
    },
    {
      "heading": "Interpretation, measurement, and limits of the framework",
      "text": "An expected utility representation is an “as if” statement about preferences, not necessarily a claim about conscious calculation. A person may satisfy the relevant choice restrictions without explicitly computing probabilities or utilities. Conversely, an analyst's use of an expected utility objective does not establish that the modeled decision maker possesses complete rankings over every conceivable gamble. The distinction between a formal representation and a psychological process is essential when applying the theory outside tightly specified choice problems.\n\nUtility can be elicited from indifferences between certain outcomes and lotteries. Suppose \\(x_H\\) and \\(x_L\\) are the best and worst outcomes in a bounded set, normalized so that \\(u(x_H)=1\\) and \\(u(x_L)=0\\). If a person is indifferent between receiving \\(x\\) certainly and a lottery paying \\(x_H\\) with probability \\(p\\) and \\(x_L\\) otherwise, expected utility implies \\(u(x)=p\\). This standard-gamble procedure identifies utility relative to the chosen endpoints. Its interpretation assumes the representation being tested: probability weighting, framing, or changing reference points can alter the inferred utility.\n\nConsequences must also be specified at the correct level. Treating a gamble as a change in total lifetime wealth yields different implications from treating it as a self-contained gain or loss. Matthew Rabin's 2000 calibration theorem showed that consistent rejection of modest-stakes favorable gambles across a sufficiently broad wealth range can, under concave expected utility over wealth, imply extreme aversion to much larger gambles. The result exposes limits to explaining all observed small-stakes caution solely through the curvature of utility over total wealth.\n\nExpected utility is compatible with broader goals than personal monetary gain. Consequences may include other people's welfare, environmental conditions, fairness, or the consequences of complying with a rule. This flexibility makes the framework widely applicable, but it can reduce testability if every observed violation is explained by adding an unobserved consequence. A useful empirical model specifies the outcome space and preference restrictions before interpreting the evidence. Revealed preference methods then examine whether observed choices are consistent with those restrictions.\n\nThere are also mathematical boundaries. With infinitely many outcomes, expected utility may fail to exist if utilities are unbounded and positive and negative components are not integrable. Continuity can exclude preferences that assign absolute priority to avoiding a particular outcome, however small its probability. Complete preferences can be inappropriate when a decision maker leaves some options incomparable. These issues concern different assumptions and cannot all be repaired by changing utility curvature. In applications involving extreme outcomes or poorly specified probabilities, the admissible lotteries, probability model, and utility domain are therefore substantive parts of the model rather than incidental technical details."
    }
  ],
  "links": [
    "utility-economics",
    "microeconomics",
    "insurance",
    "expected-value",
    "probability-distribution",
    "utility-function",
    "consumer-choice",
    "risk-aversion",
    "kenneth-arrow",
    "variance",
    "bayesian-inference",
    "present-bias",
    "deductible",
    "moral-hazard",
    "principal-agent-problem",
    "loss-function",
    "welfare-economics",
    "loss-aversion",
    "mental-accounting",
    "prospect-theory",
    "behavioral-economics",
    "bounded-rationality"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}