Prospect theory is a descriptive theory of decision-making under risk developed by Daniel Kahneman and Amos Tversky. First published in 1979, it explains choices that systematically depart from expected utility theory, including the tendency to evaluate outcomes as gains or losses relative to a reference point, to respond more strongly to losses than to comparable gains, and to give probabilities nonlinear decision weights. Its central claim is that choices depend not only on final outcomes and their probabilities but also on how those outcomes are represented.
The theory became a foundation of behavioral economics by combining economic models of choice with experimental findings from psychology. A revised formulation, cumulative prospect theory, appeared in 1992 and extended the approach to a wider class of uncertain prospects. Applications range from investment and insurance to consumer demand, labor supply, and policy communication. Prospect theory is primarily an account of how people choose, rather than a prescription for how they ought to choose; its explanatory success depends on specifying reference points, decision weights, and the context in which alternatives are evaluated.
Intellectual origins and the challenge to expected utility
The conventional benchmark for risky choice is expected utility theory. In its standard monetary formulation, a decision-maker evaluates a lottery by averaging the utility of its possible final wealth levels, weighted by their probabilities. If initial wealth is , possible monetary changes are , and their probabilities are , the evaluation is
A concave utility function represents risk aversion: the person prefers a certain amount to a risky prospect with the same expected monetary value. The framework allows substantial differences in preferences, but it imposes consistency conditions on how lotteries are ranked.
One important condition is independence: a preference between two lotteries should not reverse when both are mixed in the same proportions with an identical third lottery. Maurice Allais challenged the descriptive adequacy of this condition in 1953. Choices associated with the Allais paradox indicate that people may attach special importance to certainty, rather than evaluate all probability changes in the manner required by expected utility. Earlier work by Daniel Bernoulli had already distinguished the utility of money from money itself, but it did not contain prospect theory’s account of reference-dependent gains and losses.
Kahneman and Tversky’s 1979 article, “Prospect Theory: An Analysis of Decision under Risk,” brought several recurring experimental patterns into a single framework. Respondents often preferred a sure gain over a larger uncertain gain, while preferring an uncertain loss over a smaller certain loss. They also reacted differently to formally equivalent problems when the descriptions changed. Many demonstrations used hypothetical monetary choices with stated probabilities, allowing the researchers to compare observed response patterns with precise theoretical predictions.
The proposed explanation changed both what receives value and how uncertainty enters valuation. Instead of assigning utility only to final wealth, the theory assigns value to changes from a psychologically relevant baseline. Instead of multiplying these values directly by objective probabilities, it uses decision weights. These modifications were intended to describe systematic choice patterns, not random mistakes around an otherwise unchanged model.
The theory developed alongside broader challenges to fully rational choice, including bounded rationality, but it is not simply a claim that people have limited computational ability. It provides a structured mathematical account of particular departures from conventional choice axioms. Kahneman received the 2002 Nobel Memorial Prize in Economic Sciences for integrating psychological research into economics, especially concerning judgment and decision-making under uncertainty. Tversky, who died in 1996, was not a recipient.
Prospects, reference points, and the editing of choices
A prospect is a set of possible outcomes together with their probabilities. A simple example offers a gain of 100 with probability and no change otherwise. Prospect theory distinguishes the objective description of such an option from the psychological representation used to evaluate it. The same final monetary position can be experienced differently depending on whether it follows a gain, a loss, or no departure from an expected outcome.
A reference point separates gains from losses. It may be current wealth, an existing possession, a purchase price, an earnings target, or an expectation about what will occur. If an outcome is and the reference point is , the relevant change is
Reference dependence therefore concerns the origin of the value scale, not merely its curvature. A payment of 900 may be a gain relative to an expectation of 800 but a loss relative to an expectation of 1,000. The final payment is identical; its classification differs.
The original 1979 theory distinguished an editing phase from an evaluation phase. Editing organizes and simplifies the available prospects before values and decision weights are applied. Proposed operations included coding outcomes as gains or losses, combining probabilities attached to identical outcomes, separating a sure component from a risky component, and canceling components shared by the alternatives. Rounding and the detection of clearly dominated options could also simplify the representation.
For example, a prospect offering 100 with probability , 100 with probability , and zero otherwise can be represented as a single 50 percent chance of receiving 100. More consequentially, two alternatives with a common initial stage may be compared mainly through the stage at which they differ. This “isolation effect” can make a multistage problem psychologically different from an equivalent problem described directly in terms of final probabilities.
Editing should not be understood as a completely specified computational procedure with a fixed sequence of operations. The original account identified psychologically plausible transformations but did not give an algorithm that uniquely determines the representation of every problem. Much subsequent research has therefore focused on how reference points and choice representations are formed.
Reference points need not be fixed or externally supplied. Expectations can change with experience, and repeated gains or losses can gradually become incorporated into a new baseline. Later reference-dependent models, including work by Botond Kőszegi and Matthew Rabin, made expectations central to reference-point formation. These models extend the broader research program rather than merely restating the original theory. The distinction matters because different assumptions about the baseline can produce different predictions even when the value function is unchanged.
The value function: diminishing sensitivity and loss aversion
Prospect theory’s value function is normally drawn around a reference point at which value is zero. It is increasing in outcomes, usually concave for gains, usually convex for losses, and steeper on the loss side near the reference point. These features describe three distinct properties: larger outcomes are preferred, sensitivity decreases with distance from the baseline, and losses carry disproportionate weight relative to comparable gains.
Diminishing sensitivity means that a fixed increment has less psychological impact when it occurs farther from the reference point. The difference between gains of 100 and 200 may matter more than the difference between gains of 1,100 and 1,200. On the loss side, the difference between losing 100 and losing 200 may matter more than the difference between losing 1,100 and losing 1,200. This produces the familiar S-shaped value function, with its change in curvature at the reference point.
A widely used specification is
where describe curvature and scales the loss branch. With equal exponents, means that a loss has greater absolute value than an equally large gain. In Tversky and Kahneman’s 1992 experimental estimates, the reported median curvature parameters were , and the median loss-aversion parameter was . These estimates are frequently used as illustrations, but they are not universal psychological constants.
Loss aversion is different from risk aversion. Risk aversion compares a lottery with a sure amount, whereas loss aversion compares the treatment of negative and positive changes around a reference point. A person may reject a mixed gamble because its possible loss is heavily weighted while also seeking risk among alternatives consisting entirely of losses. Consequently, neither the statement that people dislike risk nor the statement that losses matter more is sufficient to characterize the theory.
For a symmetric gamble offering a gain or a loss , each with probability one-half, equal decision weights and linear value on each side would make acceptance depend on whether exceeds . Curvature and unequal weighting complicate that threshold. Such calculations illustrate why an observed rejection cannot identify loss aversion without assumptions about the rest of the model.
The geometry also requires care. With power exponents below one, derivatives can become unbounded as outcomes approach zero, so “steeper at the reference point” is best interpreted through comparisons of equally sized gains and losses or an appropriately defined local measure. Across specifications, the numerical meaning of a loss-aversion coefficient depends on normalization, curvature, and the monetary scale used.
Probability weighting and the special status of certainty
Prospect theory separates an event’s probability from its influence on choice. A decision weight is not necessarily a mistaken belief about how likely an event is. Someone may correctly state that a prize has a one percent chance of occurring yet give the possibility disproportionate influence when deciding whether to buy a ticket. Probability weighting concerns the evaluation of uncertainty, although distorted beliefs can independently affect real decisions.
A common empirical pattern is an inverse-S-shaped weighting function. Small probabilities receive more weight than their numerical size would suggest, while moderate and high probabilities receive less. Sensitivity is especially marked near impossibility and certainty. Changing a probability from zero to a small positive number may have a larger effect than an equally sized change between two intermediate probabilities; similarly, eliminating the last remaining possibility of an adverse event can be especially influential.
The certainty effect is a related choice pattern: an outcome that is certain can be disproportionately attractive relative to one that is merely very likely. This helps explain common-ratio reversals. A person may prefer a sure payment of 3,000 to an 80 percent chance of 4,000, yet prefer a 20 percent chance of 4,000 to a 25 percent chance of 3,000. Expected utility, with the same zero payoff otherwise, does not permit this particular reversal because the second comparison multiplies both relevant probabilities in the first comparison by the same factor.
In their 1992 formulation, Tversky and Kahneman used a weighting function of the form
with separate estimated functions for gains and losses. Over empirically relevant parameter ranges, values of below one can generate the characteristic inverse-S pattern. Other functional forms are also common. Drazen Prelec’s later formulation, for example, allows the curvature and elevation of probability weighting to be represented in a different way. No single functional form defines prospect theory.
Decision weights in the original theory need not sum to one. This is one reason they should not simply be interpreted as subjective probabilities. In cumulative prospect theory, weighting is applied to cumulative or tail probabilities, and differences between transformed probabilities determine the weights assigned to outcomes. The distinction between these formulations is mathematically important, particularly for prospects with many possible outcomes.
Probability weighting is also sensitive to information format. Small-probability events may receive less influence when people learn about them through limited experience rather than read an explicit description. Rare events can be missed in small samples, and recent observations can dominate attention. This qualification limits any claim that people invariably overweight rare events: the characteristic pattern is most closely associated with many description-based choice tasks.
Cumulative prospect theory and rank-dependent evaluation
The original theory was developed chiefly for relatively simple prospects. Applying a nonlinear transformation independently to every outcome probability can create difficulties when a prospect contains many outcomes, including violations of stochastic dominance. First-order stochastic dominance expresses a basic requirement: if one prospect yields at least as high an outcome at every corresponding probability position, and a higher outcome somewhere, it should not be ranked below the other by someone who prefers more to less.
Cumulative prospect theory, introduced by Tversky and Kahneman in 1992, addressed this problem by combining reference-dependent valuation with ideas from rank-dependent utility. Instead of weighting each individual probability in isolation, it transforms cumulative probabilities. An outcome’s decision weight consequently depends on its rank among the possible outcomes, not just on the probability attached to it.
Consider ordered outcomes
with probabilities . Losses lie below the reference point and gains at or above it. For gains, define the upper-tail probabilities . The decision weight for gain is
where . For losses, define lower-tail probabilities , with . Their weights are
The overall valuation is
The two weighting functions, and , can differ. Both are normally increasing and satisfy and .
For a simple all-gain example, suppose a prospect pays 100 with probability , 20 with probability , and zero otherwise. The weight on 100 is , while the weight on 20 is . The second weight is not generally . This difference shows how rank enters the model and why probability weighting is not merely a separate adjustment to each listed probability.
Under the appropriate monotonicity conditions, cumulative weighting preserves first-order stochastic dominance while retaining nonlinear responses to probability. It also accommodates prospects with arbitrary finite numbers of outcomes and has continuous-outcome extensions. The 1992 treatment extended the framework beyond risk with known probabilities to uncertainty represented through nonadditive weighting of events. That extension is related to, but distinct from, the empirical question of how people form beliefs when objective probabilities are unavailable.
The fourfold pattern, framing, and preference reversals
The interaction of value-function curvature and probability weighting produces the fourfold pattern of risk attitudes. For moderate or high probabilities, people often show risk aversion over gains and risk seeking over losses. For low probabilities, the pattern often reverses: risk seeking over gains and risk aversion over losses. These are characteristic tendencies under particular conditions, not claims that every individual displays all four patterns in every task.
For a highly probable gain, a certain alternative can be attractive because of both diminishing sensitivity and the special weight attached to certainty. For a highly probable loss, accepting a gamble may preserve a chance of avoiding the loss altogether. At low probabilities, overweighting a large gain can make a lottery attractive, while overweighting a large loss can make insurance attractive. Thus gambling and insurance need not require a person to possess an inconsistent global attitude toward risk.
The familiar reflection effect concerns a change in risk preference when positive outcomes are replaced by corresponding negative outcomes. A preference for a certain gain over a risky gain can coexist with a preference for a risky loss over a certain loss. However, reflection does not follow from loss aversion alone. In comparisons containing only gains or only losses, the relative slope between the two branches need not determine the ranking. Curvature and probability weights remain essential.
Framing effects concern changes in choice produced by different descriptions of equivalent outcomes. In Tversky and Kahneman’s 1981 “Asian disease” problem, respondents considered programs responding to a hypothetical outbreak expected to kill 600 people. In one group, 72 percent preferred saving 200 people for certain over a one-third chance of saving all 600 and a two-thirds chance of saving none. In another group, 78 percent preferred the equivalent risky option when the alternatives were described in terms of deaths: 400 deaths for certain versus a one-third chance of no deaths and a two-thirds chance of 600 deaths.
The experiment illustrates how equivalent consequences can be coded as gains or losses. It does not establish that every framing effect is generated by prospect theory’s value function. Wording can also change perceived implications, attention, or interpretation of what information the communicator intends to convey. A framing result therefore provides evidence about representation without uniquely identifying all underlying mechanisms.
Preference reversals can also arise when the response procedure changes. Pricing a gamble and choosing between gambles may direct attention toward different attributes. Prospect theory offers tools for analyzing some such reversals, especially those involving reference points or probability weighting, but it does not by itself explain every inconsistency between valuation and choice.
Experimental evidence and parameter estimation
Empirical tests of prospect theory use several methods. Binary-choice experiments ask participants to choose between lotteries. Certainty-equivalent tasks identify the guaranteed payment considered equivalent to a risky prospect. Other studies elicit acceptable mixed gambles, willingness to pay for objects, or choices under changing reference points. Each method identifies different combinations of model components, making experimental design central to interpretation.
Estimating curvature from gain-only choices does not establish loss aversion, because no losses enter those choices. Conversely, rejection of a mixed gamble can reflect loss aversion, value curvature, probability weighting, background risk, or misunderstanding. Reliable estimation therefore requires multiple tasks and explicit assumptions about the reference point. A gain-only task and a loss-only task may help identify curvature, while mixed prospects help identify the relative valuation of gains and losses.
Modern estimation commonly combines the deterministic valuation formula with a stochastic choice rule. For example, a logistic rule may represent the probability of choosing prospect over as
where controls sensitivity to value differences. This is an additional modeling assumption, not a component fixed by prospect theory itself. Different noise specifications can lead to different parameter estimates, especially when choice differences are small.
Large replication projects have tested classic findings across populations and locations. The 2020 study “Replicating Patterns of Prospect Theory for Decision under Risk,” by Kai Ruggeri and colleagues, examined 4,098 participants across 19 countries and provided broad support for many original choice patterns. Such results establish the recurrence of important regularities, but they do not imply that one numerical parameter vector describes every population, stake size, or decision environment.
Incentives are another issue. Much early evidence used hypothetical outcomes, whereas later experiments often paid participants according to randomly selected decisions. Real payments can improve engagement, but they do not automatically solve every design problem. A procedure that randomly selects one decision for payment introduces an additional compound lottery, and its theoretical neutrality depends on assumptions about how participants evaluate that procedure.
Evidence also varies between description and experience, laboratory and field, and individual and aggregate data. Repeated feedback may alter beliefs, reference points, or attention. Aggregating people with different parameters can produce a group pattern that describes no particular individual. Consequently, experimental support is best evaluated component by component: reference dependence, diminishing sensitivity, loss aversion, and probability weighting have related but distinct evidential bases.
Applications to finance and insurance
In financial markets, prospect theory helps analyze why purchase prices, past performance, and the frequency of evaluation may affect investment decisions. Conventional portfolio models generally emphasize distributions of final wealth or consumption. A reference-dependent investor may instead evaluate a position relative to its purchase price or a recent account balance. The same asset can then appear as a gain to one holder and a loss to another.
A prominent application is the disposition effect, the tendency to realize gains more readily than losses. Hersh Shefrin and Meir Statman discussed the phenomenon in 1985, and Terrance Odean documented it in brokerage-account data in 1998. If investors evaluate each holding separately relative to its purchase price, concavity over gains and convexity over losses can contribute to a willingness to sell winners and retain losers. However, a static value function does not automatically generate the full dynamic trading pattern. Beliefs about future returns, taxes, transaction costs, and whether a sale changes the reference point also matter.
Mental accounting is especially important in these applications. Investors may evaluate individual holdings separately rather than integrate all risks into total wealth. This narrow evaluation can make losses salient even when a diversified portfolio has gained overall. Prospect theory specifies valuation conditional on a representation; mental accounting helps explain which outcomes are grouped together for that valuation.
Shlomo Benartzi and Richard Thaler’s 1995 account of myopic loss aversion combined loss aversion with frequent evaluation of returns to address the equity premium puzzle. Equities have historically offered higher average returns than relatively safe assets, yet their short-horizon returns frequently include losses. Investors who evaluate portfolios often and react strongly to losses can demand substantial compensation for holding equities. The explanation depends on the evaluation horizon and on what is included in the relevant account; it is not a parameter-free implication of prospect theory.
Probability weighting also informs models of demand for positively skewed assets with a small chance of a large payoff. Some investors may value such “lottery-like” returns disproportionately. In equilibrium models, this demand can affect prices and expected returns, but linking individual weighting functions to market prices requires assumptions about investor heterogeneity, trading constraints, and arbitrage.
Insurance provides a complementary application. Overweighting a small chance of a large loss can increase willingness to pay for protection. Yet insurance demand also depends on wealth, contractual exclusions, trust, liquidity, and the perceived probability of a claim. Prospect theory does not imply that all small-probability risks will be insured. Underinsurance against rare disasters and the purchase of low-deductible coverage for smaller risks show why salience, learning, and the construction of reference points must accompany the basic valuation model.
Consumer behavior, work, and public policy
Reference dependence has been applied to consumer demand by treating an expected or familiar price as a baseline. A price above that baseline can be coded as a loss, while an equally sized reduction is coded as a gain. With loss aversion, the price increase can have a larger effect on perceived value than the price decrease. Such models help organize evidence on asymmetric reactions to discounts and surcharges, although ordinary substitution effects, quality inferences, and search costs can also generate asymmetries.
The endowment effect is another important application. Experiments have often found that people require more money to surrender an object they possess than they would pay to acquire it. Daniel Kahneman, Jack Knetsch, and Richard Thaler’s 1990 experiments with mugs became a prominent example. Ownership can establish a reference point, making relinquishment a loss. Nevertheless, the observed gap between willingness to accept and willingness to pay can also depend on elicitation procedures, market experience, transaction costs, and participants’ understanding of the task.
Labor-supply models sometimes use a daily income target as the reference point. A worker below the target may place greater value on additional earnings than one who has already exceeded it. In such a model, unexpectedly high earnings per hour can lead the worker to stop earlier, rather than work longer. The 1997 study by Colin Camerer and colleagues on New York City taxi drivers helped make this application influential. Subsequent work produced differing results and emphasized identification problems, including the joint determination of hours, earnings opportunities, and the decision to end a shift.
In public communication, gain and loss frames have been studied in areas such as conservation, taxation, and public health. A policy can be described as preserving an existing benefit or avoiding a deterioration. Prospect theory predicts that descriptions may influence evaluation when they change the perceived reference point, but it does not establish that loss framing is always more effective. Effects depend on the action, the outcome measure, the population, and whether people perceive the alternatives as risky.
The theory also contributes to choice architecture, particularly where defaults or existing arrangements become reference points. Remaining in a current plan may avoid a perceived loss even when an alternative offers advantages. However, default effects may also arise from inertia, switching costs, limited attention, or an inference that the default is recommended. Reference dependence is one possible mechanism rather than a complete explanation.
These applications raise a distinct issue for welfare economics. If equivalent descriptions produce different choices, observed choice alone may not reveal a unique underlying welfare ranking. Evaluating a policy then requires assumptions about experienced outcomes, informed preferences, or which reference points should count. Prospect theory describes how a frame can change choice; it does not independently determine which frame yields the normatively correct decision.
Limitations, competing explanations, and unresolved questions
A central unresolved issue is the determination of the reference point. Current wealth, expectations, aspirations, purchase prices, and social comparisons can all be plausible baselines. If the baseline is selected only after observing a choice, the theory can become difficult to test. Stronger applications specify how the reference point is formed before deriving predictions, or manipulate it independently while holding objective outcomes constant.
The scope and magnitude of loss aversion remain subjects of empirical debate. Many studies find asymmetries between gains and losses, but estimates vary with stakes, procedures, experience, and population. Some tasks produce little loss aversion or patterns better explained by other mechanisms. Moreover, an asymmetry in attention or emotional response is not automatically the same as an asymmetry in the value function governing choice. The claim that losses universally matter about twice as much as gains is therefore stronger than the evidence warrants.
Model identification is another difficulty. Curvature, probability weighting, reference points, and stochastic errors can compensate for one another when fitted to limited data. A flexible model may explain observed choices well without predicting new choices accurately. Comparisons with expected utility, rank-dependent models, regret-based theories, and other behavioral models therefore benefit from out-of-sample prediction and experimental tasks designed to separate mechanisms.
Prospect theory also distinguishes imperfectly between decisions made from explicit descriptions and those learned through experience. Description-based weighting functions do not automatically transfer to environments involving sampling, memory, and changing beliefs. Under genuine uncertainty, subjective probabilities may be unknown to the researcher as well as to the decision-maker. An apparent weighting effect can then reflect beliefs, valuation, or both.
Dynamic choice presents additional problems. Reference points can adapt after outcomes occur, and a plan chosen in advance may no longer be preferred when a later decision is reached. Repeated opportunities to accept attractive-looking gambles can produce strategies very different from a one-shot evaluation. Modeling these settings requires assumptions about commitment, anticipation of future preferences, the grouping of outcomes over time, and how realized gains and losses enter the baseline.
Finally, prospect theory is not a comprehensive theory of human decision-making. It does not by itself explain beliefs, strategic interaction, social preferences, attention, or learning. Nor does it replace normative analysis: a descriptive account of why someone chooses an option does not establish that the option improves that person’s welfare. Contemporary research often treats its components as separable hypotheses, testing when reference dependence, loss aversion, and nonlinear weighting are necessary and when simpler or alternative mechanisms better account for the evidence.