Risk aversion is a preference for avoiding uncertainty about outcomes when doing so does not change their expected value. A risk-averse person prefers receiving a fixed amount of money to entering a gamble whose average monetary payoff is the same amount. The concept does not imply that all risky activities are rejected: a sufficiently favorable expected return, a low price, or protection against other risks can make an uncertain prospect attractive.
In economics, risk aversion helps explain the purchase of insurance, the diversification of investments, compensation for hazardous employment, and the allocation of uncertain income through contracts and public institutions. Its standard mathematical treatment connects preferences over uncertain prospects to the curvature of a utility function. Research also distinguishes risk aversion from mistaken beliefs, dislike of unknown probabilities, and special sensitivity to losses.
Definition and the expected-utility framework
Consider a person choosing between a guaranteed payment of 100 monetary units and a lottery paying either 50 or 150, each with probability one-half. Both options have an expected monetary value of 100. A risk-averse person weakly prefers the guaranteed payment; a strictly risk-averse person strictly prefers it when the lottery has genuine uncertainty. A risk-neutral person is indifferent, while a risk-seeking person prefers the lottery. These classifications concern preferences over a specified range of outcomes, rather than necessarily describing a person's behavior in every setting.
The standard framework is expected utility theory. Let denote uncertain final wealth and let be an increasing utility function. Expected utility is
Risk aversion requires that, for every relevant lottery,
Under the usual regularity conditions, this is equivalent to concavity of . If utility is twice differentiable, concavity means . Economically, additional wealth produces progressively smaller utility gains: losing 50 from an initial wealth of 100 reduces utility more than gaining 50 increases it. This is the connection between risk aversion and diminishing marginal utility.
The inequality is an application of Jensen's inequality. Strict concavity produces a strict preference for certainty whenever the lottery is nonconstant and lies within the utility function's domain. An affine utility function represents risk neutrality. Expected-utility functions may be replaced by positive affine transformations without changing the preferences they represent: , with , yields the same choices. By contrast, an arbitrary increasing transformation generally changes rankings of lotteries.
A broader distributional interpretation uses a mean-preserving spread: a change that increases dispersion while leaving the mean unchanged. Every expected-utility maximizer with an increasing concave utility function weakly dislikes such a change. This provides a comparison that does not depend on one particular utility specification. It also shows why risk cannot always be identified with variance alone. Two distributions can have equal variances but different downside tails, and a higher variance by itself does not establish that every risk-averse person will prefer the alternative. The full distribution of outcomes can matter.
Historical development
The analysis of risk aversion developed from problems in probability, gambling, and insurance. The St. Petersburg paradox was particularly influential. In its familiar version, a fair coin is tossed until the first head, and the payoff doubles with each additional toss required. The game's mathematical expected monetary payoff is infinite, yet ordinary valuations of participation are finite. The discrepancy challenged the idea that uncertain prospects should always be evaluated by their average monetary return.
In 1738, Daniel Bernoulli proposed that the utility of an additional monetary gain depends on the recipient's existing resources. His logarithmic utility formulation made the expected utility of the St. Petersburg gamble finite for a fixed positive initial wealth. Gabriel Cramer had independently advanced related reasoning in correspondence in 1728. Bernoulli's contribution established a durable distinction between money and the satisfaction associated with it, although logarithmic utility did not become a universal empirical law and does not eliminate every possible version of an unbounded-payoff paradox.
A systematic axiomatic foundation came with John von Neumann and Oskar Morgenstern's Theory of Games and Economic Behavior, published in 1944. They showed how preferences satisfying conditions including completeness, transitivity, continuity, and independence can be represented by expected utility. Their framework made risk aversion a property of a representation derived from preferences rather than simply an assumed psychological attitude. Leonard J. Savage's The Foundations of Statistics in 1954 extended the analysis to choices involving subjective probabilities.
Modern measurement emerged in work by John W. Pratt and Kenneth Arrow. Pratt's 1964 article “Risk Aversion in the Small and in the Large,” together with Arrow's work in the 1960s, connected utility curvature to observable willingness to bear risk. Their measures made it possible to distinguish absolute from relative risk aversion and to compare people independently of arbitrary utility scales. Related developments included Harry Markowitz's 1952 portfolio analysis and the study of optimal insurance and risk-bearing contracts.
These developments did not eliminate alternative theories. Maurice Allais's 1953 challenge to expected utility and Daniel Ellsberg's 1961 experiments demonstrated systematic choice patterns that its simplest interpretation could not accommodate. Later research preserved expected utility as a central benchmark while investigating probability weighting, reference dependence, incomplete information, and uncertainty about the probability model itself.
Certainty equivalents and measures of risk aversion
The certainty equivalent converts an uncertain prospect into a comparable sure amount. For uncertain final wealth , its certainty equivalent satisfies
If utility is increasing, the certainty equivalent is uniquely determined whenever the expected utility lies in its range. A risk-averse person has . The difference
is the risk premium for that lottery: the amount of expected wealth the person would surrender to replace it with certainty. This individual valuation differs from a market risk premium, which usually refers to an expected return above a benchmark return.
The Arrow–Pratt coefficient of absolute risk aversion is
It describes sensitivity to small monetary risks around wealth . Its units are the inverse of the units in which wealth is measured. Relative risk aversion is
which is dimensionless and describes sensitivity to risks measured as proportions of wealth. Both coefficients remain unchanged under positive affine transformations of utility.
For a small, mean-zero monetary risk added to wealth , a second-order approximation gives
This relationship makes the role of curvature concrete. If absolute risk aversion at a particular wealth level is per monetary unit and a sufficiently small gamble has variance , the approximate risk premium is monetary units. The approximation becomes less reliable when risks are large relative to wealth, utility changes rapidly across possible outcomes, or extreme outcomes dominate valuation.
Arrow–Pratt measures also support comparisons “in the large.” Under suitable smoothness and common-domain conditions, one person is everywhere more risk averse than another if the first person's utility is an increasing concave transformation of the second person's utility. Equivalently, the first has weakly higher absolute risk aversion at every wealth level. This implies weakly lower certainty equivalents for all admissible lotteries, not merely greater reluctance toward a single observed gamble.
Changing wealth introduces another distinction. Decreasing absolute risk aversion means falls as wealth rises; decreasing relative risk aversion means falls. These properties describe different comparisons. A person can become more willing to accept a fixed monetary gamble as wealth increases while remaining equally reluctant to expose a fixed percentage of wealth.
Common utility specifications and their implications
Several utility families are widely used because they make risk preferences tractable. Constant absolute risk aversion, or CARA, is commonly represented by exponential utility:
Here , regardless of wealth. An independent addition to certain wealth does not change the ranking of fixed monetary gambles. With unconstrained portfolio choice and an independent investment opportunity, this often implies that the optimal monetary amount invested in the risky asset is independent of initial wealth. CARA utility is defined for all real wealth, making it convenient when normally distributed payoffs can include negative outcomes.
Exponential utility has a particularly simple relationship with normal distributions. If final wealth is normally distributed with mean and variance , the certainty equivalent is exactly
Outside this setting, mean and variance generally do not fully determine expected utility. The formula's convenience therefore depends on both the utility specification and the distributional assumption.
Constant relative risk aversion, or CRRA, is represented for positive wealth by
with the logarithmic case
when . Relative risk aversion equals , while absolute risk aversion equals . Fixed monetary risks become less consequential as wealth increases, but proportional risks are evaluated consistently across wealth levels. In standard investment models without important constraints or nonfinancial income, this structure can produce a constant optimal fraction of wealth in risky assets.
Quadratic utility, often written for , provides another route to mean–variance analysis. Expected utility depends only on the mean and variance of wealth. However, marginal utility becomes negative above , so the specification is economically plausible only on a restricted domain. Within its increasing region it also exhibits increasing absolute risk aversion, which is often an unattractive description of household investment behavior.
These forms are assumptions about preferences, not interchangeable definitions of risk aversion. Their usefulness depends on whether the object being evaluated is consumption, final wealth, or a narrowly framed payment. Applying a utility function to one experimental payoff instead of total resources can yield very different parameter estimates. Likewise, a specification appropriate for moderate positive consumption changes may behave implausibly near zero consumption or under catastrophic losses. Domain restrictions are part of the model rather than merely technical details.
Insurance, contracts, and the sharing of risk
Insurance illustrates how risk aversion can create gains from trade without increasing expected aggregate resources. Suppose wealth is , a loss occurs with probability , and coverage costs a premium . Final wealth is without a loss and with a loss. At full coverage, , wealth is in both states. If premiums are actuarially fair, coverage is unrestricted, and no other relevant distortions are present, a strictly risk-averse person prefers full insurance.
The insurer need not be indifferent to risk for this exchange to occur. Pooling many approximately independent policies reduces the variability of average claims. However, pooling does not eliminate correlated losses from earthquakes, epidemics, or widespread financial disruption. Capital, reinsurance, and access to broader investors help determine how much risk an insurer can absorb and at what price.
Actual contracts usually depart from the fair-price benchmark. Administrative costs, capital requirements, and markups raise premiums above expected payouts. With loaded premiums, partial coverage may be optimal. A deductible leaves small losses with the policyholder while insuring larger ones. Moral hazard can also make full coverage inefficient if insurance weakens incentives to prevent losses or control claims. Adverse selection arises when buyers know more about their risk than insurers do, complicating pricing and the availability of coverage. Observing incomplete insurance therefore does not establish that people are risk neutral.
Risk aversion also shapes contract theory. In a standard principal–agent problem, a risk-neutral employer would prefer to insure a risk-averse employee against random output fluctuations if effort could be directly observed and contracted upon. When effort is hidden, making pay depend on output provides incentives but exposes the employee to risk. Optimal compensation balances incentive provision against the cost of risk-bearing. Fixed salaries, performance bonuses, and commissions allocate these burdens differently.
Efficient risk sharing can be expressed through marginal utilities. With complete contingent contracts and fixed welfare weights , an interior efficient allocation satisfies
for each person in each state , where is the shadow value of resources in that state. The condition does not require identical consumption. It requires that resources be allocated so their weighted marginal value is equalized. Credit constraints, limited enforcement, private information, and missing markets prevent this benchmark from being fully realized.
Portfolio choice and the pricing of financial risk
In financial markets, risk aversion helps determine both individual portfolios and equilibrium prices. A risky investment must be evaluated not only by its own payoff variability but also by how its payoff interacts with the investor's other resources. An investment that performs well when labor income or other assets perform poorly may be desirable even if it is highly volatile in isolation. Conversely, a relatively stable security can be unattractive if its losses occur precisely when funds are most valuable.
Portfolio diversification exploits imperfect correlation among returns. Combining investments can reduce portfolio risk without a proportionate reduction in expected return. Firm-specific risks may largely disappear in a broad portfolio, while economy-wide risks remain. This distinction underlies the capital asset pricing model, in which expected excess returns depend on covariance with the market portfolio rather than on each security's total variance. Risk aversion helps determine the aggregate reward required for bearing market risk, while the model's covariance structure determines how that reward is distributed across securities.
A simple one-period example makes the role of curvature explicit. Let a risky investment offer an excess payoff per unit invested, with mean and variance . If monetary units are invested, a local approximation to its certainty-equivalent gain is
Under the approximation and without binding constraints, the optimum is
Higher expected excess payoff raises the desired holding; greater variance or absolute risk aversion lowers it. Correlated background income, borrowing restrictions, transaction costs, and large risks alter this expression.
More generally, an investor who optimally chooses consumption and assets satisfies an asset-pricing condition of the form
where is a gross return and is the stochastic discount factor. In a standard time-additive consumption model,
Payoffs received when consumption is low have especially high value because marginal utility is high. An asset that pays mainly in prosperous states generally needs a higher expected return than one that protects consumption in adverse states.
Risk aversion alone does not determine market prices. Expectations, available securities, borrowing capacity, heterogeneous preferences, and institutional constraints also matter. A rise in observed market risk premiums can reflect greater expected volatility, tighter intermediary balance sheets, or changing beliefs rather than a change in underlying preferences.
Saving, background risk, and decisions over time
Risk aversion interacts with time because households choose consumption and saving before future income is known. A person facing uncertain earnings may accumulate a buffer of assets, but this precautionary saving motive is not equivalent to risk aversion. In a standard two-period model, the relevant curvature condition concerns marginal utility: prudence is associated with . Its local coefficient is
Risk aversion concerns a preference for smoothing consumption across uncertain states; prudence concerns how uncertainty changes the incentive to transfer resources into the future.
The distinction can be seen in the saving condition
where is a certain gross return. If marginal utility is convex, greater future consumption risk can raise expected marginal utility and encourage more saving, under the model's other conditions. Concavity of utility alone does not guarantee this response. A utility function may represent risk aversion without generating a positive precautionary saving motive.
Background risk also affects willingness to accept new risks. Uninsurable labor income, uncertain medical spending, or exposure to a family business can make a financial gamble more burdensome. If background income is correlated with asset returns, the interaction is especially direct: an employee whose earnings depend on one industry may already have substantial exposure to it before purchasing any securities. Even statistically independent background risk can alter desired risk-taking because it changes the distribution of marginal utility. The direction of this effect requires additional restrictions on preferences; it does not follow from risk aversion alone.
Risk preferences must also be distinguished from willingness to substitute consumption between dates. In the standard time-additive CRRA model, the elasticity of intertemporal substitution is , mechanically tying it to relative risk aversion. These are conceptually different attitudes: one concerns variation across uncertain states, and the other concerns variation across time. Epstein–Zin preferences, developed in 1989, provide a recursive framework that separates them.
This distinction matters for macroeconomic models and asset pricing. The equity premium puzzle, formulated by Rajnish Mehra and Edward Prescott in 1985, concerns the difficulty of reconciling the historically high average return on equities relative to relatively safe assets with standard consumption-based models using conventional parameter values. Proposed explanations include rare disasters, habit formation, incomplete markets, heterogeneous households, and alternative preference structures. A high observed equity premium does not by itself uniquely identify a high coefficient of risk aversion.
Measurement and empirical evidence
Risk aversion is not directly observable. Researchers infer it from choices, conditional on assumptions about beliefs, resources, and the decision environment. Laboratory experiments commonly offer choices between a relatively safe lottery and a riskier lottery, varying their probabilities or payments. The point at which a participant switches identifies a range of risk-aversion parameters within a chosen utility family. Multiple price lists, especially the design associated with Charles Holt and Susan Laury's 2002 study, became widely used because they are comparatively straightforward to administer and interpret.
Experimental results depend on implementation. Real rather than hypothetical payments, the scale of incentives, the number of decisions, and the procedure for choosing which decision is paid can affect behavior. Participants may misunderstand probabilities or switch inconsistently. An analyst must decide whether such choices reflect random error, inattention, probability weighting, or preferences outside the assumed model. A fitted coefficient is therefore jointly an estimate of preferences and a consequence of the statistical specification.
Field evidence uses insurance deductibles, portfolio holdings, occupational choices, farming decisions, and participation in risky markets. These choices often involve economically substantial stakes, but they also contain confounding influences. A household may hold few equities because of limited financial access, participation costs, distrust, or insufficient liquid savings. A high insurance deductible may reflect liquidity constraints rather than weak dislike of risk. Agricultural choices combine risk preferences with expectations about weather, crop prices, and access to credit.
Surveys provide a complementary method. Respondents may report willingness to take risks generally or in domains such as finance, health, and work. Representative surveys, including research using the German Socio-Economic Panel, have connected such reports to experimental choices and observed behavior. However, domain-specific differences remain important. Someone willing to take professional risks may avoid financial gambles, and a single scalar parameter may not summarize both kinds of decision.
Identification is particularly difficult when probabilities are subjective. Rejecting an investment could indicate pessimistic beliefs rather than strong risk aversion. Reliable inference therefore benefits from separately eliciting beliefs, observing repeated decisions, and accounting for background wealth and income. Reported relationships between risk-taking and age, wealth, education, or other characteristics are not automatically causal. Changes in resources, experience, selection into occupations, and the measurement instrument can all contribute to an observed pattern. Experimental estimates from small stakes should consequently not be transferred mechanically to retirement, business formation, or catastrophic risks.
Behavioral distinctions and public decision-making
Several attitudes that resemble risk aversion have different theoretical meanings. Loss aversion is greater sensitivity to losses than to equal-sized gains relative to a reference point. A person can be loss averse without having a globally concave utility function over final wealth. Prospect theory, introduced by Daniel Kahneman and Amos Tversky in 1979, combines reference-dependent valuation with transformed decision weights. Its value function is commonly concave for gains and convex for losses, allowing risk aversion in some choices and risk-seeking behavior in others.
Ambiguity aversion concerns uncertainty about probabilities rather than merely uncertain outcomes with specified probabilities. In an Ellsberg-style choice, people may prefer betting on an urn with known proportions of colored balls to one with unknown proportions, even when no single subjective probability explains all their choices. Such behavior cannot generally be reduced to increasing the curvature of a conventional expected-utility function. Likewise, mental accounting and narrow framing can cause a person to evaluate a gamble separately from overall wealth, changing its apparent riskiness.
A major theoretical limitation arises when small-stakes choices are interpreted through utility over total wealth. Matthew Rabin's 2000 calibration result showed that rejecting modest favorable gambles over a broad range of wealth can imply implausibly extreme aversion to much larger gambles under expected utility. This is not a demonstration that concave utility is mathematically inconsistent. It shows that a single wealth-based curvature explanation can be quantitatively inadequate when applied across very different stakes.
Public decisions introduce an additional distinction between individual risk aversion and social evaluation. A government project may distribute uncertain benefits and costs across many people and many years. Under restrictive conditions, independent small risks can be spread widely enough that the public risk premium per person becomes small. Kenneth Arrow and Robert Lind's 1970 analysis formalized this argument. It does not establish that governments should ignore correlated disasters, risks borne disproportionately by vulnerable groups, or project outcomes linked to aggregate consumption.
In cost–benefit analysis, risk adjustment must therefore account for who bears uncertainty and when adverse outcomes occur. Climate change, for example, involves long time horizons, uncertain damages, and losses that may be concentrated in already disadvantaged populations. These issues combine risk preferences with discounting, distributional judgments, and uncertainty about the model itself. Similarly, concavity of a social welfare function may express concern about inequality across people rather than an individual's dislike of uncertainty. Keeping these concepts separate is necessary for interpreting what a policy's apparent “risk aversion” actually represents.