Why Most People Reject a Bet With Positive Expected Value

People reject positive expected value bets for two different reasons. Loss aversion makes a loss feel roughly twice as heavy as an equal gain, which is a bias at small stakes. But a stake that is large relative to wealth can lower expected log growth, so declining it is rational. Kelly sizing separates the two.

The bet most people turn down

Offer someone a fair coin flip: heads they win $200, tails they lose $100. The expected value is plus $50 per flip. Most people still say no. Paul Samuelson described a colleague who refused exactly this bet in his 1963 paper "Risk and Uncertainty: A Fallacy of Large Numbers" in Scientia, while saying he would happily take a hundred of them.

That refusal looks like an error to anyone who has just learned to compute expected value. It is not always one. Whether a positive expected value bet is worth taking depends on three things expected value ignores: how losses feel, how large the stake is relative to wealth, and how sure the bettor is that the edge exists. The first is mostly bias. The second and third are often sound reasoning.

Loss aversion: the behavioural reason

Daniel Kahneman and Amos Tversky set out prospect theory in Econometrica in 1979. Its central claim is that people judge outcomes as gains and losses relative to a reference point, not as final levels of wealth, and that losses loom larger than gains of the same size. In their 1992 follow up, "Advances in Prospect Theory" in the Journal of Risk and Uncertainty, Tversky and Kahneman estimated a median loss aversion coefficient of 2.25 from their experimental subjects, with a value function exponent of 0.88.

Apply a simplified version to the coin flip, ignoring the probability weighting that full cumulative prospect theory adds. With linear value and a coefficient of 2.25, the gain counts 0.5 times 200, which is 100, and the loss counts 0.5 times 2.25 times 100, which is 112.5. The bet scores minus 12.5 and is rejected. Under these parameters a typical person needs the win to be more than about 2.25 times the loss before a 50/50 bet feels acceptable. The 2.25 figure is a median from one set of experiments, not a constant of nature, and later studies report a wide spread across people and contexts.

Utility of wealth: the rational reason

Classical economics explains risk aversion differently. If each extra dollar is worth a little less than the last, a fair bet lowers expected utility and only a sufficiently favourable bet raises it. The usual benchmark is logarithmic utility, because maximising expected log wealth is the same as maximising long run growth.

Under log utility the coin flip has a clean break-even. Accepting is better when the average of ln(W + 200) and ln(W - 100) exceeds ln(W). That holds when (W + 200)(W - 100) exceeds W squared, which simplifies to W above $200. Someone with $1,000 should take the flip. Someone whose entire bankroll is $150 should not, even though the expected value is identical in both cases. The arithmetic is the same; what differs is the share of wealth at risk.

Kelly sizing turns this into a number

The Kelly criterion, from John Kelly's 1956 paper in the Bell System Technical Journal, gives the stake that maximises expected log growth. For this flip the net odds b are 2, the win probability p is 0.5, and the Kelly fraction is (bp minus q) divided by b, which is (1 minus 0.5) divided by 2, or 25 percent of bankroll. A $100 stake is exactly full Kelly at a bankroll of $400.

Growth per flip at stake fraction f is 0.5 ln(1 + 2f) plus 0.5 ln(1 - f). It peaks at f of 0.25 and falls back to zero at f of 0.5, double Kelly. That is the $200 bankroll from the previous section: at twice the Kelly stake, a positive expected value bet produces no growth at all, and beyond it, negative growth. A bet can have positive expected value and still shrink the typical bankroll if it is too large.

Rabin's calibration problem

If utility curvature explained everyday risk aversion, small favourable bets would almost always be accepted, because utility is close to linear over small amounts. Matthew Rabin made this precise in "Risk Aversion and Expected-Utility Theory: A Calibration Theorem", Econometrica, 2000. His headline example: an expected utility maximiser who turns down a 50/50 bet to lose $100 or gain $110 at every wealth level must also turn down a 50/50 bet to lose $1,000 or gain any amount at all. Nobody is actually that risk averse.

Rabin's conclusion was that expected utility is a poor account of attitudes to modest stakes, and he pointed to loss aversion as the better explanation. So the refusal of small positive expected value bets is largely behavioural, while caution about large bets relative to wealth is largely mathematical. Both are real, and they apply at different scales.

Why the reason matters in betting

Sports bettors usually face the reverse problem from Samuelson's colleague: they accept negative expected value bets readily and question positive ones. But the same three factors apply.

First, narrow framing. Judging each bet in isolation magnifies loss aversion. Samuelson's colleague would refuse one flip but accept a hundred, because the aggregate rarely loses. Shlomo Benartzi and Richard Thaler called the investing version myopic loss aversion in 1995. Evaluating a process over hundreds of bets, rather than bet by bet, is the standard correction.

Second, stake size. A positive expected value price is only worth taking at a size the bankroll can carry. Fractional Kelly and hard caps exist because the growth curve punishes overbetting far more than underbetting.

Third, and most important, uncertainty about the edge. In a textbook flip the probability is known. In a betting market it is an estimate, and an estimate that differs from a liquid market price is more often wrong than right. When another party offers a bet, the offer itself is information: they may know something the bettor does not. Declining in that situation is not loss aversion. It is a reasonable prior about adverse selection.

A short test for any refusal

Before treating a refusal as bias or as prudence, ask three questions. Is the probability known, or estimated? If estimated, how has the estimate performed against closing prices or out of sample results? Is the stake below the Kelly fraction implied by a conservative version of that estimate? Would the decision change if this bet were one of a hundred similar ones?

If the probability is well supported, the stake is small, and the answer to the last question is yes, the refusal is probably loss aversion. If the probability is shaky or the stake is large, the refusal is probably correct. This page describes decision mathematics and is not betting advice.

Frequently asked questions

Is rejecting a positive expected value bet irrational?
Not necessarily. At small stakes it is usually loss aversion, a documented bias. At stakes that are large relative to wealth it can be rational, because a big enough stake lowers expected log growth even when expected value is positive. It is also rational when the claimed edge is uncertain, which is the normal case in betting markets.
What is the loss aversion coefficient?
It is the ratio of how heavily a loss is weighted compared with an equal gain. Tversky and Kahneman's 1992 paper on cumulative prospect theory reported a median estimate of 2.25 from their subjects. Later research finds wide variation across people and settings, so it is best read as a typical figure rather than a fixed constant.
How does the Kelly criterion relate to rejecting bets?
Kelly gives the stake that maximises long run growth for a known edge. Staking more than Kelly lowers growth, and staking twice Kelly on a simple binary bet drives expected growth to zero. So a positive expected value bet offered at a fixed stake can be worth declining if that stake is far above the Kelly fraction of the bankroll.
What did Samuelson's colleague get wrong, if anything?
He refused one 50/50 flip paying $200 or costing $100 but said he would accept a hundred of them. Samuelson argued this was inconsistent under expected utility. Behavioural economists later read it as narrow framing: judged alone, one flip feels like a probable loss, while a bundle of a hundred almost never loses money overall.