Kelly criterion
Also called Kelly formula, Kelly bet, fractional Kelly
J. L. Kelly Jr.'s 1956 formula for the fraction of capital to stake on a favorable bet for the fastest long-run growth. Full Kelly brings large swings; staking a fraction of it trades some growth for safety.

On this page· 5 sections
Staking everything on every favorable bet maximizes the expected value of capital. In Kelly’s setup, the same gambler is broke with probability one if the betting goes on indefinitely[1]. The Kelly criterion finds the stake in between: large enough to grow, small enough to survive.
J. L. Kelly Jr. published it in the Bell System Technical Journal in July 1956, in a paper on information rate and gambling on a noisy communication channel[1].
How it works
Suppose the gambler stakes the same fraction of capital on every bet. After N bets with W wins and L losses, capital is the starting amount times (1 + f)^W times (1 − f)^L, for an even-money stake f. Kelly measures the exponential rate of growth of capital, G, and picks the fraction that makes G as large as possible[1].
For an even-money bet that wins with probability q, the answer is f = 2q − 1[1]. For other payoffs, the Kelly wager is the edge divided by the odds, where the edge is the expected return per dollar staked[2].
The reward is long-run growth. Kelly showed that in a game that never ends, a gambler using this fraction will, with probability one, eventually get ahead of one using any other fraction, and stay ahead[1]. The cost is short-term risk, because the suggested wagers may be very large[2].
Betting less than Kelly is called fractional Kelly, a blend of Kelly and cash[2]. In the authors’ table for blackjack with a 2% advantage (p = 0.51), betting half of Kelly gave a 0.89 probability of doubling before halving, against 0.67 at full Kelly, with 75% of the growth rate. At double Kelly the growth rate fell to zero[2].
Example
Say a bet pays even money and wins 55% of the time. Then f = 2 × 0.55 − 1 = 0.10, so the Kelly stake is 10% of capital: $1,000 from $10,000.
- Full Kelly, 10%: long-run growth of about 0.50% per bet.
- Half Kelly, 5% ($500): about 0.38% per bet, or 75% of the full-Kelly rate.
- Double Kelly, 20%: about −0.01% per bet. The extra risk buys no growth.
- After five losses in a row, full Kelly leaves 0.9⁵ = 59.0% of capital, and half Kelly leaves 0.95⁵ = 77.4%.
Watch out
The formula needs the true edge. MacLean, Thorp and Ziemba call Kelly calculations extremely sensitive to errors in the estimated mean, and advise that wagers be reduced to be on the safe side[2]. In investing, realized returns may differ greatly from estimates drawn from history[2].
Kelly is not a sure thing. Like any other rule, it can produce a poor result over a finite run of bets[2].
Staking more than Kelly adds risk and lowers growth[2].
A coin toss is a poor model of a market. There, win rates and payoffs are estimates from history, not known odds[2].
Sources
- ResearchBell System Technical Journal (copy hosted by Princeton University)A New Interpretation of Information Rate (J. L. Kelly Jr., Bell System Technical Journal, July 1956)Checked November 9, 2026
- ResearchL. C. MacLean, E. O. Thorp, W. T. Ziemba (copy hosted by UC Berkeley Statistics)Good and bad properties of the Kelly criterion (MacLean, Thorp, Ziemba, 2010)Checked November 9, 2026
