Definition
Kelly sizing chooses the fraction of wealth that maximizes expected logarithmic wealth growth under a specified repeated-bet model. For a binary bet with win probability p and net win odds b, the unconstrained full Kelly fraction is (bp minus q) divided by b, where q equals 1 minus p. The result depends on the payoff model and is not a personal allocation instruction.
How it is calculated
Binary Kelly fraction f-star = (b times p minus q) divided by b, where p is win probability, q is loss probability 1 minus p, and b is net profit per unit stake on a win.
| Symbol | Meaning | Unit |
|---|---|---|
| f | fraction of wealth staked | fraction |
| b | net win odds | fraction per fraction staked |
| p | win probability | probability |
| q | loss probability | probability |
| g | expected log growth | dimensionless |
-
Write expected log growth from the win and loss outcomes.
-
Set its first derivative with respect to the stake fraction to zero.
-
Solve for the fraction while retaining the stated domain constraints.
Worked example
Synthetic example.
-
inputs:
-
win probability: 0.6
-
loss probability: 0.4
-
net win odds: 1
-
-
result:
- kelly fraction: 0.2
Interpretation
Kelly balances the growth from winning against the shrinkage from losing on a logarithmic wealth scale. A small change in the estimated edge can move the fraction materially, which is why estimation uncertainty matters. The formula is a model result, not a command to stake a particular amount.
Assumptions and conventions
-
Outcomes are binary with known or specified p and b.
-
For the interior derivation, b>0, 0<p<1 and -1/b<f<1; a loss forfeits the stake, and constraints change the solution.
-
Bet outcomes are repeated under a stable payoff model and wealth is positive.
-
No leverage, ruin, drawdown or estimation-risk constraint is added to the unconstrained formula.
All quantities are dimensionless fractions or probabilities; b is net profit per unit staked.
Limitations and common misreadings
-
p is a probability, not a historical win percentage without a model for future trials.
-
Full Kelly ignores practical risk limits and estimation error unless those are added separately.
References
- J. L. Kelly. (1956). A New Interpretation of Information Rate. Bell System Technical Journal. Sections 2–4. doi:10.1002/j.1538-7305.1956.tb03809.x. Source

