Kelly Criterion
The Kelly Criterion is a mathematical formula developed by John L. Kelly Jr. (1956) that determines the optimal fraction of capital to allocate to a bet or investment in order to maximize the long-run growth rate of wealth, balancing the trade-off between investing too little (foregone return) and too much (excessive drawdown and ruin risk). For a binary bet with win probability p and win/loss payoffs of b-to-1, the Kelly fraction equals f* = p - (1-p)/b = (bp - q)/b, where q = 1-p.
Key takeaways
- The Kelly formula maximizes the expected value of the logarithm of wealth, which is equivalent to maximizing the long-run geometric growth rate of capital—the correct objective for investors with infinite investment horizons.
- Betting more than the Kelly fraction ('over-betting') reduces long-run growth rate despite higher expected value, and full Kelly can lead to dramatic drawdowns that are psychologically and practically intolerable.
- Fractional Kelly (typically 25-50% of the full Kelly bet) is widely used in practice, sacrificing some long-run growth for dramatically reduced volatility and drawdowns.
- For continuous returns, the Kelly fraction equals the Sharpe ratio squared divided by the variance of returns (f* = μ/σ²), equivalently μ/σ per unit of Sharpe ratio.
- Kelly sizing requires accurate estimates of edge (expected return) and risk—overestimation of edge with Kelly sizing can lead to rapid ruin, making conservative 'fractional Kelly' approaches appropriate when parameter uncertainty is high.
Explanation
The Kelly Criterion represents the mathematically optimal solution to the question of how much to bet on a positive-expected-value proposition when the goal is long-run wealth maximization rather than short-run expected value. It was originally developed by John L. Kelly Jr. at Bell Labs in 1956 as a solution to a gambling problem analogous to information transmission over noisy channels, and was subsequently recognized by Claude Shannon and Edward Thorp as directly applicable to investment management. Thorp, who later ran the hedge fund Princeton-Newport Partners, applied Kelly sizing successfully to blackjack card counting and subsequently to warrant and convertible arbitrage.
The derivation of the Kelly Criterion begins with the observation that long-run wealth maximization is equivalent to maximizing the expected value of the logarithm of wealth, E[ln(W)]. This is because wealth after many periods is the product of many growth factors, and by the law of large numbers the geometric average of these factors converges to e^{E[ln(W)]} almost surely. For a binary bet, if the bettor wagers fraction f on a bet paying b:1 with probability p of winning and probability q = 1-p of losing, wealth after one period is either (1+bf) with probability p or (1-f) with probability q. Maximizing E[ln(W)] = p·ln(1+bf) + q·ln(1-f) with respect to f gives the first-order condition: pb/(1+bf) - q/(1-f) = 0, which solves to f* = (bp-q)/b.
For continuous return distributions (more relevant to financial markets), the Kelly fraction takes the form f* = μ/σ², where μ is the expected excess return and σ² is the variance of returns. Equivalently, f* = SR/σ where SR is the Sharpe ratio. For a strategy with a 10% expected annual excess return and 20% annual volatility (Sharpe ratio = 0.50), full Kelly implies allocating f* = 0.10/0.04 = 2.5 times capital to the strategy—implying significant leverage. For a strategy with a 2% expected annual excess return and 5% annual volatility (Sharpe ratio = 0.40), Kelly implies f* = 0.02/0.0025 = 8.0x leverage, highlighting that Kelly sizing for typical financial strategies implies extremely high leverage and volatility.
The practical limitations of full Kelly sizing explain why sophisticated practitioners invariably use fractional Kelly approaches. The Kelly Criterion's derivation assumes perfect knowledge of the true probability distribution of returns. In practice, estimates of expected return and volatility are uncertain—the confidence interval around μ and σ estimates from finite historical data is large. If the estimated edge (expected return) is higher than the true edge (a common outcome given estimation error and overfitting), full Kelly sizing will over-bet relative to the true optimum, generating excess volatility and drawdowns. A commonly stated rule is that fractional Kelly at 25-50% of the theoretically optimal fraction dramatically reduces volatility while sacrificing only modest long-run growth.
For hedge funds, Kelly sizing provides a theoretical benchmark for position sizing that accounts for both return and risk. Quant funds use Kelly-inspired frameworks to set position weights in systematic strategies, with fractional Kelly adjustments reflecting estimation uncertainty, transaction costs, and liquidity constraints. The Kelly Criterion also provides insight into the relationship between information ratio, leverage, and growth: maximizing long-run capital growth requires understanding not just the Sharpe ratio of individual positions but the interaction between position correlations, aggregate portfolio Sharpe ratio, and optimal leverage. Platforms like Renaissance Technologies, widely regarded as the most successful quantitative hedge fund, have reportedly applied Kelly-inspired sizing frameworks alongside their proprietary signal generation to achieve extraordinary long-run compounding.
Formula
f* = (bp - q)/b (binary bet); f* = μ/σ² (continuous returns); Long-run growth rate: g = μf - σ²f²/2
Example
A systematic equity trader estimates that a specific momentum signal has an expected annual return of 5% with 15% annual standard deviation (Sharpe ratio = 0.33). The continuous Kelly fraction is f* = 0.05/(0.15)² = 0.05/0.0225 = 2.22x. Full Kelly implies allocating 2.22x capital to the strategy (222% exposure). A half-Kelly allocation of 1.11x provides a more moderate exposure. The expected annual growth rate at full Kelly is approximately μ - σ²/2 = 5% - 0.0225/2 = 3.875% per year, while the standard deviation of annual returns at full Kelly is 2.22 × 15% = 33.3%—a volatility level that most investors would find excessive. At half-Kelly (1.11x), expected growth is approximately 4.1% and standard deviation is 16.7%, a much more attractive risk-return profile despite the slight reduction in expected growth rate from the optimal.
Related terms
Arbitrage Beta Coefficient Convertible Arbitrage Diversification Drawdown Equity Factor Model Hedge Fund Information Ratio Law Of Large Numbers Leverage Liquidity