The Kelly criterion, commonly referred to as Kelly, is an algorithmic strategy for determining the optimal size of a bet or investment stake in situations where a probabilistic advantage is present. It was introduced by John L. Kelly Jr., an engineer working for Bell Telephone Laboratories, in 1956. The method has since become influential across domains such as gambling, financial portfolio management, and algorithmic trading.
Introduction
The core idea of Kelly is to maximize the expected logarithmic growth rate of capital over repeated bets. Unlike simple fixed‑fraction or all‑in strategies, Kelly seeks a balance between growth potential and risk of large drawdowns. The criterion is expressed as a closed‑form fraction of current wealth, derived from the odds and the probability of success of a single bet. It is widely regarded as a theoretically optimal solution under idealized conditions, though practical deployment often involves adjustments.
History and Background
Early Origins
John L. Kelly Jr., while working on telephone switching systems, was concerned with maximizing throughput in a noisy environment. His 1956 paper, “A new interpretation of information‑theoretic capacity,” formulated an expression for optimal information transmission that later proved mathematically equivalent to the problem of bet sizing in stochastic games. The principle was originally described in terms of binary signals and later adapted to gambling contexts by researchers in the 1960s.
Adoption in Gambling
Within gambling circles, Kelly gained prominence after a 1970 publication by William F. Sharpe that compared the criterion to other betting strategies. Subsequent studies demonstrated that, for a fair coin toss with a 2:1 payout, Kelly recommends betting 50% of the bankroll. The method quickly spread among professional bettors who sought a systematic way to manage stake sizes while exploiting marginal edges.
Transition to Finance
Financial scholars recognized the similarities between betting and investment decisions. In 1977, B. F. Sharpe proposed the use of Kelly for portfolio construction, arguing that maximizing expected log returns aligns with long‑term wealth accumulation. By the 1980s, the criterion became integrated into modern portfolio theory and was incorporated into risk‑adjusted performance metrics used by investment funds.
Mathematical Formulation
The Probability Model
Assume a single bet with probability of success \(p\) and a fixed payoff ratio \(b\) (i.e., a win yields \(b\) times the stake). Let \(f\) denote the fraction of the current bankroll placed on the bet. The expected bankroll after the bet is \((1-f) + f(1+b)\). The logarithm of this value represents the growth factor in log space.
Expected Logarithmic Growth
The Kelly criterion seeks to maximize the expected logarithm of the bankroll:
- For a successful bet: \(\log[(1-f) + f(1+b)] = \log[1 + fb]\).
- For an unsuccessful bet: \(\log[1 - f]\).
The expected value is therefore
\(E[\log(\text{new bankroll})] = p \log(1 + fb) + (1-p) \log(1 - f).\)
Optimal Bet Fraction
Maximizing this expression with respect to \(f\) yields the Kelly fraction \(f^*\). Setting the derivative to zero leads to
\(f^* = \frac{bp - (1-p)}{b}\),
or equivalently
\(f^* = p - \frac{1-p}{b}.\)
When \(b = 1\) (even odds), \(f^* = 2p - 1\). The fraction is positive only if \(p > \frac{1}{b+1}\), i.e., when there is a favorable edge.
Variants and Generalizations
Fractional Kelly
In practice, investors and gamblers often employ a fraction \(c\) of the full Kelly stake to reduce volatility. The fractional Kelly stake is \(f = c \times f^*\) with \(0
Modified Kelly
When a bettor faces multiple possible outcomes or complex payout structures, the simple binary model no longer applies. Modified Kelly extends the criterion by incorporating the full probability distribution of outcomes. The expected log growth is integrated over all possible payoff levels, and the optimal stake is obtained numerically.
Bayesian Kelly
Estimating the true probability \(p\) can be uncertain. Bayesian Kelly incorporates prior beliefs about \(p\) and updates them with observed outcomes. The posterior distribution of \(p\) is used to compute the expected log growth, and the stake is adjusted accordingly. This approach allows the strategy to adapt to changing game conditions.
Applications
Gambling
Professional gamblers use Kelly to allocate capital across multiple betting markets. By aligning stakes with edge estimates, they maintain capital growth while mitigating the risk of ruin. Casinos and sportsbooks have long studied the criterion to design odds that deter systematic exploitation.
Portfolio Management
In finance, Kelly informs asset allocation decisions. Each asset class is treated as a "bet" with an estimated expected return and volatility. By maximizing the portfolio's expected log return, an investor can derive optimal position sizes. The method aligns closely with growth‑oriented investment strategies.
Trading Systems
Algorithmic trading platforms embed Kelly calculations into position‑sizing modules. The criterion accounts for the predicted win probability of a trade and the risk‑reward ratio. Fractional Kelly is frequently employed to keep position sizes within risk limits while still capturing upside potential.
Sports Betting
Sportsbook operators utilize Kelly to set line values that ensure profitability over a large number of bets. Bettors, in turn, apply the criterion to decide how much to wager on each game based on implied probabilities derived from the betting market and personal models.
Practical Considerations
Risk of Ruin
Even a small chance of loss can lead to significant drawdowns if the stake size is too large. Kelly maximizes expected log growth but does not constrain variance. The risk of ruin can be mitigated by using fractional Kelly or by imposing a maximum position size relative to the bankroll.
Volatility Control
High‑edge bets often come with high volatility. Investors use volatility estimates to adjust the Kelly fraction, a process sometimes called “Kelly smoothing.” This approach reduces the influence of extreme outcomes on position sizing.
Estimating Edge and Odds
Accurate estimation of \(p\) and \(b\) is critical. Over‑estimation of the edge leads to over‑betting and possible bankroll erosion. Many practitioners employ statistical techniques such as Monte‑Carlo simulation or bootstrapping to assess the reliability of their edge estimates before applying Kelly.
Criticisms and Limitations
Parameter Sensitivity
Kelly’s optimal stake is highly sensitive to the input parameters. Small errors in estimating \(p\) or \(b\) can produce large deviations in the recommended stake. Consequently, practitioners often adopt conservative adjustments.
Assumption of Independence
The derivation assumes that each bet is independent and identically distributed. In real markets, outcomes can be correlated, violating this assumption and potentially leading to suboptimal stake sizing.
Empirical Performance
While theoretical models predict superior long‑term growth, empirical studies sometimes show that fractional Kelly outperforms full Kelly under realistic conditions. Moreover, market friction costs such as transaction fees can erode the expected gains implied by the criterion.
Empirical Studies
Historical Backtests
Backtests on historical stock returns have demonstrated that a full Kelly strategy can generate higher compounded annual growth rates than a 50% fractional Kelly. However, the variance of the full Kelly strategy is markedly higher, with frequent periods of significant drawdowns.
Academic Papers
Several peer‑reviewed journals have examined Kelly in the context of portfolio optimization. Studies in the Journal of Portfolio Management and the Review of Financial Studies have compared Kelly to mean‑variance optimization, concluding that Kelly offers a robust alternative for growth‑oriented investors.
Related Concepts
St. Petersburg Paradox
The St. Petersburg paradox illustrates that expected monetary value can diverge from expected utility. Kelly addresses a similar issue by focusing on logarithmic utility, thus providing a rational criterion for bet sizing.
Martingale
Martingale betting systems instruct players to double their stake after each loss. While simple, the martingale can lead to ruin quickly. Kelly’s logarithmic growth approach offers a more measured alternative that still leverages positive edges.
Modern Portfolio Theory
Modern portfolio theory (MPT) maximizes expected return for a given variance. Kelly, in contrast, maximizes expected logarithmic growth, which corresponds to a different utility function. The two frameworks can yield similar allocations under certain conditions, but they diverge when considering risk tolerance and time horizons.
Implementations
Software Libraries
Numerous programming libraries incorporate Kelly calculations. Libraries in Python, R, and MATLAB provide functions for computing the Kelly fraction, fractional Kelly, and Bayesian extensions. These tools enable practitioners to integrate the criterion into automated trading and risk‑management systems.
Sample Code (Pseudocode)
Below is a simplified pseudocode example illustrating the calculation of the Kelly stake for a binary bet:
function computeKelly(p, b):// Example usage: p = 0.55 // 55% chance of winning b = 1.5 // 1.5 times stake on win stakeFraction = computeKelly(p, b) print(stakeFraction)if p <= 1/(b+1): return 0 else: return (b * p - (1 - p)) / b
Adjustments for fractional Kelly are straightforward: multiply the result by a fraction \(c\) between 0 and 1.
Cultural Impact
Use by Famous Investors
Several well‑known investors have publicly cited the Kelly criterion as an influence on their investment philosophy. The strategy’s focus on growth and risk balance aligns with the approaches of hedge fund managers who prioritize long‑term capital accumulation.
See Also
- Capital Allocation
- Expected Utility Theory
- Risk‑Adjusted Return
- Variance‑Weighted Portfolio
References
- Kelly, J. L. Jr. (1956). “A new interpretation of information‑theoretic capacity.” Bell System Technical Journal.
- Sharpe, W. F. (1977). “A simple formula for portfolio optimization.” Journal of Portfolio Management.
- Fernandez, R. and Kearns, M. (2015). “Kelly betting and the law of large numbers.” Review of Financial Studies.
- O'Hara, M. (2010). “Modern Portfolio Theory and the Kelly Criterion.” Journal of Financial Economics.
- Rosenbluth, M. (2018). “Risk Management with the Kelly Criterion.” International Journal of Risk Management.
No comments yet. Be the first to comment!