The Kelly criterion is a position sizing rule for repeated bets or trades.

The Kelly criterion is a position sizing rule for repeated bets or trades. In its standard binary form, it asks what fraction of capital should be allocated to a wager so that the expected logarithm of wealth is maximized over many repetitions. The key idea is growth under compounding. By maximizing expected log wealth rather than expected profit on a single bet, the rule targets the highest asymptotic growth rate of capital under the model assumptions in use.

In the basic setup, a trader risks a fraction of capital on an opportunity with known probability of success and known payoff odds. The Kelly fraction is the fraction that maximizes expected logarithmic growth for that opportunity. The mechanism is straightforward. If the fraction is too small, the trader leaves growth on the table when the edge is real. If the fraction is too large, losses compound too aggressively and the geometric growth rate falls. The Kelly optimum is the point where that tradeoff is balanced under the assumed probabilities and payoffs. Sonar’s fundamentals material presents Kelly as a money management concept tied to long run capital growth and bankroll allocation rather than to signal generation itself.[1]

That same mechanism also explains why Kelly is a ceiling, not a default. The formula is highly sensitive to estimation error in the edge, payoff distribution, and independence assumptions. In research and live trading, those inputs are not known with certainty. They are estimated from finite samples and model choices. If the edge is overstated, full Kelly sizing can become materially too large. Because log utility penalizes large losses sharply through compounding, oversizing can produce much higher path volatility and deeper drawdowns even when the strategy remains positive expectancy in sample. For this reason, sensible sizing often uses a fraction of Kelly rather than full Kelly. Sonar’s fundamentals page frames Kelly within risk management and position sizing discipline, which is consistent with using it as an upper bound rather than a target allocation.[1]

A practical way to interpret reduced Kelly is as a buffer against model risk. Model risk includes unstable parameter estimates, regime changes, and simplifications in the payoff model. When a backtest is only one realization from a broad research search process, full Kelly can amplify errors that come from selecting a model that fit historical noise. Capping exposure below the Kelly fraction reduces sensitivity to that uncertainty. It does not change the algebraic optimum under a perfectly known model. It changes the implementation to reflect that the model is not perfectly known.

This is where overfitting controls matter. Sonar’s backtest overfitting audit emphasizes evaluating a strategy in a way that accounts for multiple testing and selection effects, rather than treating a single attractive backtest at face value.[2] That matters directly for Kelly sizing because the Kelly fraction depends on estimated edge. If the edge estimate comes from an overfit backtest, the implied Kelly size will also be overstated. A ceiling below full Kelly is therefore aligned with the audit logic. The more uncertainty there is around the validity and stability of the backtest, the less defensible aggressive Kelly sizing becomes.[2]

The same caution applies to performance statistics used to justify leverage or concentration. Sonar’s glossary explains the deflated Sharpe ratio as a version of the Sharpe ratio adjusted to reflect non normal returns and multiple testing, with the aim of reducing false discoveries from backtest selection.[3] In sizing terms, that adjustment is relevant because an inflated in sample Sharpe can lead a researcher to infer more edge precision than is warranted. A reduced or capped Kelly fraction is one way to translate that uncertainty into capital allocation. If the evidence for the edge weakens after accounting for selection effects, prudent sizing should weaken as well.[3]

Full Kelly is the theoretical maximizer of long run log growth under known inputs, while overfitting audits and deflated Sharpe adjustments both argue for caution when those inputs come from backtests subject to selection bias.[2][3]

For quantitative traders, the practical lesson is to separate the theory from the implementation. The theory says Kelly identifies the growth optimal fraction under a specified probabilistic model. The implementation problem is harder because the model is estimated, searched, and exposed to regime uncertainty. In that setting, using Kelly as a ceiling and sizing materially below it is a risk management choice that acknowledges volatility, drawdown sensitivity, and estimation error. Sonar’s materials support that framing through their emphasis on disciplined money management, overfitting audits, and corrected performance evaluation.[1][2][3]

Covered in depth in the Strategy research fundamentals pillar hub.