Kelly Criterion in Practice: Why Everyone Uses Fractional Kelly

Full Kelly is mathematically optimal and behaviorally untradeable. It maximizes the expected logarithm of wealth under the assumption that you know your edge exactly, which you never do. Estimation error in the inputs compounds into catastrophic overbetting on the output. Every practitioner who has run Kelly on real strategies eventually arrives at the same conclusion: bet a fraction of what the formula tells you.

The formula and what it actually optimizes

For a binary bet with win probability p, loss probability q = 1-p, and payoff ratio b (units won per unit risked), Kelly's fraction is:

For continuous returns with mean μ and variance σ², the analog is:

Kelly maximizes E[log(W)], the expected log of terminal wealth. This is not the same as maximizing expected wealth — that objective is satisfied by betting 100% on any positive-edge wager, which guarantees eventual ruin. Log utility penalizes drawdowns symmetrically with gains, which is why Kelly produces a finite optimum. The geometric growth rate of a Kelly-sized strategy beats any other fixed-fraction strategy over infinite horizons, given correct inputs.

Why nobody actually uses full Kelly

The problem is the phrase "given correct inputs." Your estimate of μ comes from a backtest with finite samples, look-ahead bias risk, regime dependence, and survivorship in the underlying data. Your estimate of σ² is more stable but still understates tail behavior. If your true edge is half of what you measured, full Kelly puts you at 2x the optimal leverage — and the penalty for overbetting is asymmetric and brutal.

The growth-rate curve as a function of bet size is concave. Below f*, you give up some return. Above f*, you give up more return per unit of oversizing, and beyond 2·f* the geometric growth rate turns negative even though the arithmetic edge is still positive. You can be right about direction and still go broke through sizing alone.

What fractional Kelly buys you

Half-Kelly is the most common practitioner default. It captures roughly 75% of the geometric growth of full Kelly while cutting the volatility of returns by half. The drawdown profile is dramatically more tolerable: full Kelly produces a median maximum drawdown approaching 50% of peak wealth over long horizons. Half-Kelly cuts that to roughly 25%, and quarter-Kelly to roughly 12%.

The trade-off is not purely psychological. Lower leverage means you survive parameter drift, regime change, and the inevitable discovery that your backtested Sharpe was inflated. It also means you can keep position-sizing constant when your strategy underperforms for a year, rather than being forced to delever after a drawdown — which locks in losses and degrades the geometric return further.

Estimation error: the real reason for the fraction

If you treat μ and σ² as known constants, full Kelly is optimal. If you treat them as estimates with their own distributions, the optimal bet size shrinks. There is a clean Bayesian result here: under a normal prior on μ with estimation variance σ_μ², the adjusted Kelly fraction is approximately:

For a strategy with 1000 trades, σ_μ² can easily equal σ²/1000, which seems negligible. But the relevant uncertainty is not just sampling error in the backtest — it is model uncertainty, regime uncertainty, and the unknown gap between historical and forward performance. Most practitioners encode this by shrinking the Kelly fraction by 50-75%, which corresponds to an implicit belief that σ_μ² is on the order of σ² itself.

Practical implementation

Compute Kelly per-strategy using out-of-sample performance estimates, not in-sample backtest numbers. Apply a fraction — typically between 0.25 and 0.5 — to that figure. Cap the result at a hard maximum leverage you are willing to tolerate regardless of what the math says. Recompute as new data arrives, but smooth the position size to avoid whipsawing leverage on noise.

For multi-strategy portfolios, the Kelly extension involves the inverse covariance matrix of strategy returns multiplied by the vector of expected returns. The same fractional discipline applies, and the same estimation-error problem is worse because covariance matrices are notoriously unstable. Kestrel Signal exposes per-strategy Kelly statistics on the equity-curve output, but the framework leaves the choice of fraction to the operator — there is no universal correct value.

Most retail traders discover Kelly, run it on their best backtest, blow up, and conclude the formula is broken. The formula is fine. The inputs are not, and the fraction exists to absorb that gap.