How to Size Positions with Kelly Under Uncertainty

Estimate the edge and its uncertainty

Kelly needs an expected edge, but the honest version also needs how confident you are in that edge. Estimate both from out-of-sample data, not the in-sample fit that produced the strategy. A small sample or a wide spread of results means high uncertainty, which should pull your bet size down. Treating the edge as a point estimate when it is really a distribution is the root cause of Kelly overbetting.

Use the lower end of a confidence interval for the edge, not the central estimate. Sizing on the optimistic case is how good strategies get overbet into ruin.

Use the Position Sizing under Edge Variance: Bayesian-Kelly bet sizing when your edge is itself uncertain. Compare deterministic Kelly, Bayesian-adjusted, and conservative lower-bound versions.

Compute the full Kelly fraction

Full Kelly is the bet size that maximizes the expected logarithm of wealth. For a simple bet it is the edge divided by the odds; for a continuous strategy it is roughly the expected excess return divided by the variance of returns. This number is the theoretical ceiling. It is correct only if your edge estimate is exact, which is why it is a starting point for discounting rather than a sizing decision in itself.

Full Kelly is famously aggressive: it routinely implies drawdowns of fifty percent or more even when the edge is real. Treat it as a maximum you will scale down from.

Use the Kelly Criterion Calculator: Size positions with full, half, quarter, or eighth Kelly and stress-test the choice with a drawdown Monte Carlo simulator. Client-side. Private by default.

Apply a fractional Kelly discount

Bet a fraction of full Kelly, commonly one half or one quarter. The reason is mathematical, not timid: because growth is concave near the Kelly optimum, halving the bet sacrifices only a small slice of expected growth while roughly halving the volatility of wealth. Under estimation uncertainty, fractional Kelly also corrects for the systematic overbetting that comes from sizing on an edge you measured with error.

The more uncertain your edge, the smaller the fraction. A well-measured edge from thousands of trades justifies a larger fraction than a fresh strategy with fifty trades.

Use a Bayesian-adjusted size when the edge is shaky

When the edge estimate is genuinely uncertain, a Bayesian approach formalizes the discount: it places a prior on the edge, updates it with your data, and sizes on the resulting posterior rather than a single point. With little data the posterior pulls heavily toward the prior, automatically shrinking the bet. As evidence accumulates, the size grows toward the data-driven Kelly. This is the principled version of fractional Kelly when you can quantify the uncertainty.

Compare the deterministic Kelly, the Bayesian-adjusted Kelly, and a conservative lower-bound version. The gap between them is a direct readout of how much your uncertainty is costing in size.

Stress the drawdown before committing the fraction

Even a correct Kelly fraction produces deep drawdowns, and the relevant question is whether you can survive them psychologically and financially. Simulate the path of wealth under your chosen fraction, given the strategy's return distribution including skew and kurtosis, and read the time and depth of the typical and tail drawdowns. If the recovery time or the depth would force you to abandon the strategy, cut the fraction until it would not.

Size to the drawdown you can sit through, not the one you can describe in a spreadsheet. The constraint that matters is the one that makes you capitulate at the bottom.

Use the Drawdown-Recovery Markov Simulator: Time to recover from an N% drawdown given monthly Sharpe + skew + kurtosis. Cornish-Fisher Monte Carlo, percentile distribution of recovery months.