Fractional Kelly in practice: taming the optimal formula

The Kelly criterion promises the mathematically optimal formula for growing capital — and that's true, on paper. In practice, applying it to the letter has ruined more traders than it has enriched. Understanding why, and how to fraction it intelligently, is one of the most underrated risk management skills in quantitative trading.

What the Kelly criterion is

Developed in 1956 by John Kelly, an engineer at Bell Labs, this criterion answers a precise question: what fraction of your capital should you bet on each wager to maximize long-term geometric growth rate? Unlike a fixed or intuitive bet size, Kelly dynamically adjusts position size based on your statistical edge and win probability.

The formula and its intuition

In its simplest form, f* = (bp − q) / b, where p is the probability of winning, q = 1 − p the probability of losing, and b the win/loss ratio. Intuitively: the larger your edge and the more asymmetrically favorable your gains, the more you should bet — but the formula severely punishes any overestimation of p or b, which is precisely the problem in practice.

Why full Kelly is dangerous in practice

Full Kelly assumes you know p and b with perfect certainty. On a finite sample of historical trades, these parameters are just estimates — often optimistic, sometimes inflated by the data-snooping discussed elsewhere on this blog. Even a modest overestimation of the edge causes full Kelly to recommend an excessive bet, with drawdowns that can exceed 50% or 70% of capital, even when the underlying strategy remains statistically valid.

Extreme sensitivity to estimation errors

This is full Kelly's central flaw: the formula is extremely sensitive to errors in its own inputs. A win probability overestimated by just a few percentage points can turn an 'optimal' bet into a ruinous one. And since parameters are always estimated from past data, some degree of statistical uncertainty is unavoidable — full Kelly, however, builds in zero safety margin for that uncertainty.

Fractional Kelly: divide and conquer

The solution commonly adopted by professionals is simple in principle: bet a fraction of full Kelly — most often half ('half-Kelly'), sometimes a quarter. This choice drastically reduces capital volatility for a surprisingly modest cost in growth, thanks to the concave shape of the geometric growth function near its optimum.

How much to fraction, concretely

A half-Kelly typically captures around 75% of the theoretical maximum growth, while cutting capital variance in half compared to full Kelly — a clearly favorable trade-off. A quarter-Kelly goes further toward caution, relevant when uncertainty about the parameters is high (small sample, recently deployed strategy, shifting market regime). The choice of fraction should depend directly on your actual statistical confidence in p and b, not an arbitrary preference.

Building in real statistical uncertainty

An even more rigorous approach weights the Kelly fraction not by a fixed rule, but by objective indicators of how reliable your estimated edge actually is: the Deflated Sharpe Ratio (was it inflated by the number of trials tested?), edge stability across Combinatorial Purged Cross-Validation (CPCV), and the effective sample size of available trades. The weaker these signals, the closer the applied Kelly fraction should move toward zero — cancelling it entirely if the statistical signal simply isn't strong enough.

Capping the bet, no matter what

Beyond fractioning, an absolute cap on position size (say, never more than 5% of capital on a single bet, regardless of what the formula suggests) protects against cases where the estimated edge is radically wrong — a model error, a market regime shift, or simply extreme statistical bad luck. Fractional Kelly reduces risk; the absolute cap eliminates the risk of catastrophic ruin.

Putting it all into practice

Concretely: calculate your theoretical full Kelly from your trading statistics (win rate, average win/loss ratio). Apply a fraction — a half or a quarter depending on your confidence in these estimates. Weight that fraction by the actual statistical robustness of your edge (DSR, CPCV stability). And set an absolute cap that doesn't depend on any calculation, as a last safety net.

A worked example of fractional Kelly

Take a strategy with a 55% win rate and an average win/loss ratio of 1.2 (b = 1.2). The theoretical full Kelly gives f* = (1.2 × 0.55 − 0.45) / 1.2 ≈ 0.175, or 17.5% of capital to risk on each position — an already dizzying size for most traders. A half-Kelly brings that bet down to roughly 8.75%, a quarter-Kelly to 4.4%. If the trade sample used to estimate that 55% win rate only counts 80 trades, statistical uncertainty around that rate remains significant — one more reason to favor quarter-Kelly, or an even more conservative fraction, until the track record grows longer.

Kelly in a portfolio: the problem of correlated bets

The Kelly calculation shown so far assumes a single, isolated bet. In reality, most traders hold several positions at once — and if those positions are correlated (several tech stocks, say, that tend to rise and fall together), applying an individual Kelly to each one seriously underestimates the portfolio's overall risk. A multi-asset version of Kelly exists (using the covariance matrix of returns), but in practice, the most robust fix remains reducing the fraction applied to each position further as soon as the portfolio holds bets that are strongly correlated with one another.

Kelly isn't built for your emotional risk tolerance

One last point, often forgotten: Kelly, even fractional, optimizes the geometric growth of capital — not your psychological comfort. A well-calibrated half-Kelly can still produce losing streaks that feel endless in real time, even when they're statistically normal. Many traders who abandon a profitable strategy mid-drawdown aren't victims of a bad Kelly calculation — they're caught in the gap between the mathematically optimal position size and what they can actually stomach emotionally without deviating from their plan. A more conservative fraction than the theoretical half-Kelly — a third, say — is often the right choice not for statistical reasons, but to stay faithful to your own strategy over time.

Key takeaway

Full Kelly is a useful theoretical reference, never a recommendation to follow literally. Fractioning it (typically by half), weighting it by real statistical confidence in your parameters, and capping it in absolute terms turns a fragile formula into a genuinely usable position-sizing tool. That's exactly the logic TrueVerdikt applies automatically to every strategy it analyzes: half-Kelly, capped, weighted by DSR and CPCV stability, and forced to zero on a NO-GO verdict — test your own to see the position size real statistical rigor actually recommends.