Kelly Criterion for a Small Bankroll: Fractional Kelly in Practice

The Kelly criterion is not a stake-size rule you can pull off a shelf. It is a formula that requires you to supply a number—your estimate of the win probability—and every dollar you risk depends on how close that number is to reality. Fractional Kelly exists precisely because the number is almost never exactly right, and the cost of being wrong scales fast when the bankroll is small.

The 1956 Bell Labs Paper That Started It

John L. Kelly Jr. published “A New Interpretation of Information Rate” in 1956, in the Bell System Technical Journal, volume 35, pages 917 through 926. The original paper is notoriously hard to obtain; later arXiv preprints note that the text is often reproduced in secondary sources precisely because the primary journal is not widely accessible.

Kelly’s problem was not about sports. He worked with a sequence of simple Bernoulli bets—wagers with two outcomes—where the bettor held an edge, meaning a positive expected return on each individual bet. The objective was to choose the fraction of bankroll that maximized long-run logarithmic growth. That objective mattered because maximizing log growth implicitly balances compound returns against the risk of ruin in a way that maximizing simple expected value does not. A 2020 arXiv preprint on Kelly and Lévy processes explicitly frames the original criterion as a long-run growth strategy derived from exactly this Bernoulli-bet setting.

The theory is clean. The application is not.

The Stake Rule the Reader Can Actually Compute

You cannot compute a Kelly stake with a hunch. The framework requires two inputs: an estimate of the win probability and the payoff or odds of the wager. A Simon Fraser University paper on modified Kelly criteria makes this explicit—both numbers must be specified before any fraction leaves the calculator.

The full Kelly fraction is the output of that calculation. For a binary bet with known probability and known odds, the formula produces a single number: the proportion of bankroll to risk. But full Kelly is only the starting convention. The standard practical modification is fractional Kelly, which means betting a fixed proportion of the full Kelly fraction instead of the entire amount.

Half-Kelly means betting one-half of the full fraction. Quarter-Kelly means betting one-quarter. The naming is literal and the arithmetic is simple: multiply the full-Kelly stake by 0.5 or 0.25. What is not simple is knowing whether you should.

Why Full Kelly Is Fragile

Full Kelly is mathematically optimal under one brutal condition: the edge estimate must be correct. A 2021 experimental review describes fractional betting as the common practical modification and gives half-Kelly with a weight parameter of 0.5 as the worked example. The reason is not theoretical elegance. It is that the edge estimate is wrong, always, by some margin.

The original Kelly criterion was derived for a setting where the probabilities are known. In gambling, they are not. The bettor estimates, models, guesses. When that estimate overshoots the true edge, full Kelly over-allocates, and the compounding math that makes log-growth optimal also amplifies the error. Later academic sources treat the fractional rule not as a footnote but as the bridge between the mathematics and the mess.

The trap is thinking of fractional Kelly as merely conservative. It is not. It is a hedge against misestimation. A bettor who is 20% wrong about the edge will do far less damage at quarter-Kelly than at full Kelly, and the difference compounds with every bet.

What the Simulations Show About Drawdowns

Simulations of bankroll curves show what halving the Kelly stake does. Halving reduces the probability of large drawdowns across several thresholds: the chance of finishing below 80% of starting bankroll drops, and the pattern holds at 60%, 40%, and 20% as well. Exact probabilities depend on the edge and the number of bets, so this desk quotes no single figure, but the direction of the effect is consistent and stark.

For a small bankroll, this is not an abstraction. A 40% drawdown on a $500 bankroll leaves $300. The same drawdown on $50,000 leaves $30,000. The mathematics is indifferent to the absolute number, but the bettor is not. Fractional sizing lets the bankroll survive long enough for the edge to express itself, and survival is the variable that the logarithmic growth formula assumes away.

The Kelly rule is only as good as the probability and payoff estimates behind it. John Kelly solved the problem of how much to bet when you know the odds exactly. The bettor with a small bankroll faces a different problem entirely: how much to bet when the odds are a guess. The answer, almost always, is less.