Kelly Criterion Explained: What It Does, What It Assumes, Where It Fails

The Kelly criterion is one of those rare formulas in finance that is simultaneously elegant, well-tested, and dangerous when applied without care. Derived by John Kelly Jr. at Bell Labs in 1956 in the context of information theory and signal transmission, it gives the bet size that maximizes the long-run geometric growth rate of wealth. Edward Thorp used it to win at blackjack and then to run one of the most successful hedge funds of the 1970s. Mathematically, it is optimal.

And yet, applying full Kelly to equity positions is a near-guarantee of catastrophic drawdowns. The reason has nothing to do with the math being wrong and everything to do with the assumptions Kelly makes, assumptions that are roughly true for blackjack and roughly false for stocks.

The formula

For a binary bet (you either win or lose), the Kelly criterion gives the fraction of bankroll to bet as:

f* = (bp − q) / b

where b is the odds (payout per unit bet, e.g., 1 for even money), p is the probability of winning, and q = 1 − p is the probability of losing. If b = 1 (even-money bet), this simplifies to f* = 2p − 1, bet a fraction equal to your edge.

For a continuous bet (a stock position), the generalized Kelly formula is:

f* = μ / σ²

where μ is the expected excess return (above the risk-free rate) and σ² is the variance of returns. The intuition is the same: bet bigger when the expected return is high relative to the variance.

For a stock with μ = 10% and σ = 25% (typical for a single-name equity), full Kelly gives f* = 0.10 / 0.0625 = 1.6, or 160% of bankroll. That is, the formula tells you to lever up 1.6x. Most people's intuition correctly screams that this is insane. The intuition is right; the formula is right; the application is wrong.

Why full Kelly is suicidal in equities

The Kelly formula assumes you know μ and σ. In a casino, where the rules are fixed and the probabilities are computable, this is true. In equities, μ and σ are estimated from past data, and the estimation error is enormous. A typical 95% confidence interval on the expected return of a single stock based on 10 years of monthly data is something like [-15%, +25%]. The point estimate is 10%, but it could plausibly be anywhere in that range.

Kelly is wildly sensitive to this error. If your true μ is half of what you estimated, the optimal Kelly is one-quarter of what you bet. Bet at the "estimated optimum" when your estimate is too high by a factor of 2 and you are over-leveraged by 4x. Real markets routinely surprise estimators by that much, especially for individual names.

The second problem is that returns are not normally distributed (we wrote about this in the Monte Carlo article). The Kelly derivation assumes log returns are normal. Real log returns are fat-tailed. Fat tails mean extreme negative returns are far more frequent than the formula assumes, which means the variance estimate σ² systematically understates the true risk. Kelly under-prices risk in fat-tailed markets, which leads to over-betting.

The third problem is that even a small over-bet relative to optimal Kelly leads to catastrophic drawdowns. The growth curve as a function of bet size is asymmetric, it rises gently as you approach optimal Kelly and falls off a cliff as you exceed it. Betting at 1.5x optimal Kelly does not give you 1.5x the returns; it gives you, on average, slightly worse returns with a much higher chance of going to zero.

Fractional Kelly: half-Kelly, quarter-Kelly

The standard fix is to bet a fraction of full Kelly, half-Kelly or quarter-Kelly. The math of fractional Kelly is well-studied. Half-Kelly gives roughly 75% of full Kelly's growth rate while cutting the worst-case drawdown roughly in half. Quarter-Kelly gives roughly half the growth rate with another large cut to drawdown.

For most retail investors, quarter-Kelly is the right starting point. The growth-rate sacrifice is real but small; the drawdown protection is huge. Professional traders sometimes operate at half-Kelly because they have access to better estimates of μ and σ (and because their psychological tolerance for drawdowns is higher than retail). Full Kelly is essentially never used outside of toy problems with known probabilities.

Drawdown tolerance and why caps appear in practice

There is a second consideration beyond growth rate: drawdown tolerance. Even at quarter-Kelly, a portfolio can experience drawdowns of 30-40% during bad regimes. If a holder's psychological tolerance is 20%, they tend to sell at the bottom and lock in losses. The largest bet size that can be held through a drawdown without panic-selling is usually well below what the math suggests is optimal.

For this reason, most disciplined practitioners apply a hard position cap, frequently 5-10% of portfolio per single name, regardless of what the formula suggests. The cap is not a refinement of Kelly. It is an admission that Kelly's assumptions do not hold cleanly and that estimation error must be absorbed somewhere.

• Estimation error: μ and σ estimates are noisier than the formula assumes; a cap absorbs that error.
• Tail risk: single-stock drawdowns can hit -50% in a quarter (see 2022 META, 2020 BA, much of biotech). A 5% position cap means such an event costs 2.5% of portfolio. A 15% cap costs 7.5%.
• Diversification: a 5% cap forces a minimum of 20 positions for a fully invested book, which captures most of the diversification benefit available in single-name equity portfolios.

Where the inputs come from in practice

Kelly needs two inputs: expected excess return μ and variance σ². Neither is observable. Both must be estimated.

σ² is the easier of the two. It is typically estimated from realized return variance over a rolling 252-day window, with adjustments for volatility clustering (GARCH-style updating). This is a well-understood estimation problem and the resulting σ is reasonably stable.

μ is the hard one. Using a trailing average return is essentially useless because the standard error of the mean is enormous on noisy data. Quant systems usually derive μ from a composite signal (factor model, scoring system, ML output) and calibrate the mapping from "signal level" to "average historical forward return at that signal level" using isotonic regression on a held-out validation set. We cover the mechanics in Isotonic Calibration. The key property is that the resulting μ is honestly calibrated rather than wishful. If the model says μ = 4%, the average historical outcome at that signal level was actually around 4%, not 8%.

A worked example, mechanically

Suppose a single name has a calibrated expected 6-month excess return of μ = 3% and an annualized σ = 22% (6-month σ ≈ 15.6%). Full Kelly would be f* = 0.03 / 0.156² ≈ 1.23, or 123% of portfolio. That is the mathematically optimal bet under known μ and σ. It is also wildly inappropriate as guidance once you remember that μ is estimated, not known.

Quarter-Kelly gives 30.8% of portfolio in a single name. Still uncomfortable for most readers, and for good reason: concentration that large only makes sense if you genuinely trust μ to within a fraction of itself, which is rare in equities.

A 5% per-position cap, common in practitioner workflows, reduces this to 5%. The Kelly machinery then mostly serves as a binary entry test: at calibrated μ and σ, does Kelly suggest more than a trivial fraction? If not, the position may not be worth taking. If yes, the cap binds and the position is sized at the cap.

Where Kelly fits and where it fails

• Where the math fits cleanly: well-defined repeated bets with known probabilities (blackjack, sports betting markets with sharp lines).
• Where it fits loosely: long-horizon discretionary or systematic equity strategies with honestly calibrated μ and σ and uncorrelated positions.
• Where it routinely fails: highly correlated positions (Kelly assumes independence; in tail scenarios correlations rise toward one and the portfolio becomes a single bet).
• Where it fails completely: short-horizon options trades where the binary-outcome assumption breaks (asymmetric payoffs, vega and theta exposure not captured by μ and σ alone).
• Where it is not defensible at all: when μ cannot be justified from anything other than narrative. "I think this stock will rip" is not a defensible μ; full Kelly built on top of it is hallucinatory.

Conclusion

The Kelly criterion answers a precise mathematical question: given known μ and σ for a repeated bet, what is the bet fraction that maximizes long-run geometric growth? The answer is mathematically elegant and routinely wrong as a literal recommendation for equity positions, because μ and σ are not known, returns are fat-tailed, and even small over-estimation of μ leads to ruinous over-betting. Quarter-Kelly with a hard position cap is the common practitioner adjustment.

What to check in your own setup: open the last analysis you ran. If the platform reports a position size, ask how μ and σ were estimated, what calibration was applied, and what the cap (if any) is. If none of those answers are visible, the number being shown is closer to a guess than a sizing recommendation.

Research, not advice. Always verify before acting.