The Kelly Criterion calculates the optimal position size for long-run growth. Here is the formula, its limits, and how to apply it practically.
Kelly Criterion trading applies a mathematical formula to determine the optimal position size for maximum long-run account growth. Unlike fixed-percentage sizing, Kelly scales position size based on the strategy's win rate and win/loss ratio. Bet too small and you underutilize edge. Bet too large and variance destroys the account before the edge can compound. Kelly identifies the exact fraction that maximizes long-run growth — and explains why most practitioners then deliberately bet less than that fraction.
John L. Kelly Jr. developed the criterion at Bell Labs in 1956. The original paper addressed information theory — specifically, how much of a transmission channel's capacity to use given a noisy signal. The connection to betting and trading came through Edward Thorp, who applied Kelly's mathematics to blackjack in the 1960s and then to financial markets in subsequent decades.
Kelly solves one precise problem: given a bet with known positive expected value, what fraction of the bankroll maximizes the long-run growth rate? The answer is neither "bet everything" nor "bet the minimum." Betting everything maximizes expected value per bet but produces ruin with certainty over many bets — one loss wipes the account. Betting too little is safe but leaves most of the edge uncaptured. Kelly finds the exact fraction that produces the highest geometric growth rate over many bets.
The criterion is directly related to trading expectancy. Kelly is the position sizing formula that expectancy feeds into. Expectancy tells you whether an edge exists. Kelly tells you how large to size positions to exploit that edge optimally over time. For the expectancy calculation, see Trading Expectancy.
The trading version of the Kelly formula:
f* = (p x W — q x L) / (W x L)
Where f* is the optimal fraction of bankroll to risk per trade, p is the win rate, q is the loss rate (1 minus p), W is the average win as a decimal (e.g., 0.10 for a 10% gain), and L is the average loss as a decimal (e.g., 0.05 for a 5% loss).
A simpler equivalent using the win/loss ratio (R):
f* = (p x R — q) / R, where R = average win / average loss
Example: Win rate 55%, average win $200, average loss $100. R = 2. f* = (0.55 x 2 — 0.45) / 2 = (1.10 — 0.45) / 2 = 0.325. Kelly says to risk 32.5% of the bankroll per trade.
This fraction seems high. That is the correct reaction — and it is why most practitioners use a fraction of Kelly rather than the full value. The formula produces the mathematically optimal fraction given perfectly known win rate and win/loss ratio. In practice, neither is known precisely.
Full Kelly maximizes the long-run growth rate given perfect knowledge of the inputs. It produces severe drawdowns in practice because estimated win rates have uncertainty, outcome distributions have fat tails, and the formula assumes unlimited ability to absorb variance. Full Kelly players experience drawdowns that most traders find psychologically impossible to hold through — even when the underlying strategy has genuine edge.
Half Kelly (0.5x full Kelly) produces approximately 75% of the growth rate of full Kelly with dramatically smaller drawdowns. The reduction in growth rate is modest. The reduction in variance is substantial. For most systematic traders, half Kelly is the most commonly used fraction. It preserves most of the compounding benefit while keeping drawdowns within sustainable range.
Quarter Kelly (0.25x full Kelly) produces approximately 44% of full Kelly growth rate but with very modest variance. Appropriate for new strategies where the estimated win rate has wide confidence intervals, for strategies with fat-tailed outcome distributions, or for any situation where the operator is near their psychological or structural risk tolerance limit.
The key insight: the difference in long-run growth between full Kelly and half Kelly is smaller than most expect (75% vs 100% of optimal growth). The difference in experienced drawdowns is much larger. Fractional Kelly almost always makes sense for trading applications, where input uncertainty alone justifies the reduction.
The Kelly formula requires two inputs: win rate and average win/loss ratio. These must come from a statistically sufficient sample — the backtest or live trading history. The critical consideration: how accurately do these estimates reflect the true underlying values?
A backtest with 50 trades produces win rate estimates with wide confidence intervals. An observed win rate of 55% might reflect a true win rate anywhere from 42% to 68% at the 95% confidence level. Applying full Kelly to an imprecise win rate estimate produces oversized positions that can destroy the account before the true win rate is established with statistical confidence.
The robustness rule: use the lower bound of the win rate confidence interval in the Kelly calculation, not the point estimate. If the observed win rate is 55% but the 90% confidence interval lower bound is 48%, use 48% in the Kelly formula. This produces a conservative Kelly fraction that underbets if the true win rate is higher — acceptable — but does not overbet if the observed win rate was partly luck.
For regime-specific strategies, calculate Kelly separately for each regime state. A strategy with 62% win rate in RANGING conditions and 44% in TRENDING conditions has fundamentally different optimal sizing for each regime. A single combined Kelly fraction averages across both and misrepresents the appropriate size in each condition. For the ATR-based sizing approach that complements Kelly, see ATR Position Sizing.
In the live signal pipeline, Kelly is used as a theoretical maximum rather than as the operating position size. Regime-segmented expectancy data — win rate and average win/loss ratio per regime state and per confidence band — feeds the Kelly calculation. The operating position size is then set at a fixed fraction of full Kelly (currently 25% to 33%) to account for estimation uncertainty, fat-tailed outcomes in crypto, and the practical observation that full Kelly produces drawdowns exceeding sustainable limits.
One practical outcome of regime-segmented Kelly: when RANGING regime shows a Kelly fraction of 0.18 and TRENDING_BULLISH shows 0.08, operating at 25% of full Kelly produces sizing of approximately 4.5% and 2% of account per trade respectively. This naturally produces larger positions in the regime with stronger risk-adjusted edge — the correct outcome — while keeping both within sustainable risk parameters. No manual adjustment is needed; the Kelly fraction does the differentiation automatically.
The Kelly fraction is recalculated at each shadow data review as sample size grows. As observed win rate confidence intervals tighten with more data, the Kelly fraction becomes more precisely estimated. A regime that starts at 25% of Kelly may move toward 30% as the regime's sample accumulates statistical confidence in the win rate estimate. The fraction never moves toward 100% of Kelly — the fat-tail risk of crypto markets always justifies maintaining a buffer below full Kelly.
Parameter estimation error. Kelly is only as accurate as its inputs. Win rates and win/loss ratios are estimated from samples. Estimation error on the upside — the observed win rate is luckier than the true win rate — produces overbetting that accelerates drawdowns exactly when the strategy is not performing at its estimated level. This is the primary risk of applying Kelly in practice. Always use conservative estimates, particularly with small sample sizes.
Fat-tailed outcomes. Kelly was derived for distributions with defined variance. Crypto's occasional extreme moves produce outcomes far outside what standard win/loss ratio estimates assume. A 3:1 win/loss ratio estimated from 200 trades may be overwhelmed in impact by a single extreme loss that never appeared in the estimation window. Full Kelly does not protect against tail events not represented in the sample. Fractional Kelly and portfolio exposure limits provide partial protection.
Serial correlation. Kelly assumes independent bets. In trading, consecutive trades on the same asset in the same market condition are correlated — they tend to win and lose together rather than independently. Kelly sizing calibrated for independent bets oversizes during correlated losing streaks. Anti-Martingale scaling (reducing size after consecutive losses) partially corrects for this, but the Kelly formula itself does not account for it.
Multiple simultaneous positions. Kelly was designed for sequential, single bets. Applying full Kelly to each of three simultaneous positions produces three times full Kelly in total exposure. The multi-position extension of Kelly requires knowing the correlation between all positions — practically difficult. In practice, use fractional Kelly per position and impose a hard portfolio-level exposure cap to prevent correlated simultaneous losses from compounding beyond sustainable limits.
Kelly Criterion is a position sizing formula — it tells you what fraction of the account to risk per trade. It is the theoretically optimal complement to expectancy: expectancy tells you whether an edge exists, Kelly tells you how large to exploit it. In practice, Kelly-derived sizes are usually scaled down (half or quarter Kelly) due to estimation uncertainty, and then further calibrated by ATR to reflect current market volatility. Kelly sets the optimal fraction; ATR determines how many units that fraction corresponds to given current conditions.
Yes, with additional caution around crypto's fat-tailed outcome distribution. The Kelly formula applies to any strategy with measurable expectancy, including crypto. The adjustments: use conservative (lower-bound) win rate estimates due to limited sample sizes in crypto's short history; apply quarter Kelly rather than half Kelly to account for crypto's occasional extreme moves not captured in the estimation sample; and always impose a portfolio exposure cap to prevent multiple correlated crypto positions from creating combined Kelly exposure that exceeds sustainable limits.
The Kelly Criterion is a formula that calculates the optimal fraction of a trading account to risk per trade to maximize long-run account growth. Developed by John Kelly at Bell Labs in 1956 and applied to trading by Edward Thorp, it solves the problem of how large to size positions given a strategy with known positive expectancy. The formula uses win rate and average win/loss ratio as inputs. Most practitioners use a fraction of the full Kelly output (half Kelly or quarter Kelly) to reduce variance.
Fixed percentage sizing risks the same percentage of account on every trade regardless of the strategy's edge quality. Kelly sizes each trade based on the strategy's specific win rate and win/loss ratio — a higher-edge setup gets a larger Kelly fraction than a lower-edge one. In practice, if a strategy has consistent expectancy across all its trades, fixed percentage and fractional Kelly produce similar results. The difference emerges when different regime states have different expectancy: Kelly naturally sizes larger in high-edge conditions and smaller in lower-edge ones without requiring manual adjustment.
Fractional Kelly means using a fraction of the full Kelly formula output as the actual position size — typically 0.5x (half Kelly) or 0.25x (quarter Kelly). Full Kelly maximizes theoretical long-run growth but produces severe drawdowns in practice because the inputs (win rate, win/loss ratio) are estimated with uncertainty, and overestimates lead to overbetting. Half Kelly produces approximately 75% of full Kelly's growth rate with dramatically lower drawdowns. The modest growth rate sacrifice is almost always worth the variance reduction for live trading.
f* = (p x R — q) / R, where f* is the fraction of bankroll to risk, p is the win rate, q is 1 minus win rate, and R is the average win divided by the average loss. Example: 55% win rate, average win $200, average loss $100. R = 2. f* = (0.55 x 2 — 0.45) / 2 = 0.325 — Kelly says to risk 32.5% per trade. In practice, most traders apply half or quarter Kelly to reduce variance: 0.5 x 0.325 = 16.25% per trade at half Kelly.
Four main limitations: (1) Parameter estimation error — win rates and win/loss ratios are estimated from samples and may not reflect the true underlying values, causing overbetting when estimates are optimistic. (2) Fat-tailed outcomes — the formula assumes defined variance, but extreme moves in crypto can exceed the historical sample's worst case. (3) Serial correlation — Kelly assumes independent bets, but consecutive trades on the same asset tend to be correlated. (4) Multiple positions — Kelly was designed for sequential single bets; applying it to multiple simultaneous positions requires correlation adjustments most implementations ignore.