The Kelly Criterion
The Kelly criterion is the position-sizing fraction that maximises the long-run geometric growth rate of capital, given by f* = W − (1 − W) ÷ b for a simple win-loss bet, beyond which betting more lowers growth while still raising risk.
Quick answer: The Kelly criterion is the position-sizing fraction that maximises the long-run geometric growth rate of capital, given by f* = W − (1 − W) ÷ b for a simple win-loss bet, beyond which betting more lowers growth while still raising risk.
In simple words
The Kelly criterion answers one precise question: what fraction of your capital should you bet to grow it fastest over the long run. Bet less than Kelly and you grow more slowly; bet more and you actually grow more slowly again while taking wilder swings. In practice traders bet only a fraction of the Kelly amount, such as half, because the formula assumes you know your edge exactly, which you never do.
Purpose
Kelly exists to define the mathematically optimal growth fraction so that sizing is anchored to an objective ceiling; knowing that ceiling lets a trader bet a deliberate fraction of it and understand exactly what growth and volatility they are trading away.
Professional explanation
What Kelly optimises and why it is a ceiling
The Kelly criterion maximises the expected logarithm of wealth, which is equivalent to maximising the long-run geometric growth rate of the account. Maximising log wealth is the correct objective for a repeated, reinvested bet because growth compounds multiplicatively, so it is the geometric mean, not the arithmetic mean, that determines terminal wealth. The Kelly fraction is the unique bet size that sits at the top of the growth curve; because that curve is concave, betting either side of it grows the account more slowly, and over-betting is punished on both counts by lower growth and higher volatility.
The formula for a simple win-loss bet
For a bet that wins with probability W and loses with probability 1 − W, paying b to 1 on wins and losing the full stake on losses, the optimal fraction is f* = W − (1 − W) ÷ b. Here b is the payoff ratio, the average win divided by the average loss. If the edge is zero or negative the formula returns zero or a negative fraction, correctly telling you not to bet. This discrete form is the one most traders quote, though real trading with variable win and loss sizes requires the continuous version that maximises expected log return over the actual return distribution.
Why full Kelly is rarely used
Full Kelly is optimal only if the win probability and payoff are known exactly and stationary, which is never true of a market edge estimated from a finite, noisy backtest. Because the growth curve is flat near its peak but falls steeply beyond it, a small overestimate of the edge pushes you past optimal into the region of lower growth and much higher drawdown. Full Kelly also implies brutal equity swings: drawdowns on the order of the win probability are routine. For these reasons practitioners treat the Kelly fraction as a ceiling to stay well under, not a target to hit.
Half-Kelly and fractional Kelly
Betting a fixed fraction of Kelly, most commonly one half, is the standard compromise. Half-Kelly captures about three quarters of the full growth rate while roughly halving the volatility and cutting the drawdown dramatically, because growth is concave near the peak but variance falls linearly with the fraction. Quarter-Kelly is common where the edge estimate is especially uncertain. The general principle is that fractional Kelly trades a little growth for a large gain in robustness against the estimation error and non-stationarity that plague every real strategy.
Kelly in a backtesting workflow
In validation, Kelly is best used to interpret a fixed-fractional study rather than to set the live bet directly. You estimate W and b from the out-of-sample trade record, compute f*, and note where your chosen fraction sits relative to it: a strategy sized at, say, one third of estimated Kelly has a comfortable buffer, while one sized above Kelly is over-betting even if the backtest looks strong. Because W and b are themselves uncertain, responsible practice recomputes Kelly with conservative, downward-adjusted estimates, checks its sensitivity to small changes in the inputs, and never lets a single in-sample estimate drive real leverage.
Formula
f* = W − (1 − W) ÷ b
f* = optimal fraction of capital to risk; W = probability of a winning trade (e.g. 0.55 for 55 percent); 1 − W = probability of a loss; b = payoff ratio = average win ÷ average loss. If f* is zero or negative there is no edge and you should not bet; most traders then use a fraction of f*, such as 0.5 × f* for half-Kelly.
Full Kelly vs Half-Kelly
| Aspect | Full Kelly | Half-Kelly |
|---|---|---|
| Long-run growth | Maximum possible | About three quarters of maximum |
| Volatility of equity | Very high | Roughly half |
| Sensitivity to bad estimates | Severe; easy to over-bet | Much more forgiving |
| Typical drawdowns | Deep and frequent | Substantially shallower |
| Practical use | Rare, theoretical benchmark | Common working choice |
Practical example
Illustrative example (Indian market)
From an out-of-sample record of a Nifty options strategy you estimate a win rate W = 0.55 and a payoff ratio b = 1.5, meaning average wins are 1.5 times average losses. Kelly gives f* = 0.55 − (0.45 ÷ 1.5) = 0.55 − 0.30 = 0.25, so full Kelly says risk 25 percent of capital per trade. On Rs 5,00,000 that is Rs 1,25,000 at risk on a single trade, which would produce violent equity swings, so you bet half-Kelly at 12.5 percent, or Rs 62,500, and more likely quarter-Kelly given the estimates are noisy. The exercise shows the ceiling; the discipline is choosing to sit well below it.
Indian F&O edges are often estimated from short samples across a handful of expiries, so a win rate measured on 60 trades has a wide confidence interval. Plugging that noisy W into Kelly can overstate the optimal fraction badly, which is why treating the computed f* as an upper bound and betting a quarter of it is the safer reading for a retail options book.
Limitations
- It assumes the win probability and payoff are known exactly and are stationary, which markets never satisfy
- The growth curve falls steeply past the peak, so overestimating edge pushes you into lower growth and deeper drawdowns
- Full Kelly produces equity swings most traders cannot psychologically or operationally tolerate
- The simple formula assumes a single fixed payoff, whereas real trades have a full distribution of win and loss sizes
- Estimation error from short backtests can make the computed fraction meaningless without conservative adjustment
Why it matters in practice
- It provides an objective ceiling that anchors every fractional-Kelly or fixed-fractional sizing decision
- It formalises why over-betting reduces long-run growth, correcting the intuition that bigger bets always grow faster
Common mistakes
- Betting full Kelly from an in-sample estimate, then over-betting because the real edge was smaller
- Using a single fixed payoff b when trades have a wide, skewed distribution of outcomes
- Treating a negative or near-zero f* as a rounding artefact instead of a signal that there is no edge
- Recomputing Kelly on the same data used to build the strategy, inheriting its optimism
- Ignoring that correlated positions make the aggregate Kelly far smaller than the per-trade figure
- Confusing the Kelly fraction of capital to risk with the fraction of capital to deploy as notional
Professional usage
Serious quantitative traders use Kelly as a diagnostic ceiling rather than an execution setting. They estimate the win probability and payoff from out-of-sample results, deliberately shade those inputs downward to account for estimation error, and then bet a fixed fraction of the resulting f*, commonly a half or a quarter. For portfolios they compute a joint Kelly that accounts for correlations between strategies, since summing independent per-strategy Kelly fractions dangerously over-bets the whole book. The unifying discipline is that Kelly informs how far below the optimum to sit, never a licence to bet the optimum itself.
Key takeaways
- Kelly gives the fraction that maximises long-run geometric growth: f* = W − (1 − W) ÷ b
- Betting more than Kelly lowers growth and raises volatility, so it is a ceiling not a target
- Full Kelly assumes a perfectly known, stationary edge, which no real strategy has
- Half or quarter Kelly captures most of the growth with far smaller drawdowns and more robustness
- In backtesting, use Kelly to interpret how aggressive a fixed-fractional bet is, with conservative inputs
Frequently asked questions
What is the Kelly criterion?
What is the Kelly formula?
Why does Kelly maximise growth?
What is half-Kelly?
Why do traders not use full Kelly?
What happens if I bet more than Kelly?
How do I estimate W and b for Kelly?
Does Kelly work for trades with variable sizes?
Is the Kelly fraction the fraction of capital to deploy?
How does Kelly apply to a portfolio?
What drawdowns does full Kelly imply?
Can Kelly tell me not to trade?
Is Kelly the same as fixed-fractional sizing?
Why is Kelly fragile in real trading?
Voice search & related questions
Natural-language questions people ask about The Kelly Criterion.
What is the Kelly criterion in simple terms?
What is the Kelly formula?
Why do people use half-Kelly?
Is full Kelly a good idea?
Can betting too much actually lose money?
Does Kelly work for a whole portfolio at once?
Sources & references
Last reviewed 12 July 2026. Educational content only — not investment advice. Markets and rules change; verify current conventions with SEBI, NSE/BSE and your broker.