Earlier lessons covered position sizing and risk of ruin, which argue for keeping risk per trade small. The Kelly criterion approaches the same question from the other side: if you do have an edge, what is the largest fraction of your account it makes sense to risk? This lesson explains the formula, what it assumes, and why it is better used as a ceiling than as a target.
Where it comes from
The idea comes from John L. Kelly Jr., a Bell Labs researcher who published it in 1956 in the context of information theory. It was later popularized for gambling and investing by the mathematician Edward Thorp, who applied it to blackjack and to markets. It answers a specific question: which bet size gives the highest long-run growth rate of capital when the same favorable bet is repeated?
The formula
For a trade that wins a multiple R of the amount risked with probability p, and otherwise loses the amount risked (1 − p = q):
Kelly fraction = p − q / R
In trading terms, the result is the percentage of your balance to risk per trade, measured to your stop-loss. A few examples:
- 55% win rate, 1:1 reward:risk: 0.55 − 0.45 / 1 = 10% of the balance per trade.
- 45% win rate, 2:1 reward:risk: 0.45 − 0.55 / 2 = 17.5%.
- 40% win rate, 1.5:1 reward:risk: 0.40 − 0.60 / 1.5 = 0%. There is no edge, so Kelly says to risk nothing.
You can try your own numbers in the Kelly criterion calculator. These percentages are far above what most traders risk, which is the first hint that full Kelly is not meant for everyday use.
Growth rate versus risk per trade (55% win rate, 1:1 reward:risk)
Illustrative model: expected log-growth per trade for a 55% win rate and a 1:1 reward:risk, where full Kelly is 10%. Growth peaks at Kelly, half Kelly keeps about three quarters of it, and at roughly twice Kelly the edge is cancelled out by volatility. It assumes known odds and independent trades, which real trading does not offer.
What it actually optimizes
Kelly maximizes the long-run growth rate of the account, not comfort or safety. Two features of the curve above matter most:
- Half Kelly keeps most of the growth. In this simple model, risking half the Kelly amount keeps roughly three quarters of the growth rate while cutting the size of the swings sharply.
- Overbetting is punished. The curve is not symmetric. Risking much more than Kelly lowers growth, and at around twice Kelly the expected growth is no better than zero, even though the edge is real. Past that point volatility eats the edge.
Why full Kelly is too aggressive in practice
- You do not know your true win rate and reward:risk. They are estimated from a limited sample, and estimates are noisy. If the edge you plug in is too good, the formula tells you to overbet, and being too high costs more than being too low by the same amount.
- The drawdowns are large. Risking 10% per trade means five losses in a row cost about 41% of the account (1 − 0.95). Even with a real edge, full Kelly is a rough ride, and many traders cannot stick with it. See the losing streak lesson.
- Real trades break the assumptions. Kelly assumes independent trades with fixed odds. In practice, rewards and risks vary, trades overlap or are correlated, and spreads, slippage and gaps change the result.
- Edges change. A strategy that worked in one market regime can stop working, which makes a past win rate a shaky input.
Fractional Kelly: a safer way to use it
Because of these problems, people who use Kelly at all usually risk a fraction of it, such as half or a quarter, and often cap the result at a fixed percentage as well. Many traders risk between 0.5% and 2% per trade, which is well below Kelly for most plausible edges. Here is a sensible way to use the number:
- Treat Kelly as an upper bound, not a target. If it says 10%, that does not mean 10% is a good idea. It means risking more than that is likely to hurt.
- Use conservative inputs. Take a win rate and reward:risk a little worse than your records show, ideally from at least 100 trades (see the data collection lesson).
- If Kelly is zero or negative, you have no measured edge. The answer is not a smaller size but to fix or stop the approach (see win rate and profit factor).
- Check the drawdown you could live with. Use the losing streak planner and the drawdown recovery calculator to see what the size means in a bad run.
Limits of the formula
- It needs known, stable odds, which markets do not provide.
- It ignores costs, leverage limits, margin requirements and the chance of gaps past your stop.
- It says nothing about whether your edge exists. A positive Kelly figure computed from a lucky streak is still luck.
- It describes the long run, and you live in the short run.