The Kelly criterion

The sizing formula that maximises long-run growth in theory, and why traders use only a fraction of it.

All courses · Lesson 5 of 10 · Intermediate · 4 min

The Kelly criterion gives the fraction of capital that maximises long-run growth when you know your win rate and reward-to-risk. In practice the inputs are only estimates, so Kelly is best read as an upper reference point, not a target.

The formula

f* = W − (1 − W) / R
same as f* = (b × p − q) / b, with b = R, p = W, q = 1 − W
W is the win rate and R is the reward-to-risk ratio, so a win pays R times the amount risked.

A positive f* means the strategy has a positive expectancy; zero or negative means Kelly says to risk nothing. Betting exactly f* each time maximises the long-term growth rate of capital in the model, but also produces large swings.

Worked example

Full Kelly and half Kelly

  1. Win rate W = 0.50 and reward-to-risk R = 2.
  2. Loss probability = 1 − 0.50 = 0.50.
  3. f* = 0.50 − 0.50 / 2 = 0.25, so full Kelly risks 25% of equity per trade.
  4. Half Kelly risks 0.5 × 25% = 12.5%.
  5. Check with the other form: (2 × 0.5 − 0.5) / 2 = 0.25.

Full Kelly is 25% of equity per trade here; half Kelly is 12.5%.

A few properties are worth knowing. Kelly rises with the win rate and with reward-to-risk, and it is zero exactly when the expectancy is zero: at W = 0.40 and R = 1.5 it is 0.40 − 0.60 / 1.5 = 0, which is the break-even point. Because it is a share of current equity, it also shrinks automatically after losses, like percentage sizing.

In the simulator, the risk is f* multiplied by the Kelly fraction you choose (0.5 is half Kelly), and never below zero. Try fractions from 0.25 to 1 and compare the drawdowns.

Why full Kelly is dangerous in practice

Kelly assumes the win rate and reward-to-risk are known exactly. In reality they are estimated from a limited sample and may change over time, and errors in the estimate matter a great deal when you bet at the maximum.

Worked example

What overestimating the win rate does

  1. You estimate W = 0.50 with R = 2, giving full Kelly of 25%.
  2. The true win rate is 0.40, so the true Kelly is 0.40 − 0.60 / 2 = 0.10, or 10%.
  3. Expected log growth per trade is W × ln(1 + f × R) + (1 − W) × ln(1 − f).
  4. At f = 25%: 0.4 × ln(1.5) + 0.6 × ln(0.75) = 0.1622 − 0.1726 = −0.0104.
  5. At f = 10%: 0.4 × ln(1.2) + 0.6 × ln(0.9) = 0.0729 − 0.0632 = +0.0097.

Risking 2.5 times the true Kelly gives a negative growth rate of about −1.0% per trade, even though the strategy has an edge.

The size of the error matters more than its direction. Underestimating the edge costs you some growth, because you risk less than the optimum, but the loss of growth is gentle. Overestimating it by the same amount costs far more, because the growth curve falls steeply once you pass the optimum. The asymmetry is the main argument for erring on the low side.

Sample size is the usual source of error. A win rate measured on 40 trades has a wide margin of uncertainty, and the reward-to-risk of real trades varies from trade to trade, whereas the formula assumes a single fixed value. Costs, slippage and changing market conditions push live results further from the backtest.

This is why fractional Kelly, commonly half or a quarter, is used: it gives up some theoretical growth in return for smaller drawdowns and some protection against estimation error.

Key takeaways

  • Kelly's f* = W − (1 − W) / R, and it is zero or negative when the expectancy is not positive.
  • Full Kelly maximises growth only if the win rate and reward-to-risk are known exactly.
  • Overestimating your edge can push full Kelly into negative growth, so most traders use a fraction.
  • Treat QuantRaider's adjusted Kelly and recommended risk as sample-based estimates, not guarantees.

Sign in to save your progress and quiz results.