The formula
f* = p − q / b
Where p is the win rate, q is 1 − p, and b is the average win divided by the average loss.
A 45% win rate with wins averaging 320 and losses 200 gives b = 1.6, so f* = 0.45 − 0.55/1.6 = 0.106 — about 10.6% of the account per trade.
That number is correct and almost nobody should use it.
Why full Kelly is a reference point, not an instruction
Kelly maximises the long-run growth rate if the inputs are exactly right. Two things break that in practice.
The drawdowns are extreme. Full Kelly produces peak-to-trough falls above 50% as a routine feature, not a tail event. The mathematics is indifferent to that; the person watching it is not, and a sizing scheme abandoned halfway through is worse than a smaller one followed.
The inputs are estimates. This is the more serious problem. Kelly amplifies input error rather than smoothing it — overbetting is punished far more harshly than underbetting, and the penalty is asymmetric. A win rate five points too optimistic can roughly double the recommended fraction, and betting twice the true Kelly has a negative expected growth rate despite a genuine edge.
Which is why the practical settings are:
| Size | Growth vs full Kelly | Volatility |
|---|---|---|
| Full | 100% | reference |
| Half | ~75% | about half |
| Quarter | ~44% | about a quarter |
Half Kelly keeps three quarters of the growth for half the variance. That is why it is the standard, and it is also a hedge against your own estimates being wrong.
When it returns zero or negative
That is a real answer: at these numbers there is no edge, and no position size makes a negative expectancy positive. The row showing the win rate needed for any edge tells you how far away you are.
What it does not know
Kelly assumes each bet is independent and repeated. Trading violates that in the ways that matter: losses cluster, correlated positions are one bet, and a bad regime is a run of related outcomes rather than a string of independent draws.
It also assumes you can size continuously. Whole shares, whole contracts and minimum lot sizes all round the answer, sometimes substantially — futures sizing is the clearest case.
Treat the result as a ceiling on sensible size rather than a target. If your current risk per trade is well below quarter Kelly you have room; if it is above full Kelly, the arithmetic says you are overbetting a genuine edge into a losing one.