What Is the Kelly Criterion, and Why Full Kelly Is Rarely Used
The Kelly Criterion is a formula that converts an edge — stated as a win probability and a win/loss ratio — into one number: the fraction of your stake to risk on the next trade. It isn't a forecasting tool and it doesn't find an edge for you. It only answers what to do with an edge you already believe you have, and it answers that question unforgivingly: feed it a wrong edge and it will tell you, with total confidence, to bet too much.
What is the Kelly Criterion formula, in plain terms?
It's f = W - (1 - W) / R, where f is the fraction of your stake to risk, W is your win rate as a decimal, and R is your average win divided by your average loss.
Run through what each input actually is before the formula means anything:
- W (win rate): the share of trades that come out ahead, out of every trade taken — not the ones that "worked," the whole set, including the ones that quietly didn't.
- R (win/loss ratio): the average size of a winning trade divided by the average size of a losing trade, in the same units (percent of account, or R-multiples). A ratio of 2 means the average win is twice the average loss.
- f (the output): the fraction of capital Kelly says to risk on one trade to maximize the long-run growth rate of the account, assuming every future trade carries the identical W and R forever.
That last assumption — identical, forever — is where most of the real-world trouble starts, and it's covered below.
Why does Kelly call for betting MORE as the edge gets smaller, not less?
It doesn't — a smaller edge always produces a smaller f, never a larger one. The formula is doing the opposite of what it sometimes gets blamed for: it shrinks your position size automatically as either input gets worse, and it goes to zero or negative the moment the edge disappears.
Walk the formula through three edges side by side, holding the win/loss ratio fixed at 2:1 and varying only the win rate:
| Win rate (W) | Win/loss ratio (R) | Kelly fraction (f) |
|---|---|---|
| 45% | 2:1 | 17.5% of stake |
| 40% | 2:1 | 10% of stake |
| 33% | 2:1 | 0% (no edge — formula goes to zero) |
At W = 33% with a 2:1 payoff, the math break-even point, Kelly correctly says to bet nothing — there's no edge left to size. Below that, f goes negative, which in practice means "this isn't a bet to take, or you should be on the other side of it." The formula isn't aggressive by nature; it's precisely as aggressive as the stated edge, no more.
Why is "full Kelly" almost never traded in practice?
Because the formula that maximizes long-run growth also happens to produce some of the worst live drawdowns a trader will ever sit through on the way there, and because it only works if the edge inputs are exactly right — which they never quite are.
Two separate problems compound here:
- Full Kelly's ride is brutal even with a perfectly correct edge. The math that proves Kelly maximizes long-run growth says nothing about the path getting there — simulated full-Kelly equity curves routinely show 50%+ peak-to-trough drawdowns along the way to the eventual gain, which is the kind of stretch that gets most live accounts closed by the trader long before "long run" arrives. What maximum drawdown actually measures is the honest way to look at exactly that kind of stretch before deciding you can sit through it.
- The inputs are estimates, not facts, and Kelly has no humility built in.
WandRcome from a measured track record — which means they carry all the same honesty problems a track record carries anywhere else. Overstate the win rate by even a little and Kelly doesn't gently overshoot; it compounds the error, because the fraction it recommends is itself a function of an input that was already wrong.
How much does an overstated edge actually distort the Kelly number?
Materially — a small error in the win-rate input produces position sizes that aren't just a little too big, because f moves faster than W does near the edges of the formula's range.
Take the row where W = 40% and the ratio is 2:1: true Kelly calls for 10% of stake. Now suppose the real edge was overstated by a measurement problem that's common to small or curated samples — a cherry-picked track record or a sample too small to mean much can each inflate a measured edge without the trader doing anything deliberately dishonest. Push the believed W from 40% up to 48% on the same 2:1 ratio and Kelly's answer jumps to 22% of stake — more than double, off an edge-measurement error that would look unremarkable on a short track record. This is the real argument for sizing down from the formula's raw output, not a hunch about caution for its own sake.
What is "fractional Kelly," and why do most practitioners use it?
Fractional Kelly means deliberately betting some fraction of what the formula outputs — commonly half ("half Kelly") — which gives up some of the theoretical long-run growth rate in exchange for a dramatically smoother ride and much more room for the edge estimate to be wrong.
Half Kelly is the common compromise for a specific, calculable reason: it captures roughly 75% of full Kelly's theoretical growth rate while cutting the worst expected drawdown by roughly half. The trade being made explicitly is growth rate for survivability — betting less than the formula's own mathematically optimal number, on purpose, because the optimal number assumes an edge you can never measure with total certainty and a stomach for drawdowns almost no live account has.
Questions traders ask
Does the Kelly Criterion tell me whether I have a real trading edge?
No — it has nothing to say about that question. It takes W and R as given inputs and tells you what to do with them; whether those two numbers reflect a real, repeatable edge or a lucky, overfit, or cherry-picked stretch of history is a separate question entirely, covered in what curve-fitting actually is and how many calls before a track record means anything.
Is a higher win rate always better for the Kelly fraction than a higher win/loss ratio?
Not necessarily — the two inputs trade off against each other, and win rate vs. expectancy walks through exactly why a lower win rate paired with a strong win/loss ratio can carry more real edge than a high win rate with small, choppy wins.
What happens if I use the Kelly fraction but my edge changes over time?
The formula has no memory of that — it assumes the W and R you feed it hold steady forever, and a real edge that decays, as most eventually do, means yesterday's Kelly fraction is sizing tomorrow's trade off a stale number. This is one more reason a fraction of Kelly, re-measured on a rolling basis, is used far more often than a single full-Kelly calculation set once and left alone.
Can the Kelly Criterion put a number on position size directly from the risk framework I already use?
It's a narrower tool than a full risk framework, not a replacement for one — risk in four dimensions covers size, stop, target, and time together; Kelly only answers the size question, and only once the other three and an honestly measured edge are already in hand.
The house teaches this as one honest input among several, inside the same method documented at the methodology hub — a formula is only as good as the track record it's fed, which is the whole subject of this lane.
Keep reading
Get them free