Finding a good price tells you what to bet on. It says nothing about how much. Stake too little and an edge earns almost nothing; stake too much and a normal losing run takes you out entirely. The Kelly criterion is the formula that answers the question properly.
The formula
fraction = (p × b − q) ÷ b
Where p is your estimated chance of winning, q is 1 − p, and b is the decimal odds minus 1 — the profit per dollar staked.
Say you rate something a 55% chance and the price is $2.00, so b = 1.00:
- (0.55 × 1.00 − 0.45) ÷ 1.00 = 0.10
- Kelly says stake 10% of your bankroll
On a $2,000 bankroll that is a $200 bet. If your 55% estimate is right, that is the stake that grows the bankroll fastest over time. If the formula returns zero or a negative number, it is telling you there is no edge and the correct stake is nothing.
Why nobody sensible uses full Kelly
Kelly maximises long-run growth, which is not the same as being comfortable to live through. A full-Kelly bettor swings violently: drawdowns of 50% or more are entirely normal, not a sign anything has gone wrong. Most people stop betting during one, which converts a theoretical optimum into a real loss.
The deeper problem is that Kelly assumes p is correct. It never is — it is an estimate, usually derived from no-vig fair odds that carry their own error. Overestimate your edge by a little and full Kelly overstakes by a lot, because the formula scales directly with the number you got wrong.
Fractional Kelly
The standard fix is to stake a fixed fraction of what Kelly says — half, quarter or less. Quarter Kelly on the example above turns a $200 bet into $50.
| Fraction | Stake on the example | Character |
|---|---|---|
| Full | $200 | Theoretically optimal, brutal in practice |
| Half | $100 | Most of the growth, much less swing |
| Quarter | $50 | Common default; forgiving of a wrong estimate |
Quarter Kelly gives up surprisingly little long-run growth for a large reduction in volatility, and — more importantly — it stays sensible when your probability estimate is off. It is what EdgeBoard's EV Finder shows for that reason.
Things worth being honest about
- Kelly is only as good as your probability. Applied to estimates you cannot justify, it is precise-looking nonsense — a formula does not make a guess accurate.
- It assumes you can bet repeatedly. The maths is about compounding over many bets. On a handful it says very little.
- It says nothing about correlated bets. Six bets in the same match are not six independent bets, and staking each at Kelly size stacks far more risk than it appears.
- No stake size makes a bad bet good. Sizing is what you do after finding an edge, not instead of finding one.