The Kelly Criterion: How Much to Bet
Finding a good bet is half the job. The other half — the one that quietly decides whether you survive — is how much to stake. The Kelly criterion is the math that answers it: bet more when your edge is bigger, nothing when it's gone, and never so much that one bad run wipes you out.
The problem Kelly solves
Imagine you have a real, repeatable edge. Bet too little and you barely grow your bankroll. Bet too much and a normal losing streak — and they will come — busts you before your edge ever pays off. There's a sweet spot in between that grows your money fastest over the long run without risking ruin. The Kelly criterion is the formula for that sweet spot.
The formula
Kelly tells you what fraction f of your bankroll to put on a bet:
b = decimal odds − 1 · p = your win probability · q = 1 − p
There's a cleaner way to read it. The top of that fraction is just your edge — your probability minus the price's implied probability — so:
Two things fall right out of this. If your edge is zero or negative, f is zero or negative — you bet nothing. And for the same edge, longer odds get a smaller fraction, because the bet is more volatile. Kelly bakes in "no edge, no bet" automatically.
A worked example
You back a team at +120 (decimal 2.20, so b = 1.20). Your model says they win 52% of the time; the price implies 45.5%. Your edge is +6.5 points.
| Step | Value |
|---|---|
| b (decimal − 1) | 1.20 |
| p / q | 0.52 / 0.48 |
| Full Kelly f = (1.20×0.52 − 0.48) / 1.20 | 0.12 → 12% of bankroll |
| Half Kelly | 6% |
| Quarter Kelly | 3% |
Full Kelly says bet 12% of your bankroll on one game. That already feels aggressive — and it is, which is the whole reason almost nobody bets full Kelly.
Why full Kelly is too much
Kelly is only "optimal" if your p is exactly right. In sports betting it never is — you're estimating a probability, not reading it off a card. And Kelly is brutally sensitive to that error: overestimate your edge and full Kelly overbets dramatically, turning a real edge into a blown bankroll. Even when your edge is perfectly estimated, full Kelly produces gut-churning swings — drawdowns of 50% or more are routine.
Practical bankroll rules
- Define your bankroll as money set aside for betting that you can afford to lose — not your rent.
- Use fractional Kelly (quarter to half) on top of a genuine edge, not a hunch.
- Cap any single bet at ~1–2% of bankroll regardless of what the formula spits out — a guard against a wildly overestimated edge.
- Re-base periodically, not after every bet. Recomputing your bankroll constantly amplifies swings.
- No edge, no bet. Kelly returns zero for a fair price. Most of the board is a fair price.
How Lakeshore Edge sizes bets
Every pick on the site carries a stake suggestion built on fractional Kelly (quarter-Kelly by default), computed from the model's edge and the actual price you'd bet. When the model's probability is below a coin flip, or the edge is thin, the stake is shrunk further or the pick is skipped outright — because the cost of overbetting a shaky edge is far worse than the cost of missing it. The sizing only matters if the edge is real, which is why we obsess over calibration and grade everything against the closing line.
FAQ
What is the Kelly criterion?
A formula for how much of your bankroll to bet to grow it fastest over the long run: f = edge / (decimal odds − 1). Bigger edge means a bigger bet; no edge means no bet.
How do I calculate my Kelly stake?
f = (b×p − q) / b, where b is decimal odds minus 1, p is your win probability, and q is 1 − p. Multiply f by your bankroll. Then bet a fraction of that (quarter or half) to stay safe.
Why not bet full Kelly?
Full Kelly assumes your probability is exact. It never is, and Kelly overbets hard when you overestimate your edge — plus the swings are violent even when you're right. Fractional Kelly keeps most of the growth with far less risk.
Is Kelly better than betting the same amount every time?
If your probabilities are accurate, Kelly grows a bankroll faster than flat betting. If they're shaky, flat betting is more forgiving. Fractional Kelly is the common middle ground.