Finding an edge is half the job. The other half is a question most bettors answer with vibes: how much should I risk on it? Bet too little and the edge barely matters; bet too much and one routine losing streak ends the experiment. The Kelly criterion — a 1956 Bell Labs formula that quants and card counters have leaned on ever since — answers it with arithmetic.
What Kelly actually does
Kelly takes two inputs — your win probability and the offered price — and returns the percentage of your bankroll that maximizes long-run growth. The intuition: bigger edges and better prices earn bigger stakes, and because the stake is always a percentage of your current bankroll, bets shrink automatically during a downswing. That built-in throttle is why full Kelly can never technically bust you — each loss makes the next stake smaller.
A worked example
You like a side at +100 (even money) and your backtest says it wins 54% of the time. Kelly's formula — edge over odds — gives (0.54 × 1 − 0.46) / 1 = 8% of bankroll. At −110 with the same 54%, the thinner price cuts the stake to roughly 3.4%. Change the win probability to 52.4% at −110 — exactly break-even — and Kelly says 0%. No edge, no bet: the formula refuses to stake a coin flip against the vig. The Kelly calculator does all of this for any price and probability.
Why nobody sharp bets full Kelly
Full Kelly is optimal only if your win probability is exactly right — and in sports betting it never is. A backtested 54% might truly be 52%, and overbetting a real-but-smaller edge is far more destructive than underbetting it. Full Kelly also rides a brutal rollercoaster: a one-third bankroll drawdown is routine. The professional compromise:
- Half Kelly — about 75% of the optimal growth rate with roughly half the volatility.
- Quarter Kelly — the most common pro sizing, and what Covrr's own stake suggestions are built on. Most of the growth, a survivable worst week.
When to use it — and when flat staking wins
Kelly rewards bettors whose win probabilities come from somewhere real: a backtested model, hundreds of matched games, a measured hit rate. If your probabilities are guesses, Kelly just scales the guessing — flat one-unit staking will serve you better until the estimates earn trust. That's the honest order of operations: build the record first, then let Kelly size it. Start the record here.