The Kelly criterion in football betting: how much to stake on an edge
Published
Most betting content obsesses over what to back and ignores the question that actually decides whether a bettor survives: how much. The Kelly criterion in football betting is the classic answer — a formula from a 1956 Bell Labs paper that converts a probability and a price into a stake size. Used carefully, it is the most principled staking framework there is. Used naively, it is one of the fastest known ways to destroy a bankroll, because it amplifies every error in the probabilities you feed it. This guide covers both halves: the math, and the honesty the math demands.
What the Kelly criterion actually says
For a simple back bet at decimal odds, the Kelly stake — as a fraction of your total bankroll — is:
f = (p × o − 1) / (o − 1)
where p is your estimated probability of the bet winning and o is the decimal odds. The numerator, p × o − 1, is your expected profit per unit staked — your edge. The denominator, o − 1, is the net odds you’re being paid. Kelly stakes your edge divided by the payout.
Three properties fall straight out of the formula:
- No edge, no bet. If p × o ≤ 1, the numerator is zero or negative and Kelly says stake nothing. The formula never bets for entertainment.
- Bigger edge, bigger stake. Double the edge at the same odds and the stake doubles.
- Longer odds, smaller stake. The same edge at odds of 5.00 gets a much smaller stake than at 1.50, because the payout structure is more volatile.
Kelly isn’t arbitrary: it is the stake size that maximises the long-run growth rate of a bankroll, assuming the probabilities are correct. Bet more than Kelly and long-run growth falls; bet far enough beyond it and growth turns negative even though every individual bet has positive expected value. That last sentence is worth rereading — you can lose money over time on winning bets, purely through oversizing.
A worked football example
Suppose a model prices an away win at 40%, and a bookmaker offers 2.80 on it. Converting the price to a probability (the arithmetic is in our guide to converting odds to implied probability), 2.80 implies about 35.7% before margin — so if the 40% estimate is right, there’s a genuine edge.
Plug into the formula:
- Edge: 0.40 × 2.80 − 1 = 0.12 (12% expected profit per unit staked)
- Net odds: 2.80 − 1 = 1.80
- Kelly fraction: 0.12 / 1.80 = 6.7% of bankroll
On a €1,000 bankroll, full Kelly says stake €67 — on a bet that loses 60% of the time. That discomfort you feel is correct, and it is the central practical problem with Kelly: for edges that look modest on paper, the recommended stakes are aggressive.
Note also what happens if the edge shrinks. At a 38% estimate the Kelly stake drops to 3.6%; at 36% it’s 0.4%; at 35.7% — the market’s own de-vigged view — it’s zero. The entire stake rests on a few percentage points of disagreement with the market.
Why full Kelly is more dangerous than it looks
The formula’s optimality proof contains an assumption that never holds in football: that p is the true probability. Your 40% is an estimate from a model or a judgment, and it carries error. Kelly is brutally asymmetric about that error — overestimating your edge is far more costly than underestimating it, because oversized stakes compound losses geometrically. Staking double the true Kelly fraction has an expected growth rate of zero: all the edge is burned as variance.
And betting markets make overestimation the default state. The odds you beat are set by a market that is, on average, well informed. When your probability disagrees with the market, some of that gap is your insight; some of it is your error. Full Kelly stakes as if the gap were 100% insight.
Even with perfectly correct probabilities, full Kelly is a rough ride. Its bankroll paths routinely pass through drawdowns of half the bankroll or worse — mathematically expected behaviour, not bad luck — and few people can keep executing a strategy through that.
Fractional Kelly: paying for your own uncertainty
The standard remedy is fractional Kelly: compute the full Kelly stake, then bet a fixed fraction of it — a half, a quarter, an eighth. In the example above, quarter Kelly turns €67 into about €17.
The trade-off is heavily in your favour. Growth scales down roughly linearly with the fraction, but variance falls faster — half Kelly keeps around three-quarters of full Kelly’s growth rate at half the volatility. More importantly, fractional Kelly is an implicit hedge against estimation error: if your true edge is smaller than you think (it usually is), a quarter-Kelly stake may be close to the actual optimum. A useful way to read it: the fraction you choose is a confession of how much you trust your own probabilities. Quarter Kelly says “I believe my model, but I know it’s a model.”
Fractional Kelly still preserves the properties that make Kelly worth using — stakes proportional to edge, smaller stakes at longer odds, zero stake without an edge — which is more than flat staking can say.
Kelly criterion football betting only works with calibrated probabilities
Here is the part most Kelly guides skip: the formula is only as good as the p you feed it, and almost nobody measures the quality of their probabilities. A bettor whose “60%” bets actually win 53% of the time isn’t slightly wrong — under Kelly staking they are systematically oversized on every bet, and the formula converts that miscalibration into losses with maximum efficiency.
The fix is measurement. Calibration — whether your 60% claims win 60% of the time — is checkable against any recorded history of probabilistic predictions, and proper scoring rules like the Brier score summarise it in a single number. This is exactly why we publish every FootInsights prediction, win or lose, on our track record page: probabilities that feed staking decisions should be audited, not asserted. Before trusting any probability source with Kelly-sized stakes — a model, a tipster, your own judgment — demand the same evidence: a complete, append-only history and a scoring rule computed over it.
A practical sequence, in order:
- Record probabilities first, stakes later. Log a few hundred predictions with explicit probabilities before risking Kelly sizing on them.
- Check calibration. If your 60% bets win 53% of the time, shrink your probabilities toward the market’s before computing any stake.
- Start at a small fraction. An eighth or a quarter of Kelly; earn your way up only as the evidence accumulates.
- Recompute per bet. Kelly is a function of the current bankroll and the specific odds — it is not a flat percentage.
The right framework, pointed at the wrong inputs
The Kelly criterion is the right framework — stake in proportion to edge, never bet without one, respect the geometry of losses. But it is a lens that magnifies whatever you point it at: genuine edges and self-deception alike. The formula takes one input you control, and everything depends on whether that input has been tested against reality. Verified, calibrated probabilities plus fractional Kelly is a coherent strategy. Confident guesses plus full Kelly is just variance with a bibliography.
Want the arithmetic done for you? The Kelly criterion calculator turns your probability, the price and your bankroll into full, half and quarter Kelly stakes — and says “no bet” out loud when the edge isn’t there.