The Kelly Criterion in Football — and Why Full Kelly Is Almost Always Wrong

Kelly is the formula most cited and least correctly applied in this field. It comes from information theory, it has an elegant optimality property, and it becomes dangerous the moment you stop reading the assumptions.

The formula

f = (b · p − q) / b

• f — the fraction of bankroll the formula suggests
• p — your probability for the outcome
• q — 1 − p
• b — decimal odds minus 1, the net profit per unit

Worked example. Probability 45%, odds 2.40, so b = 1.40:

f = (1.40 × 0.45 − 0.55) / 1.40
  = (0.63 − 0.55) / 1.40
  = 0.08 / 1.40
  = 5.7%

The formula suggests 5.7% of the bankroll.

What Kelly actually promises

Kelly maximises the expected logarithmic growth rate of a bankroll given that your probability is exact and that you repeat the same class of wager indefinitely.

Both conditions are heavy in football:

• Your probability is a model estimate, not a fact. It carries an error bar.
• You do not repeat the same wager. Every match differs, and the model's bias differs between leagues and between data regimes.

Feed a false assumption into a formula that maximises a growth rate, and you get a tool that maximises the damage from the falsehood too.

The core problem: asymmetry

This is what makes Kelly dangerous rather than merely imprecise.

Suppose the true probability is 42% but the model says 45%, at odds of 2.40.

Belief	Kelly stake	Actual EV at the true 42%
p = 45%	5.7% of bankroll	0.42 × 2.40 − 1 = +0.008
p = 42% (truth)	0.8% of bankroll	+0.008

A three-point overestimate multiplied the stake sevenfold while the real edge was close to zero. The estimation error and the stake size move in the same direction — the single worst property a risk-management rule can have.

In the opposite direction, underestimating only costs you an opportunity. There is no symmetry.

Why fractional Kelly is the sensible default

The standard response is to divide the output. Half Kelly, quarter Kelly, sometimes less.

Fraction	Relative volatility	Growth rate if the estimate is right
Full Kelly	Highest	Maximum
Half Kelly	Substantially lower	Roughly three quarters of maximum
Quarter Kelly	Very low	Roughly half of maximum

The property that makes this a good trade: growth falls more slowly than volatility does. Halving the stake costs about a quarter of the theoretical growth rate while dramatically reducing drawdown depth.

And when your estimate is wrong — which happens — half Kelly turns an expensive mistake into a survivable one.

An under-appreciated wrinkle: same edge, different sizes

People assume Kelly responds only to the size of the edge. It also responds to the odds, in a way that is not intuitive.

In every row below, the edge in percentage points is identical — five points above the implied probability:

Odds	Implied	Your probability	Kelly stake
1.50	66.7%	71.7%	~15.0%
2.00	50.0%	55.0%	~10.0%
3.00	33.3%	38.3%	~7.5%
6.00	16.7%	21.7%	~6.0%

Same edge, sizes differing by a factor of two and a half. At short odds each unit risks a lot to win a little, so the formula is willing to commit more of the bankroll when the hit probability is high.

The practical problem: short odds are exactly where a small overestimate is most destructive, and they are where Kelly proposes the largest fraction. Fifteen percent of a bankroll on a single decision is a number very few people would choose deliberately.

Where Kelly simply does not apply

Accumulators. Kelly assumes one binary event. A five-leg slip breaks that twice over: the legs are usually correlated, and the combined probability fed into the formula is a product that assumes independence. If the dependence is not handled, the input is wrong and the output is wrong — generally too high.

Small edges. A two-point gap produces a small suggested stake. The temptation is "small stake, small risk". But the edge underlying it sits inside the error bar, so the small stake reflects noise. The right answer is not a small bet, it is no bet.

Thin markets. Kelly assumes you know something. In a market with little money, both the price and our estimate are less reliable — neither party knows much.

Simultaneous decisions. Kelly sizes one wager at a time. Three decisions of 6% each in a single weekend expose 18% of the bankroll, and the formula was unaware of the other two. Worse, if all three depend on the same match script, they are one exposure rather than three. The accepted remedy is not a multi-dimensional Kelly but a hard cap on total weekend exposure — a risk decision that lives outside the formula by design.

When the bankroll is not really a bankroll. Kelly discusses a fraction of capital allocated to this purpose. If the money is needed for something else, the percentage conversation is theoretical and the real decision belongs elsewhere.

What we do not do

To be explicit: the system does not recommend stake sizes and does not manage a bankroll. No suggested amount, no built-in Kelly output presented as a recommendation, no connection to any operator account, no stored passwords, no bets placed.

The reason is not technical. Stake size is the decision where a model error converts directly into financial damage, and the responsibility for it belongs to whoever owns the money. What the product does provide — calibrated probability, fair odds, a minimum acceptable price, a reasoned decision state and slip diagnostics — is the input. The size is not.

Two lessons worth keeping even if you never use the formula

Size should track the size of the edge, not the strength of the feeling. If one decision rests on a six-point gap and another on a two-point gap, there is no reason for them to be the same size. Most people vary size according to how much they like a match, which is precisely the wrong input.

When the edge is unclear, the right size is zero. Kelly returns a negative number when there is no edge — that is, it advises not acting. That is not a quirk, it is the answer. And it is the same logic behind the PASS states: a gap that fails to clear the uncertainty budget is not a small edge deserving a small stake, it is the absence of an edge.

Summary

The formula is mathematically correct and extremely sensitive to input quality. In football the input is an estimate, so full Kelly is almost always too large. Fractional Kelly is a reasonable compromise where a decision is warranted at all. In accumulators the assumptions break and creative adaptation is worse than abstention. And no staking rule turns a bad decision into a good one, or guarantees any outcome.

---

18+. WinPIQ is an analysis tool, not advice and not a promise. Betting can be addictive and money can be lost. Only stake what you can afford to lose, and if betting stops being entertainment, seek help. WinPIQ is not affiliated with Winner or the Israeli Council for the Regulation of Sports Betting.