The Kelly Criterion in Sports Betting: Arithmetic, Not Prophecy

The Kelly Criterion: Arithmetic, Not Prophecy

The Kelly Criterion answers one question with precision: given a win probability and a posted price, what fraction of a bankroll maximizes long-run growth? It is arithmetic, not a prediction. The hard part is the estimate.

The formula, without the mystique

Convert the American price to decimal odds d. The edge per unit is p × (d − 1) − (1 − p), where p is your win probability. The full Kelly fraction is that edge divided by (d − 1). Two lines. No oracle.

Consider a −110 price, where d ≈ 1.909. If your model says p = 0.54, the edge is 0.54 × 0.909 − 0.46 ≈ 0.031. Full Kelly is 0.031 / 0.909 ≈ 3.4% of the bankroll.

Now move the probability two points to p = 0.52. The edge goes negative. The stake becomes zero.

At −110, the break-even probability is 1 / 1.909 ≈ 52.4%. Two points of opinion is the entire distance between a 3.4% stake and walking away. That is how tight the margins are.

Kelly is brutally sensitive to its input

Feed the formula a number that is confidently wrong and it will not politely lose a little. It will recommend a stake large enough to damage the bankroll on the way to being corrected.

That sensitivity is not a flaw in the math. It is the math telling you exactly how much your estimate is worth. Most casual criticism of Kelly is really criticism of unexamined inputs.

Why fractional Kelly exists

Full Kelly does not survive contact with reality. Practitioners use a fraction, typically one-quarter. The bankroll grows slower, but the strategy survives the gap between a model's stated probability and its real accuracy.

We run quarter-Kelly—a fraction of 0.25—with a hard ceiling of 2% of the bankroll per position, enforced in the sizing code itself. No projection can bet its own confidence up beyond that. The honest answer to "how wrong could this number be" is always "wronger than the formula assumes."

Zero is a result, not a failure

If the model detects no edge at the posted price, the output is zero. This is the most underused result in Kelly. It forces you to walk away when the math is not there. It makes an empty edge mathematically uninteresting rather than emotionally tempting.

A sizing system that never returns zero is not a sizing system. It is a habit with decimals.

Price is an input, so price quality is part of the discipline

The same 54% estimate that produces zero at −110 can produce a small positive guide at a better number elsewhere. That is why line shopping is a modeling behavior, not a bargain-hunting one. Measuring whether your estimates beat the closing price is a separate, checkable habit—the closing line value explainer covers how that audit works.

The danger of misuse

Most bad reputations for Kelly stem from three errors.

First, treating the output as a pick. The formula has no opinion on who wins, only on how much an edge is worth at a given price.

Second, feeding it the market's own probability. If your p is just the price restated, the guide is zero by construction. That zero is the formula working, not failing.

Third, applying full Kelly to correlated positions. Four bets that move together are really one oversized bet in a costume. That is exactly why per-position ceilings and portfolio-level checks exist beside the arithmetic.

A fourth misuse is quieter and more seductive: updating the probability after the result lands. The grade belongs to the written probability at stake time. Resulting—judging the size by the outcome instead of the process—turns a discipline back into a story.

Think of Kelly as discipline, not optimization

It forces a written probability before a size. It converts stake decisions from feeling into something a reader can check. It is a sizing guide built out of the assumptions you were already making, which is exactly why it exposes them: once the probability is written down, it can be graded.

Ours are, publicly, in the performance ledger, with the methodology that produced them beside them.

What Kelly cannot do is fix a bad estimator, price correlated legs as if they were independent, or substitute for honest uncertainty about your own uncertainty. It sizes the belief you bring. Garbage in, precisely sized garbage out.

The operating standard

If you are not sizing by a consistent standard, you are gambling. If you are, you are operating a desk.

Educational research only — model estimates and sizing guides, not picks or betting advice.

All figures are estimates. Past analysis is not a guarantee of future results. Not betting advice.