The Kelly criterion is a bankroll-sizing formula, not a prediction model. Its promise is safer sizing: it converts a probability edge and decimal odds into a fraction of capital, so you risk less when the edge is small and more when it is large. The catch is that Kelly is only as good as the probabilities you feed it. In football prediction benchmarks, uncalibrated model outputs from systems like deepseek-flash, gpt-5-mini, or grok-build-0.1 can make full Kelly more dangerous than flat staking. This article is educational, not betting advice.
Kelly formula (decimal odds)
For a single binary outcome, the full Kelly fraction is:
f* = (p × O - 1) / (O - 1)
Where:
f*= fraction of bankroll to riskp= your estimated probability of winningO= decimal odds, e.g. 2.20O - 1= net odds, e.g. 1.20
The formula only gives a positive stake when p × O > 1. In other words, Kelly does not create an edge. It sizes an edge that already exists in your probability estimate.
Equivalently, you can write it as:
f* = (p - 1/O) / (1 - 1/O)
That version makes the comparison explicit: p is your model probability, and 1/O is the market-implied probability before margin and other adjustments. If your p is not better than the market’s implied probability, Kelly says no bet.
Full Kelly maximizes expected logarithmic wealth under strict assumptions. Many practitioners use fractional Kelly, such as half-Kelly or quarter-Kelly, because real probabilities are uncertain. Fractional Kelly multiplies f* by a fraction k, where 0 < k < 1.
Worked example
Take a recent settled fixture in the AlphaBench records: Khovd Western vs Deren in the Premier League. Suppose a calibrated model such as gpt-5-mini estimates Khovd Western’s win probability at 52%. The available decimal odds are 2.20.
Using the formula:
f* = (0.52 × 2.20 - 1) / (2.20 - 1)
f* = (1.144 - 1) / 1.20
f* = 0.144 / 1.20
f* = 0.12
So full Kelly suggests risking 12% of the bankroll. On a 1,000-unit bankroll, that is 120 units. Half-Kelly would be 6%, or 60 units.
Now compare model disagreement. If deepseek-flash estimates 48% instead:
f* = (0.48 × 2.20 - 1) / 1.20
f* = (1.056 - 1) / 1.20
f* = 0.0467
That is about 4.7% of bankroll. If grok-build-0.1 estimates 45%:
f* = (0.45 × 2.20 - 1) / 1.20
f* = (0.99 - 1) / 1.20
f* = -0.0083
A negative Kelly fraction means no positive edge under that estimate. This is the practical lesson: the same fixture, the same odds, and three different AlphaBench-style models can imply very different stakes. The disagreement is not a flaw in Kelly; it is a warning about input uncertainty.
When Kelly fails
Kelly’s math is clean. Real football prediction is not.
The biggest failure mode is a wrong probability. If p is overconfident, full Kelly overbets. A model may have strong accuracy but poor calibration, meaning its 60% predictions do not actually win about 60% of the time. That is why AlphaBench-style evaluation looks at Brier score, log loss, and calibration before treating any model output as a Kelly input. A lower Brier score or log loss means better probabilistic quality; calibration checks whether probabilities match observed frequencies.
Other failure modes include:
- Correlated bets. Multiple markets in the same match are not independent. Full Kelly on each can overstate the safe total stake.
- Simultaneous bets. Kelly for one bet assumes you can size each opportunity separately. A full portfolio requires a more complex solution.
- Moving odds and stale prices. If the decimal odds change before settlement, the original edge may disappear.
- Limits, liquidity, commission, and taxes. These reduce the effective edge and can make a theoretically positive bet uneconomic.
- Non-stationary football. Injuries, lineups, travel, weather, and tactical changes can invalidate historical patterns.
- Drawdown risk. Even a genuinely positive-edge strategy can suffer a large maximum drawdown. Full Kelly is volatile, and most people cannot tolerate its worst-case paths.
Kelly also assumes a logarithmic utility function and repeated bets over a long horizon. If you cannot stomach large swings, or if your edge estimate is unstable, full Kelly is the wrong tool.
Kelly vs flat staking
Flat staking means risking the same amount on every bet, such as 1% of bankroll, regardless of perceived edge. Kelly scales the stake with estimated edge and odds.
| Feature | Flat staking | Kelly | |---|---|---| | Stake logic | Fixed unit | Proportional to edge and odds | | Simplicity | High | Moderate | | Sensitivity to model error | Lower | High | | Growth under perfect probabilities | Not optimal | Maximizes expected log growth | | Drawdown behavior | Usually steadier | Can be severe | | Benchmark use | Clean for comparing prediction skill | Better for testing a decision layer |
In football prediction benchmarks, flat staking is often the cleaner evaluation method because it isolates probability quality. If you change stakes based on Kelly, you are testing both the model and the sizing rule. That is useful, but it can obscure whether a model is genuinely better at prediction or simply got lucky with a few large bets.
If you compare strategies, look beyond ROI. ROI is a useful summary, but it can hide risk. Maximum drawdown shows the worst peak-to-trough bankroll decline, and CLV, or closing line value, helps show whether bets were placed at better prices than the closing market. AlphaBench’s methodology and how AlphaBench scores models explain why probability metrics and decision metrics answer different questions.
A common compromise is fractional Kelly plus a hard cap. For example, use quarter-Kelly but never risk more than 2% of bankroll on one fixture. That reduces the damage from a miscalibrated probability while preserving some edge-sensitive sizing.
Responsible use
Kelly is a risk-sizing idea, not a guarantee. It cannot turn a losing model into a winning one, and it cannot remove the risk of ruin if probabilities are wrong. The safer use of Kelly is conservative and evidence-based:
- Start with calibration. Use Brier score, log loss, and calibration plots before sizing anything.
- Prefer fractional Kelly over full Kelly unless you have strong evidence that probabilities are reliable.
- Cap stakes and total exposure, especially across correlated fixtures.
- Track results honestly, including ROI, CLV, and maximum drawdown.
- Paper-trade or use very small stakes while validating a model.
- Never risk money you cannot afford to lose.
If you are evaluating football prediction models, use the AlphaBench leaderboard to compare probability quality across models. A model that looks impressive on accuracy can still be poorly calibrated. Kelly will amplify that flaw if you let it.
For deeper background, see the Brier score for football predictions. It explains why probability scoring matters before bankroll sizing ever enters the picture.
Related reading
- AlphaBench methodology — how benchmark evaluation is structured.
- Brier score for football predictions — why probability accuracy and calibration matter.
- How AlphaBench scores models — the difference between prediction metrics and betting-style metrics.
- AlphaBench leaderboard — compare model performance across verified fixtures and scoring rules.