Prediction models in sports betting are real, useful and nothing like the marketing. They estimate probabilities better than intuition and worse than the market, and the margin they have to beat is published on every slip.
What the models actually do
Most of them estimate a scoring rate and convert it into match probabilities, often through a Poisson distribution for goals, with team strength derived from shot quality measures such as expected goals. Modern versions add player availability, rest days, travel and weather. The output is a probability, not a result.
The bookmaker does the same thing with more data and then adds a margin. That margin — the overround — is the number a model has to beat before it earns anything, and on a typical three-way market it is several per cent of every stake.
Where the models break
Four failure modes recur, and none of them is fixed by a bigger model:
- Thin data on the events that matter. A manager’s first match, a cup tie with a rotated squad, a promoted team — exactly the situations with the loosest prices have the least history.
- Overfitting a backtest. A system tuned on past seasons describes those seasons. Published results that start at the model’s best period are the standard tell.
- Market movement. By the time a public model finds value, the price often reflects the same information. A model competes with the market, not with a static line.
- Limits. An account that consistently beats the price gets its stakes cut or closed, which ends the strategy regardless of how good the model was.
How to read any claim about predictive accuracy
Ask for three numbers: the sample size, the closing odds the bets were placed at, and the yield against those closing odds. Accuracy alone means nothing — picking favourites is highly accurate and unprofitable. What counts is whether the bets beat the price available at the time.
| Claim | What to ask | Why it matters |
|---|---|---|
| «80 % accuracy» | On what market, at what odds? | Favourites hit often and pay little |
| «Profitable in backtest» | What period, and out of sample? | Tuned systems describe the past |
| «AI-powered» | What inputs and what output? | A probability is useful; a «guaranteed tip» is not |
| «Beats the bookmaker» | Versus opening or closing odds? | Beating closing odds is the real test |
| «Used by professionals» | Are their accounts still open? | Limits end most winning strategies |
Frequently asked questions
Can a model beat the bookmaker?
Sometimes, in narrow markets, before limits arrive. It has to beat the margin as well as the result, which is why most public systems do not.
Why is high accuracy not enough?
Because accuracy ignores price. Backing favourites is accurate and loses money; what matters is the yield against the odds you actually got.
Does more data fix the failures?
Not the main ones. Thin data on unusual fixtures, market efficiency and account limits are structural, not modelling, problems.
What is the overround?
The bookmaker’s margin built into the odds — the implied probabilities sum to more than one, and the excess is the house edge on that market.
Background: Poisson models, expected goals and the overround.
Related reading: the other verifiable-sounding promise, judgment under uncertainty, by hand, belief without a model and who builds what you bet on.