A model that is right 60% of the time can still lose money, and a model that is right 53% can make it. The number that decides which is the price — here is how to read it.
April 28, 2026 • Reading time: 6 minutes • By Derek Shaw, writer on betting models and market pricing
Every few months a new tool promises to predict results with some impressive accuracy figure attached. The figure is almost always real and almost always irrelevant, because accuracy on its own does not determine whether a bet makes money. The price does.
This article is about that distinction, and about the one measurement that actually separates a model with an edge from one without.
The margin is the opponent
A bookmaker's prices imply probabilities that sum to more than one. That excess is the margin, and it is the reason a bettor who is right exactly as often as the market still loses steadily.
The calculation takes seconds. Convert each decimal price to its implied probability by dividing one by the price, add them up, and subtract one. Two sides priced at 1.91 each imply 52.4% apiece, summing to 104.8% — a margin of about 4.8%. On a three-way market, add the third.
That number is the hurdle. A model must be better than the market's implied probabilities by more than the margin before a single unit of profit exists, and the margin is charged on every bet regardless of the outcome.
Why accuracy figures mean nothing alone
Consider a model that correctly picks the winner 60% of the time. Whether it profits depends entirely on the prices it was betting into.
| Hit rate | Average price | Break-even hit rate | Result |
|---|---|---|---|
| 60% | 1.50 | 66.7% | Loses money |
| 60% | 1.80 | 55.6% | Makes money |
| 53% | 2.00 | 50.0% | Makes money |
| 85% | 1.10 | 90.9% | Loses money |
The break-even rate is simply one divided by the price. Any accuracy claim published without the average price it was achieved at is unfalsifiable — the last row is the one that gets marketed hardest, because 85% sounds extraordinary.
The only honest scoreboard
Among people who bet seriously, the accepted test is not profit over a few hundred bets. It is whether the price obtained beat the closing price — the odds available immediately before the event starts.
The reasoning is that the closing line reflects all the money and information that arrived in the market up to that point, which makes it the most accurate probability estimate publicly available. Consistently taking a better price than the close is evidence of genuine edge; profit without it is more likely to be variance, and the sample needed to distinguish the two by results alone is larger than most bettors will ever place.
This is the same problem in a different costume: short runs do not distinguish skill from noise, and closing line value is the workaround the market's own pricing provides.
What models are genuinely good at
None of the above means modelling is pointless. It means the claims worth believing are narrower than the ones being sold.
- Consistency. A model applies the same criteria to the thousandth match as to the first. Human judgement drifts with fatigue, recent results and how much is at stake on the bet in front of it.
- Volume. Processing every fixture in a league across several seasons is trivial for software and impossible for a person, and that breadth is where small pricing inefficiencies become visible.
- Calibration. A well-built model can state that events it prices at 30% occur close to 30% of the time. Intuition is systematically overconfident and has no way of checking itself.
- Freedom from the result. A model does not feel the last loss, which removes the single most expensive behaviour in betting — the stake pattern that follows one.
Where they fail, predictably
Models fail at exactly the points where the data stops. Late team news, weather at kickoff, a manager resting players before a cup tie, motivation in a dead-rubber fixture — these appear in the market price long before they appear in any dataset, which is a large part of why the closing line is hard to beat.
They also fail through overfitting, which is the standard failure of any system trained on history. A rule with enough parameters describes the past perfectly and predicts nothing, and the published version is always the one that happened to survive the sample.
And there is a failure specific to the machine-learning era: a model trained on historical closing prices is being trained to reproduce the market. It will be accurate and it will have no edge, because agreement with the price is what it was optimised for.
The gap between a model's output and a bettor's behaviour is where most of the value leaks out. A system followed on winning weeks and overridden on losing ones is not the system that was tested — it is intuition with extra steps, and it inherits none of the discipline that made the model worth building. Overriding a rule because a run feels wrong is the same instinct that produces lucky seats.
A practical position
Three things follow from all of this, and they are worth more than any prediction service.
- Compute the margin on the market before assessing any tip. If it is wide, the tip needs to be far better than it looks.
- Record the closing price on every bet placed. It costs nothing and it answers the question that results take years to answer.
- Treat any accuracy claim without an accompanying average price as marketing, and any service unwilling to publish its closing line record as having answered the question.
The staking side matters as much as the selection side, and it is the part no model handles on its own — a real edge staked badly still goes to zero. Background on how analytics has actually been applied to professional sport is collected at the MIT Sloan Sports Analytics Conference.
Published April 28, 2026 • BetBeast Blog