How Football Predictions Are Made: From Raw Data to Three Probabilities
A football prediction is produced by turning a match into measurable inputs and compressing those inputs into three probabilities. Four families of data go in — fixtures and results, performance metrics, availability and squad state, and market prices. A model uses them to estimate how many goals each side is likely to score, expands that into a distribution over every plausible scoreline, and collapses the distribution into three numbers: home win, draw, away win, summing to 100%. Those three numbers describe what would happen across many repetitions of the same situation, not what will happen in the single match in front of you. Between publication and settlement, team sheets land and prices move, which is why a prediction only counts as checkable if it carries a publication timestamp and the price it was quoted against. The rest of this guide walks that pipeline end to end.
Step One: The Four Families of Input
Fixtures and results. Who plays whom, home or away, on what date, and what happened last time. This is the cheapest layer to obtain and the one every other reader of the sport has already seen. On its own it prices nothing that is not already priced.
Performance data. Shots, shots on target, expected goals, possession, territory, set-piece volume. These describe how a side actually played rather than how the scoreboard ended. A single match carries roughly ten to fifteen shots per side, so single-match figures are estimates built on very few events; the signal stabilises over a ten to fifteen match window.
Availability and squad state. Injuries, suspensions, rotation risk, and finally the confirmed starting eleven — which in most leagues becomes public about sixty minutes before kickoff. This is the highest-value pre-match input and the one with the shortest usable window.
Market prices. Quoted prices for each outcome, updating continuously. This input is different in kind from the other three: the first three describe the match, the fourth describes what everyone else already concluded about the match. Treating them as interchangeable is the single most common analytical error in this field.
Step Two: From Inputs to Three Probabilities
Most published football models, whatever else they add, pass through three stages.
- 1.Estimate a scoring rate for each side. The model produces an expected goals figure for the ninety minutes — say 1.55 for the home side and 1.10 for the away side — built from performance data, adjusted for opponent strength, venue and squad state.
- 2.Expand the rates into a distribution over scorelines. Assuming goals arrive roughly at random through the match, a Poisson-style distribution converts those two rates into a probability for every scoreline: 0-0, 1-0, 2-1, and so on.
- 3.Collapse the scorelines into outcomes. Add up every scoreline where the home side scores more, and that total is the home win probability. Repeat for level scores and for away-heavy scores.
Run the arithmetic on 1.55 and 1.10 and you get roughly 48% home, 25% draw, 27% away. Change either rate by a tenth of a goal and all three move. That sensitivity is the point: the three headline probabilities are downstream of two estimates, and the quality of the whole prediction is the quality of those two estimates.
Step Three: A Probability Is Not a Forecast of One Match
A 48% home probability does not say the home side will win. It says that across many repetitions of this situation, something close to 48 of every 100 end in a home win. The match in front of you happens once.
Two consequences follow, and both are routinely missed.
A single result cannot grade a prediction. The 52% of outcomes where the home side fails to win are not errors. They are the stated majority of the distribution. Grading predictions one match at a time measures luck and calls it skill.
The largest of the three numbers is not "the prediction". When the split is 40 / 30 / 30, the top line is merely the least unlikely of three unlikely things. A model that reports only its favourite has thrown away most of what it computed.
What can be graded is calibration: collect every entry a model marked between 45% and 50%, count how many actually happened, and check whether the realised rate lands inside that band. That test needs hundreds of entries, not a handful — the reasoning behind it is set out in is AI football prediction accurate.
Step Four: What Happens Between Publication and Settlement
A prediction is published into a world that keeps moving. The timeline below is why a timestamp is not bureaucracy.
| Stage | Typical timing | What changes | Effect on an already-published entry |
|---|---|---|---|
| Publication | T | Nothing yet | Probability and quoted price are pinned |
| Team sheets | About 60 min before kickoff | Lineups, late fitness | Model re-estimates; the published entry does not change |
| Kickoff | T0 | Pre-match window closes | Entry is now locked against reality |
| In play | 0–90+ min | Score, cards, substitutions | Live probabilities diverge from pre-match ones |
| Full time | About T0 + 105 min | Result is final | Settlement input exists |
| Settlement | After full time | Nothing | Hits and misses are booked together |
The load-bearing rule in that table is the third column of row one: a published entry is never edited afterwards. Our entries are mirrored to a public git repository at github.com/oddsflowai-team/clawsportbot-protocol/tree/main/record, where the commit history is itself the tamper evidence, and settled entries — hits and misses alike — sit together on our public prediction record. Those counts move every day as fixtures settle.
How to Tell a Modelled Prediction From a Hand-Waved One
Five questions, in order of how much they discriminate.
- 1.Is there a probability? "I like the home side" contains no quantity that calibration can ever test. A number can be wrong; an opinion cannot even be checked.
- 2.Is there a price attached? A probability with no price is an opinion about a match. A probability against a price is a claim about a disagreement — and only the second one is falsifiable in the dimension that matters. The conversion between the two is covered in understanding football odds.
- 3.Is there a publication timestamp, and is it earlier than kickoff? Compare it to the scheduled kickoff, never to the final whistle.
- 4.Are the misses on the same page as the hits? A record that only displays its wins is a marketing surface, not a record.
- 5.Is the settlement rule written down in advance? Pushes, half-win outcomes and abandoned fixtures all need a stated treatment, or the counting can be adjusted after the fact. Ours is written out in how we count our record.
A prediction that clears all five may still be wrong about a match. A prediction that fails any of them cannot be shown to be right about anything.
FAQ
How accurate can football predictions get? There is a ceiling, and it is set by the sport rather than by the model. Football is low-scoring and heavily influenced by single events, so a large share of every result is irreducible noise. The honest target is calibration — probabilities that come true at the rate they claim — not a high hit rate on individual matches.
Why do the three quoted prices imply more than 100%? Because the quoted prices carry the operator's margin on top of the underlying probabilities. Stripping it out is arithmetic, and the different ways of doing it disagree most on the outsider — worked through in how to read football odds deeply.
Is a model better than a well-informed human? At reading one match, often not. At restating a read as a calibrated probability, against a live price, across many fixtures at once, under a fixed rule, without fatigue — yes, and operationally that is the comparison that decides anything.