How AI football predictions work

A football prediction model turns a fixture into a probability for each outcome. Trained classifiers are ordinary technology. The interesting part is what happens between a raw model output and a percentage that means what it says, and most sites describe none of it.

Last updated
17 September 2026
Reading time
7 min read

What the model actually is

There isn't one model. ScorePredicts runs 100 trained models, exported to ONNX and held in memory, and each fixture is scored by several of them at once. Each market has its own: match result, goals for each side, both teams to score, corners for each side, cards for each side. A model that predicts corners well has learned something different from one that predicts goals, and asking a goals model about corners gives a worse answer than training a second model.

Each of those questions is answered by one model, the one trained for that market and that kind of competition. The models don't vote. Splitting the work this way keeps every model focused on a single target, and a weak result in one market can be traced to one model and retrained without touching the others.

"AI" in this category almost always means supervised learning on tabular match data. It doesn't mean a language model reading the news, and any site implying that is describing a different and much less reliable thing.

Why models are grouped by competition

The 100 models are organised into 19 groups by competition type and tier: domestic leagues across several tiers (with the second tier split by region), domestic cups, continental club competitions and national-team fixtures. Each fixture goes to the models trained on its own kind of football.

This matters more than it sounds. Home advantage, scoring rates, draw frequency and card counts all differ systematically between a major European league and a third-tier national division. One global model trained on everything learns an average of those environments, so it is slightly wrong everywhere, and most wrong at the edges where data is thinnest.

The grouping also decides which markets are available. Corner and card models only exist for groups with enough historical depth to support them. Other groups get a smaller set (match result, goals and both teams to score in some, match result and goals in others, match result alone in the thinnest) instead of a forecast built on too little data. Fewer markets with a real basis beat more markets without one.

What goes into a single prediction

The match-result, goals and both-teams-to-score models read 77 features per fixture. Corner models read 72 and card models 51, because what drives a corner count is different from what drives a goal.

The 77 break down roughly like this: 44 core form and strength features, 8 from the last five head-to-head meetings, 8 expected-goals features, 5 seasonal ones, 2 for the league's scoring environment, 2 for the referee, 2 flagging a promoted side and 6 for each club's pedigree (how much continental football it has played recently and how its current strength compares with its three-year average). Referee tendencies matter most for cards, whose models have their own referee inputs. Cards forecasts are only produced once the referee appointment is confirmed, so the referee's profile is always part of them.

Rest days and match context (whether it is a derby, how much the match matters, rotation risk in cup and continental fixtures) go in alongside, together with market price movement. Forecasts are refreshed several times a day, so a fixture scored on Monday isn't still shown on Saturday with Monday's inputs.

From raw model output to a stated percentage

A classifier's raw output looks like a probability but isn't one. It is a score between 0 and 1 that happens to be in the right order (higher means more likely) while being systematically too confident or not confident enough. Publishing it directly is the most common way a prediction site ends up with numbers that mean nothing.

Five steps sit between the model and the figure on the page. First, raw output is calibrated through an isotonic regression table fitted against graded results; this is the step that makes a stated 70% correspond to roughly 70% in reality. Next, the calibrated probabilities are shrunk to compress the extreme tails, where models are least reliable. When one side is clearly favoured, a favourite-bias correction trims the probability models tend to over-assign to the underdog and the draw. The result is then blended with the market prior. Finally, double chance is recomputed from the blended match-result numbers instead of being predicted separately, so the two can never contradict each other.

Each of those steps moves the figure, and a site that skips them is showing you raw classifier output with a percent sign on it. The trust score on a Best Pick is this final probability, further adjusted for how that market has performed in the past. On a Value Analysis item it is the model's own probability before the market blend, because that view exists to compare the model with the market.

Why the output is anchored to market prices

This step is worth stating plainly, because it's the one a prediction site is least likely to admit. For match result, goals and both teams to score, model probabilities are blended with the probabilities implied by market prices before publication, weighted by competition tier. Corner and card forecasts aren't blended, because prices in those markets are the least reliable reference.

Closing market prices are the strongest publicly available forecast of a football match. They combine the judgement of a very large market and are hard to improve on by a wide margin. A model that diverges sharply from a well-informed market is far more often broken than inspired, and blending is a cheap guard against one bad feature or a stale input producing a confidently absurd number.

The weighting depends on the tier, and that's deliberate. The model always keeps most of the weight. In top competitions, where prices are best informed, the market's share is at its largest. In lower tiers and less-covered competitions, prices are thinner and less examined, so the market's share shrinks, and that is exactly where a statistical model has something to add that the market hasn't already priced in.

How one prediction gets chosen over another

Every market on every fixture has a probability, which is far more than anyone wants to read. A separate decision layer chooses what is actually published, and it doesn't just rank by the largest difference from the market.

Candidates are ranked by a combined score: the probability gap, weighted by how reliable that market has been and by the model's confidence. Data quality lowers a candidate's confidence instead of silently discarding it, so a fixture where one side has a thin statistical history gets a weaker, clearly marked prediction. It doesn't disappear, and it doesn't appear at full confidence either. Where data is materially incomplete, only the most dependable markets are offered at all. Rotation risk before a cup or continental fixture lowers confidence instead of blocking the fixture. There is also a cap on how many predictions one match can produce, so a single well-covered fixture can't crowd out a whole matchday.

The result is fewer published predictions than the model could generate, which is the right direction for that trade-off.

What this pipeline cannot do

It can't tell you what will happen in one match. A calibrated 72% is a statement about the long run: outcomes like that one land roughly 72 times in 100, and the other 28 aren't errors. They are the distribution doing what it was measured to do.

It can't see a ninth-minute red card, a goalkeeping error or a manager sacked between the prediction and kick-off. The model isn't missing those features; they are irreducible noise, and they are why the honest ceiling on match-result accuracy sits in the mid-fifties, nowhere near the figures advertised across this category.

And it can't be validated by how any single weekend went. The only meaningful test is whether the stated probabilities match observed frequencies across thousands of matches, which is why the accuracy record is published bin by bin instead of as a headline percentage.