How our football predictions work

The data

Every prediction starts from match data covering hundreds of competitions: results, goals, shots, corners, cards, lineups, injuries, referee history, venue records and head-to-head history. We keep the full history rather than a rolling window, so a model can be tested against seasons it has never seen.

The model

Our current model, v5, has been live since September 2026. It estimates the strength of every team from past results and match data, and produces one consistent set of probabilities for every market. Earlier models worked differently: a set of analysers each scored one dimension of a fixture, such as recent form, home advantage, temporal patterns, referee tendencies, injury impact and expected goals, and their outputs were combined into probabilities. Calibration matters as much as the model itself. A 60 percent probability should be right about 60 percent of the time, and we measure whether it is.

What confidence means

Every prediction carries a confidence score. It states how far the evidence separates one outcome from the others. For our earlier models, across 144,972 graded predictions, high-confidence picks landed 62.3 percent of the time on a sample of 3,103, while low-confidence picks performed close to chance. We publish both figures.

How we grade ourselves

When a match finishes, every market we predicted is compared against the real result and recorded permanently. Nothing is fitted after the fact and nothing is quietly dropped. That is why the headline number here is lower than figures advertised elsewhere. It covers every fixture in every competition, including the ones we got wrong.

What this is not

This is a statistical model, not a guarantee. The earlier models' overall figure sits above the 43.8 percent you would get by always backing the home side, which is the bar any honest model should clear. Football remains uncertain and no model changes that. We would rather show the real record than a better-looking one.