How Our AI Football Predictions Work

Short answer: a machine-learning model trained on 79444 historical matches weighs team strength, form, expected goals and bookmaker odds and outputs a probability for every market. Probabilities are recorded before kick-off and scored against real results.

Data

Results and odds for 16 leagues since 2012 (football-data.co.uk)
Expected goals, xG (Understat)
Elo team ratings and rest days (ClubElo)
Fixtures, live scores and line-ups (ESPN)

The brain: three views, one probability

Market: the bookmakers' price closest to kick-off, margin removed, and how it moved since it opened.
Archive: the 250 most similar of 79444 past matches (market view, team strength, form, expected goals) and what happened in them.
Model: gradient-boosted trees (LightGBM) learn what form, xG, strength gap and rest add on top of the market.

The three are blended with weights learned only on earlier seasons the model had not seen; the "how was this built" card on each match shows every view. Predictions update with the latest price until kick-off, and the one the scorecard counts is recorded just before kick-off. Match result, both teams to score and over/under 1.5 / 2.5 / 3.5 are computed separately; likely scorelines come from each side's expected goals. For the European cups and internationals, probabilities are derived from market odds alone.

How accuracy is measured

The model is always trained on earlier seasons and tested on a season it has not seen, with an automated check against information leaking from the future. The key measure is calibration: picks rated 70% should land about 70% of the time. Since 28 August 2026 we have recorded 1487 matches before kick-off: the most likely pick landed 80% of the time (the model expected 80%), and 84% of the 645 picks rated 80%+ landed. Walk-forward over 2018-2026, 44,249 matches the system had never seen: the most likely pick landed 80.2% of the time, and 83.9% of the 20,404 picks rated 80%+ landed. See the scorecard and past predictions.

Limits

Football has a large share of luck; even the best model misses about half of all match-result predictions.
Our model is about as accurate as the betting market, not meaningfully better; we do not promise profit from betting.
18+. Betting loses money.