How the predictions work

In one paragraph

For each team the model estimates how strong its attack and defence are from its results, weighting recent matches more than old ones. For a fixture it combines the two sides' strengths with home advantage into an expected number of goals for each, and from those it works out the probability of every scoreline — and therefore of every market: the win, the draw, over 2.5, both teams to score, a handicap, and so on. Those probabilities are then calibrated against how often such predictions have actually come true, and the calibrated figure is what you see.

The models

• Football — a Dixon–Coles model: Poisson goal rates per side with a correction for low-scoring draws, time-decayed so a result from last season counts for less than one from last week, and regularised so a team with few matches is pulled toward the average rather than trusted blindly.
• Basketball, ice hockey, baseball — a strength-rating model on the score margin, with the sport's own scoring distribution, so a 105–98 is read the way basketball reads it.
• Tennis — a serve-hold model: each player's chance of holding serve, walked through the scoring rules game by game to the probability of winning the match, sets and games.

Thin leagues

Early in a season a league has too few results to fit on. Below twenty matches the model uses each team's last twenty matches from any season and blends in the head-to-head record, so it is never fitted on three games and called certain.

Calibration and the record

A model that says 60% should be right about 60% of the time. The calibration page checks exactly that, band by band, against every sealed prediction. The track record compares the model to a naive base rate league by league; where it fails to beat the base rate, that league is marked and its picks are withheld from the radar. Every prediction carries the time it was generated.

What the numbers mean

A probability is a frequency, not a forecast of one match. "Home 70%" means that of many fixtures the model sees this way, about seven in ten end in a home win — and three do not. Fair odds are simply 1 ÷ probability; a bookmaker's price above the fair odds is what the site calls an edge, and an edge is a long-run statement that can and will lose on any given day.

Where it stops

• It does not know about injuries, suspensions, weather, motivation or a manager's plan unless they show up in results. Where verified line-ups or events are available they are shown, not guessed.
• It is fitted on results and shots; it cannot see a referee's mood.
• It is silent where it should be: markets and leagues it cannot price honestly are left out rather than filled with a guess.

The full technical description, including the parameters in use today, is on the Model page; the data behind it on Data sources.