ProbaPredict.

How our football model works

ProbaPredict ratings is a statistical goals model. It does not watch matches, read team news or have opinions. It estimates how many goals each side is likely to score, then turns that into probabilities.

What data goes in?

Final scores from the last two seasons in each of our 10 competitions, supplied by football-data.org. Fixtures for the next 7 days are refreshed every morning at 06:00 UTC, and results are collected every two hours.

How are teams rated?

When is a prediction published?

Only when both teams have at least 5 finished matches in the model's training data. Below that, a team's rating is mostly the average prior, so we publish nothing rather than a placeholder number. Those match pages stay out of search results until a prediction exists.

How are international matches handled?

National teams have their own separate model, because clubs and nations never play each other. Results come from the open international results dataset maintained by Mart Jürisoo (CC0), refreshed at 05:30 UTC and every six hours. The model uses the last six years of men's senior internationals, with recent matches counting more (a result from two years ago carries about half the weight). Friendlies are used for training only, at half weight, because competitive internationals are sparse; we never publish or grade a friendly. When a match is played at a neutral venue, no home advantage is applied. International model versions start with "intl-". The source dataset carries dates but not kick-off times, so international match pages show the date only.

How are extra time and penalties graded?

Every pick is graded on the 90-minute result. The international dataset records scores after extra time, so any match that went to a penalty shootout is not graded at all. For World Cup and European Championship matches we use the 90-minute score from football-data.org where it is available.

How do ratings become probabilities?

For each fixture the model calculates expected goals for both teams, then builds a grid of every scoreline from 0-0 to 10-10 using the Poisson distribution. A small Dixon-Coles correction adjusts the four lowest scorelines (0-0, 1-0, 0-1, 1-1), which plain Poisson tends to misprice. Adding up the right cells gives home, draw and away chances, Over/Under 1.5, 2.5 and 3.5 goals, and the five most likely scores.

How are picks graded?

The 1X2 pick is the outcome with the highest probability. The Over/Under pick is whichever side is above 50%. After full time each pick is marked correct or incorrect, and the record is public on the accuracy page. Losing picks are never removed. We also compare our probabilities with bookmaker closing prices from football-data.co.uk on the same matches.

Known weaknesses

Changelog