A time-decayed Poisson model in the Dixon-Coles family. Each team has an attack and a defence rating, each league a home advantage and a scoring baseline. Ratings are fitted on the last three seasons with exponential decay (half-life about 107 days), refitted on every refresh. Expected goals for both sides give the scoreline distribution, with the standard correction for low-scoring draws. Only results, no injuries, lineups, or odds, go into the model.
A forecast is stored the first time a fixture appears in the schedule feed, together with the odds at that moment. It is never recomputed. After the match the result and the closing odds are attached and the row moves to the track record. Everything on this site can be re-derived from the public football-data.co.uk files.
We ran this exact model over 19,763 matches of the same five leagues from 2015 to 2026, betting only on paper against the recorded closing odds. Results:
| Log-loss | Value |
|---|---|
| Model | 0.997 |
| Pinnacle closing line | 0.965 |
| Coin flip | 1.099 |
| Flat 1-unit bets wherever model price > book price | ROI |
|---|---|
| vs Pinnacle closing (2% margin) | −5.9% |
| vs Bet365 closing (5% margin) | −9.6% |
| vs best price across books | −1.9% |
| vs a 6.5%-margin book (typical licensed RU operator) | −9.8% |
Mixing the model into the market probabilities at any weight made the forecast worse. In plain words: a model built on public results carries no information the closing line does not already have. The market is the better forecaster, and the bookmaker's margin is the whole cost.
Because almost nobody who sells football tips publishes a frozen record, and the few who do stop when it turns red. This page keeps going either way. Use it to see what a sound public model can and cannot do, and to judge anyone who claims more.
Nothing here is a recommendation to bet. The historical result of following this model is a loss roughly equal to the bookmaker's margin.