Are the models any good?
We chose each model's settings using the 2015–2018 seasons, then predicted every one of the 2,660 games of 2019–2025, using only information available before each game.
Leaderboard (lower is better)
| Model | Log-loss | Brier |
|---|---|---|
| Bookmakers' closing oddsbenchmark (only exists minutes before kick-off) | 1.0007 | 0.5990 |
| Market (ratings from odds) | 1.0153 | 0.6085 |
| Poisson | 1.0218 | 0.6129 |
| Dixon-Coles | 1.0221 | 0.6131 |
| Elo | 1.0228 | 0.6137 |
| Elo + fixed draw rate | 1.0284 | 0.6175 |
| Always the average (home/draw/away) | 1.0621 | 0.6410 |
Log-loss: how “surprised” the model was by what happened. Giving 1% to something that then happens is very costly.
Brier: mean squared gap between the predicted chances and what happened (0 = perfect).
Calibration
When a model says “30%”, it should happen about 30% of the time. Points near the dashed line = well-calibrated model.