This season's scoreboard
12 seasons of testing
We replayed 2014/15 to 2025/26 match by match. Each forecast used only results from before that match, as if we had made it at the time, and the model settings were chosen on earlier seasons. Accuracy is the Ranked Probability Score (RPS): 0 would be perfect and lower is better. "Base rates" ignores the teams and always predicts the league's usual share of home wins, draws and away wins.
Both models clearly beat guessing, and the goals model beats Elo. The bookmakers' closing odds are better still, in every season: they include team news and the views of professional bettors, while our models only know past scores.
How often each picked the right result
Hit rate: the share of matches where the outcome rated most likely actually happened. "Always home" picks a home win every time. Pure guessing would get about 33%.
Hit rate is easy to understand but a blunt measure: a 51% and a 90% favourite count the same when they win. RPS rewards getting the probabilities right, which is why we use it to choose and compare models.
Corners, cards and shots
We also predict each match's corners, yellow cards, shots and shots on target, tested the same way. "Average miss" is how far our expected total was from the real one; "league average" is what you'd get by always guessing last season's average. The last column is how often we called the over / under line correctly.
| Stat | Our average miss | League average | Over / under right |
|---|
Both models' settings were chosen on the 2008/09 to 2013/14 seasons, before the test period, and every forecast only uses results from before its match. The code is on GitHub.