How accurate is the model?

A forecast is only useful if you know how good it is. We test every version of the model on matches it has not seen, and compare it with the closing betting market — the hardest benchmark there is.

The last week of matches

1 November to 8 November: the model’s favourite won 31 of 54 matches, and its log loss was 0.932. The forecasts were made before the week began. One week is a small sample: it swings a lot either way.

Where it was most wrong:

The test

For every match since August 2023 in the Premier League, La Liga, Serie A, Bundesliga and Ligue 1, the model is refitted using only the matches played before that day, then asked for win, draw and loss probabilities. That is 5,703 matches. We score the forecasts with log loss: lower is better, and a forecast that is confidently wrong is punished hard.

Win–draw–win log loss, 2023/24 to November 2026 (goals model; adding xG leaves these figures unchanged)
LeagueMatchbornBetting marketGap
Premier League0.9850.964+0.021
La Liga0.9720.950+0.023
Serie A0.9820.960+0.022
Bundesliga0.9950.965+0.029
Ligue 11.0020.979+0.023
All five leagues0.9860.963+0.023

Guessing from league averages alone scores 1.075, so the model captures most of what can be known — but the market still knows more. Closing odds include late team news and the money of professional bettors.

What the model uses

Each team gets an attack and a defence rating, fitted to goals and to expected goals (xG) in recent matches, with older matches counting less. Home advantage is estimated for each league. Using xG alongside goals made our over/under 2.5 goals forecasts measurably better without hurting win–draw–win.

We also tested starting line-ups, injuries, rest days between matches, the referee, and shot counts. None of them improved the forecasts on unseen matches, so none of them are in the model. When one does, we will add it and update this page.