Model Backtest

The Crystal Ball

How well does the WFS predictor actually predict? Every game since 1980, scored against our pre-game forecast.

54.3%
Predicted W/D/L matched actual
Outright correct
58.4%
Higher-rated team avoided defeat
Favourite correct
0.167
Lower = better calibrated
Brier score
38,776
Every historical prediction
Backtested games

Model Calibration

When the model predicts a 60% chance, does the favourite really win 60% of the time? Perfect calibration follows the diagonal. The shaded band is the Wilson 95% interval.

Accuracy by Decade

How model accuracy has shifted across eras. More games per team in the modern era means stronger ratings and sharper predictions.

Accuracy by Rating Gap

Confidence scales with the rating delta. Narrow gaps are coin-flips; wide gaps are near-certain calls.

Accuracy by Tournament Type

1World
60.6%
2Continental
58.4%
3Major
52.9%
4Minor
52.0%
5Friendly
49.2%
6Regional
48.4%

The Biggest Upsets

Games where the model gave the eventual winner the lowest pre-game probability. The percentage badge is what the model thought of their chances.

0.1%
0 - 3
Pacific Games 22 Aug 1983
0.1%
1 - 0
Friendly 1 Apr 2000
0.2%
2 - 0
Friendly 6 Jun 1982
0.2%
2 - 3
AFC Asian Cup 21 Mar 2003
0.2%
0 - 6
FIFA World Cup 21 Mar 2000
0.2%
2 - 1
CFU Caribbean Cup 6 May 1990
0.2%
1 - 0
Friendly 5 Oct 1981
0.3%
0 - 6
FIFA World Cup 28 May 2001
0.3%
4 - 0
Friendly 21 Jan 2022
0.3%
0 - 6
FIFA World Cup 15 Feb 2001
0.3%
3 - 0
Friendly 14 May 1980
0.3%
0 - 3
Friendly 3 Aug 1998