Accuracy
One test matters: when the model says something is 60% likely, does it happen about six times in ten? Every published number is graded on the final whistle and nothing is edited afterwards.
Model saidPredictionsActually happenedDifference
90–100% likely6180%−12.5
80–90% likely37980%−3.6
70–80% likely100475%+0.2
60–70% likely89763%−2.1
50–60% likely96159%+4.1
40–50% likely120045%+0.6
30–40% likely193138%+4.0
40704
Predictions graded, each published before kick-off.
±2.6
Average gap between what the model said and what happened, across 6,433 predictions at 30% or likelier, weighted by band size.
0
Numbers changed, hidden or backdated after a result.