Model record

Predictions frozen at kick-off and graded once the match finishes. Kept separate from the backtest.

How much less error the model makes than always guessing the market's average rate (Brier skill score). Above 0 beats that guess, below 0 does worse. A few percent is a real gain for football.

When the model gives a probability, does it come true that often? Close numbers mean honest probabilities. Each tile checks the model's surest 10% of calls; the calibration chart shows every band.