Model accountability
Performance
Only published, non-demo 1X2 predictions made before kickoff and matched to completed results are counted.
Performance numbers appear once published real-data predictions have final results. Demo matches do not enter this calculation.
Definitions
How these scores are calculated
The multiclass Brier score averages the sum of squared errors across home win, draw and away win probabilities for each match. Mean log loss averages the negative natural log of the probability assigned to the outcome that occurred. Lower scores are better for both.
The sample includes the latest eligible published prediction for each fixture, published no later than kickoff. It excludes demo fixtures, incomplete results and unpublished drafts. The baseline has not been independently calibrated; see methodology.