Backtest performance
The model is evaluated walk-forward: each completed game is predicted from ratings available before it, then ratings update chronologically. Headline numbers below are the holdout (games on/after ); the full-history figures are shown alongside.
Holdout games
74
Log loss
0.639
Brier
0.225
Accuracy
63.5%
All games
178
Log loss (all)
0.668
Brier (all)
0.238
Accuracy (all)
57.9%
Tuned parameters
Grid search selected homeElo=30, K=32, temperature=0.9 by holdout log loss with Brier tiebreak.
Generated 2026-06-09T09:30:44.185Z from ESPN Core API data.