Model performance
Metrics are from a chronological walk-forward backtest: each game is predicted before that game's result updates the ratings.
Games
6,560
Brier
0.2383
Log loss
0.6692
Accuracy
58.6%
Out-of-sample backtest
Scored 6,560 games. Model Brier 0.2383 vs coin-flip baseline 0.25 (4.7% better than chance). Accuracy 58.6%.
Selected configuration
Chosen by lowest Brier score from a small grid: K=12, home advantage=35 Elo points, season carryover=0.65.
| K | Home Elo | Carryover | Brier | Log loss | Accuracy |
|---|---|---|---|---|---|
| 12 | 35 | 0.65 | 0.2383 | 0.6692 | 58.6% |
| 16 | 45 | 0.7 | 0.2386 | 0.6698 | 58.6% |
| 20 | 55 | 0.75 | 0.2393 | 0.6714 | 58.5% |
| 24 | 65 | 0.8 | 0.2405 | 0.6739 | 58.5% |
| 28 | 75 | 0.85 | 0.2419 | 0.6772 | 58.4% |