Week of 2026-09-07 · performance + feature importances for every live model.
⚾ MLB: 61-57 (51.7%) · ⚽ Soccer: 52-43 (54.7%) · 🎓 College Football: 73-9 (89.0%) · 🏈 NFL: off season (back 2026-09) · 🏀 NBA: off season (back 2026-10)
| Window | Record | Accuracy | Log loss | Brier |
|---|---|---|---|---|
| Season to date | 61/118 | 51.7% | 0.6967 | 0.2516 |
| Last 7 graded days | 44/88 | 50.0% | 0.6923 | — |
Edge vs. always-pick-home: Δlog-loss +0.0047 ± 0.0129 per game (positive = model beats the baseline). Graded through 2026-08-15. The MLB model is betting-blind Elo + home advantage + the starting-pitcher adjustment — it has no ML feature vector, so there are no SHAP/permutation importances to report; the record vs. baseline above is the model check.
| Window / league | Graded | Pick accuracy | Log loss |
|---|---|---|---|
| Last 7 days | 5 | 100.0% | 0.5615 |
| Last 30 days | 95 | 54.7% | 0.9548 |
| Season to date | 95 | 54.7% | 0.9548 |
| bundesliga | 9 | 55.6% | 0.8712 |
| championship | 23 | 43.5% | 1.0805 |
| epl | 19 | 52.6% | 1.0098 |
| la_liga | 19 | 57.9% | 0.8519 |
| ligue_1 | 9 | 44.4% | 1.1113 |
| serie_a | 16 | 75.0% | 0.7903 |
| Feature | Permutation Δlog-loss | Mean |SHAP| | Home-win coef (association) |
|---|---|---|---|
| elo_gap | +0.0614 ± 0.0015 | 0.1591 | +0.003 (raises home-win odds) |
| value_diff_z | +0.0109 ± 0.0002 | 0.0559 | +0.070 (raises home-win odds) |
| xg_net_diff | +0.0029 ± 0.0006 | 0.0175 | +0.132 (raises home-win odds) |
| spend_diff_z | +0.0000 ± 0.0000 | 0.0000 | +0.042 (raises home-win odds) |
| net_diff_z | +0.0000 ± 0.0000 | 0.0000 | +0.003 (raises home-win odds) |
| wage_diff_z | +0.0000 ± 0.0000 | 0.0000 | +0.000 (lowers home-win odds) |
Multinomial logistic refit on matches before 2024-04-06 and evaluated on the 8855 most recent (15%) held-out matches. Permutation = how much held-out log loss degrades when the feature is shuffled (5 repeats); SHAP is the exact linear attribution |coef · (x − x̄)| averaged over rows and classes. The coefficient column is the H-class weight on the z-scored feature — its sign is the direction of association.
| Model | Games | Accuracy | Log loss | Brier | Cal slope | ECE |
|---|---|---|---|---|---|---|
| production (qb+talent Elo) | 3018 | 66.5% | 0.6140 | 0.2127 | 0.978 | 0.013 |
| previous (base Elo) | 3018 | 66.5% | 0.6140 | 0.2128 | 0.981 | 0.015 |
| market (closing line) | 3018 | 66.1% | 0.6121 | 0.2122 | 1.056 | 0.018 |
| qb+talent Elo alone | 3018 | 65.2% | 0.6319 | 0.2202 | 0.806 | 0.034 |
| always base rate | 3018 | 55.0% | 0.6882 | 0.2475 | 0.038 | 0.000 |
Walk-forward out-of-sample scorecard (NFL/model/v2). The market row is the closing line — the bar the model is judged against.
| # | Feature | Avg (perm+SHAP) rank | Norm. score |
|---|---|---|---|
| 1 | spread_line | 1.6 | 0.856 |
| 2 | market_home_prob | 2.3 | 0.847 |
| 3 | elo_diff | 3.9 | 0.359 |
| 4 | roll_margin_diff | 6.0 | 0.212 |
| 5 | elo_vs_spread | 10.3 | 0.227 |
| 6 | elo_spread | 12.6 | 0.279 |
| 7 | elo_home_prob | 13.4 | 0.240 |
| 8 | away_roll_pf | 13.8 | 0.110 |
| 9 | away_ppg_diff_std | 15.9 | 0.103 |
| 10 | total_line | 17.4 | 0.068 |
| 11 | home_roll_total_vs_line | 17.6 | 0.065 |
| 12 | travel_miles | 18.0 | 0.069 |
| 13 | home_roll_margin | 18.2 | 0.119 |
| 14 | away_roll_ats | 19.0 | 0.069 |
| 15 | away_winpct_std | 19.6 | 0.054 |
| Feature | Permutation Δlog-loss | Mean |SHAP| |
|---|---|---|
| spread_line | +0.0260 ± 0.0048 | 0.3091 |
| market_home_prob | +0.0253 ± 0.0042 | 0.2860 |
| roll_margin_diff | +0.0023 ± 0.0018 | 0.0634 |
| elo_diff | +0.0021 ± 0.0016 | 0.1043 |
| away_roll_total_vs_line | +0.0020 ± 0.0018 | 0.0621 |
| home_roll_total_vs_line | +0.0019 ± 0.0014 | 0.0472 |
| total_line | +0.0019 ± 0.0018 | 0.0622 |
| travel_miles | +0.0013 ± 0.0021 | 0.0606 |
| elo_home_prob | +0.0007 ± 0.0008 | 0.0461 |
| home_rest | +0.0004 ± 0.0003 | 0.0052 |
Fit 2002–2023 with production recency weighting, measured on a held-out 2024–25 window the model never saw. Regenerated offline by NFL/model/v2/feature_importance.py; beeswarm plots showing each feature's direction of association live in NFL/model/v2/artifacts/importance/.
Can Tre Beat Vegas · whosyurgoat.app/vegas