We are like Greeks. We are like Romans. / Centaurs and monkeys just cluster 'round us "Nemesis" - Shriekback

Accuracy all seasons, every FINAL game scored on leakage-free pre-game predictions

Metric Baseline Plus Book Closing Line
Spread MAE (n=30,426) 8.96 8.83
Winner % 71.7% 72.8%
Total MAE (n=30,426) 13.07 13.02
Log loss 0.5221 0.5251
Brier 0.1740 0.1771

Head-to-head on identical games: model spread MAE 8.96 vs book 8.78 over 30,264 games; taking the model against the closing spread hits 53.2% (52.4% is breakeven at −110). Full breakdowns on Compare Models. Not betting advice.

What it knows — the 41 inputs per game

Form & schedule (15)

home_win_pct_l5, home_avg_margin_l5, home_avg_total_l5, home_margin_stddev_l5, away_win_pct_l5, away_avg_margin_l5, away_avg_total_l5, away_margin_stddev_l5, home_games_played, away_games_played, home_days_rest, away_days_rest, season_week, is_neutral_site, is_conference_game

Ratings (14)

massey_beta_home, massey_beta_away, massey_beta_diff, massey_gamma_home, massey_gamma_away, massey_gamma_sum, bt_theta_home, bt_theta_away, bt_logodds, bt_theta_weighted_home, bt_theta_weighted_away, bt_logodds_weighted, home_rpi, away_rpi

Box-score profile (12)

home_pace, away_pace, home_off_eff, away_off_eff, home_def_eff, away_def_eff, home_efg_pct, away_efg_pct, home_opp_efg_pct, away_opp_efg_pct, home_tov_rate, away_tov_rate

Training

Trained
2026-07-10
Train seasons
Held-out test
Test spread RMSE
11.79
Test Brier
0.1882

Walk-forward (each season held out in turn)

SeasonSpread RMSEBrier
2022 11.79 0.1986
2023 11.60 0.1954
2024 11.64 0.1903
2025 11.70 0.1857
2026 11.79 0.1877

Calibration when it says 70%, does the home team win 70% of the time?