Performance

2026 Season Performance

Graded against final scores as games are played, updated every Sunday alongside the new predictions. ATS is whether the model picked the covering side of the Vegas home line.

Games graded

209

Winner accuracy

76.1%

Brier

0.146

Spread MAE

13.1

Spread RMSE

17.3

Spread sign accuracy

77.5%

ATS vs Vegas

55.4%

ATS games

204

By week

Week 1 (45 games)

73.3% win · 19.3 spread MAE · 46.5% ATS

Week 2 (49 games)

77.5% win · 11.9 spread MAE · 57.1% ATS

Week 3 (57 games)

82.5% win · 9.0 spread MAE · 63.6% ATS

Week 4 (58 games)

70.7% win · 13.4 spread MAE · 52.6% ATS

Training

Both models are scored on an honest holdout: train on seasons through 2023, early-stop on the next season, then measure 2025 — a year the booster never sees while trees or hyperparameters are chosen. The production models are then retrained on every season through 2025.

Winner

LightGBM classifier on 844 2025 games. Production fit uses 8,513 games, 325 features, 109 trees.

Accuracy

71.6%

AUC

0.759

Brier

0.194

Log-loss

0.570

Isotonic calibration on 2024 raised holdout Brier from 0.194 to 0.199, so the shipped model uses raw probabilities.

Spread

LightGBM regressor predicting home minus away, scored on the same 844 2025 games. MAE and RMSE are in points.

MAE

12.9

RMSE

16.3

Sign accuracy

71.7%