NFL model performance

Walk-forward: each season is predicted by models trained only on earlier seasons. Settings were chosen on 2016–2020; the test seasons were scored once. Log loss and Brier score measure how good the probabilities are; lower is better.

Test log loss (model)

0.634

market 0.610

Winner picked (test)

63.9%

market favourite 66.3%

Test games

1,420

2021 – 2025

Live record

1 games

log loss 0.482

Test seasons 2021 – 2025 (1420 games)

Model	Log loss ↓	Brier ↓	Winner picked ↑
Always home team (base rate)	0.6896	0.2482	54.6%
Elo (logistic on rating gap)	0.6404	0.2248	63.1%
Margin model (ridge regression)	0.6345	0.2222	63.5%
Gradient boosting	0.6347	0.2222	63.5%
Ensemble (final model)	0.6341	0.2220	63.9%
Bookmaker moneyline (margin removed)	0.6103	0.2115	66.3%

Average miss on the final margin

10.13 pts model · 9.77 pts betting spread

Average miss on total points

10.51 pts model · 10.31 pts betting total

1420 games with a closing line.

The final model averages elo 10%, margin 75%, gbm 15%. The betting market is better than the model on every measure, which is normal for a model without injury, lineup or rest-decision information. Bold marks the best non-market value.

Calibration on test seasons (home-win probability)

Validation seasons 2016 – 2020

Model	Log loss ↓	Brier ↓	Winner picked ↑
Always home team (base rate)	0.6880	0.2474	55.4%
Elo (logistic on rating gap)	0.6287	0.2194	64.7%
Margin model (ridge regression)	0.6244	0.2177	65.0%
Gradient boosting	0.6267	0.2186	64.3%
Ensemble (final model)	0.6242	0.2175	65.0%
Bookmaker moneyline (margin removed)	0.6069	0.2097	66.6%

Used to choose every setting, so slightly optimistic.

Log loss by season

Season	Games	Elo	Margin	Boosting	Ensemble	Market
2016 validation	265	0.622	0.626	0.633	0.626	0.622
2017 validation	266	0.623	0.618	0.622	0.618	0.593
2018 validation	265	0.630	0.618	0.624	0.620	0.614
2019 validation	266	0.642	0.636	0.638	0.636	0.616
2020 validation	268	0.626	0.624	0.618	0.622	0.590
2021 test	284	0.664	0.651	0.639	0.649	0.623
2022 test	282	0.631	0.628	0.637	0.629	0.605
2023 test	285	0.660	0.651	0.658	0.652	0.627
2024 test	285	0.615	0.616	0.614	0.614	0.589
2025 test	284	0.633	0.627	0.626	0.626	0.607
2026 in progress	106	0.664	0.673	0.676	0.672	0.662

How the model works

Elo with a margin-of-victory multiplier (K = 20, home advantage 35 Elo points, 50% pulled back to average each off-season). A margin model (ridge regression) predicts the final point margin from Elo, recent point differential, points scored and allowed, home/road form, season record and rest and quarterback changes; the margin becomes a win probability through a normal distribution. A matching model predicts total points.

A gradient-boosted classifier uses the same inputs. The final probability averages the three with weights chosen on validation seasons; the projected score combines the predicted total and margin. Betting lines are never a model input. Model v20261029-e80118, trained Oct 29, 2026 on 6,606 games.