Skip to content

Model Performance

Frozen walk-forward backtest, 20162025

Every prediction below was made walking forward through each season with only the data available at the time, then frozen. The market benchmark is the closing spread: the strongest public forecast of a game's margin. Beating it consistently is rare, and the model is measured against it, not against a naive baseline. The published model line blends the pure model with the market at a capped weight, so the two are correlated by construction; the pure model's own error runs about half a point higher.

Games graded2761
Model MAE9.99points
Market MAE9.78points
Gap to market+0.21points (lower is better)
Model closer45.6%
Backtest accuracy by season
SeasonGamesModel MAEMarket MAEGapModel closerBias
20162679.039.04-0.0148.7%-1.03
201726710.2810.08+0.2047.2%-0.53
201826710.179.87+0.3047.9%+0.10
201926710.4210.18+0.2443.1%+1.68
20202699.929.79+0.1346.5%+0.94
202128511.1410.67+0.4736.1%-0.16
20222849.018.78+0.2247.2%-0.41
202328510.389.98+0.3942.1%-1.03
20242859.859.70+0.1547.0%-0.51
20252859.669.67-0.0150.5%-0.53
All seasons27619.999.78+0.2145.6%-0.16

Bias is the mean signed error of the model's home margin: positive means the model leans toward home teams. Seasons 2016–2021 tuned the model's parameters; 2022 onward was never touched by selection and is the honest read on accuracy.