What a preseason model knows. And what it misses.

The useful question is not simply who will win. It is what evidence supports the estimate—and what could make it wrong.

Our first football model learns team strength from final scores. Each team receives a coefficient; home field adds another effect. Ridge regularization keeps a short run of unusual scores from moving a coefficient too far. More recent seasons receive greater weight. A second regression estimates the combined score.

A real test requires a boundary

To evaluate the approach, we trained on 2022, 2023, 2024 and predicted 2025 without feeding any results from that season back into the coefficients. Across 784 eligible games, average margin error was 14.24 points. The simple constant-home-margin baseline missed by 15.99 points on the same games.

That improvement is useful evidence for the baseline. It does not prove accuracy for every team, and it says nothing about a sportsbook advantage. Programs with new quarterbacks or major roster turnover can look very different from their historical identity.

Read the range before the score

A predicted three-point margin can coexist with a wide range of plausible outcomes. Our displayed 80% range uses the 80th percentile of absolute forecast errors on 2024 games. A logistic curve fitted on that same earlier window maps projected margins to home win probabilities. Neither uses the 2025 test results for calibration. The frozen range covered 80.6% of those test margins. A probability like 60% remains an estimate, not a guarantee. The model notebook shows the reliability groups and downloadable test evidence.

The next research steps are to incorporate dated roster and efficiency features and evaluate on rolling time splits. Use the score model to organize film study and identify questions worth investigating.

Data attribution: SportsDataverse, CC BY 4.0. See the source receipts and model notebook.