More information.
A better forecast?
Test prior-game EPA and yards per play against a score-only correction on the same 784 historical games. Correction models train on 2023; probabilities are calibrated in 2024; this comparison scores the 2025 season.
No clear gain from the extra features.
Score + efficiency: 13.34 points of margin error. Score-only correction: 13.36. The difference is -0.018 points; the approximate 95% week-block range is -0.182 to 0.155.
This is an exploratory, retrospective comparison. Settings were recorded after the historical season and after the original weekly benchmark was known. Source corrections and publication delays are not reconstructed. Current forecasts and prospective ledger records are unchanged.
Fit both residual models and their feature scaling. Freeze those coefficients for later seasons.
Fit a separate probability curve and nominal 80% margin range for each model.
Compare the same matchups. Earlier results update later weekly score fits and feature inputs.
Where did the predictions change?
Margin change compares score + efficiency with the score-only correction. All tables use the same selected games. Monthly bars keep program and change filters and let you select a month; the full-test uncertainty range above is fixed.
Loading paired historical predictions…
More fields need stronger evidence.
The score-only correction and efficiency correction use ridge regression with penalty 100, a free intercept, and means and standard deviations learned only from 2023. Both predict the original weekly model’s margin error. The efficiency model adds four home-minus-away gaps: offensive EPA, EPA allowed, offensive yards and yards allowed, all per play.
Each weekly input pools paired, scored FBS-versus-FBS games from the current and preceding season whose kickoff is before Sunday 00:00 UTC, 24 hours before the Monday bucket. Prior-season totals receive weight 0.5. Each team rate is shrunk toward that cutoff’s pooled league rate with 300 equivalent plays. A team with no usable history receives the league rate. 0 of the test games needed that fallback.
The feature archive contains 3,124 paired advanced games and 65 dated states. Missing or incomplete advanced pairs do not enter those rate pools. The score fits still use their full eligible score samples. The 2025 evaluation retains the original weekly benchmark’s 784 games and its 24 exclusions for teams outside the frozen field.
Single-season residual training, unadjusted efficiency inputs, repeated teams, later source corrections and an already-observed historical test season limit this result. The 5,000-resample week-block interval retains whole weeks and recomputes game-weighted error; teams recur across weeks, so it does not remove all dependence. No live betting advantage or new production model is claimed.
Inspect fitted correction coefficients
| Model term | Score-only correction | Score + efficiency |
|---|---|---|
| Intercept | -0.0694 | -0.0694 |
| Weekly score marginPoints per training standard deviation | 4.1878 | 0.8704 |
| Offensive EPA / play gapPoints per training standard deviation | Not included | 1.6536 |
| EPA / play allowed gapPoints per training standard deviation | Not included | 1.2581 |
| Offensive yards / play gapPoints per training standard deviation | Not included | 2.0048 |
| Yards / play allowed gapPoints per training standard deviation | Not included | -2.8341 |
Every transformation has a trail.
Experiment 71db695277c78cf286e7d50ba5aebec74d4e2d644e839cdc004b6997fce49d81
Design recorded 2026-09-05T09:23:44.181193+00:00. Generated 2026-09-12T13:31:31.391550Z. Both clocks are after the historical evaluation season.
Advanced source receipts
SportsDataverse · team_advanced / 2022 ↗
Retrieved 2026-09-09T00:21:05.575694Z
SHA-256 9c17e442ba3f25f1ec25546a0570afbd280f250b133ea5f7f9cecb3de26d9511
SportsDataverse · team_advanced / 2023 ↗
Retrieved 2026-09-09T00:21:08.573527Z
SHA-256 7b97a24ba7f4aa7c04eed5ac9f72359255a2288ddf9094ee668600dbba9f1ab6
SportsDataverse · team_advanced / 2024 ↗
Retrieved 2026-09-09T00:21:11.785691Z
SHA-256 26728558bf04a0d2b8cd0ad1f6c79c43aaea42d6b03f0d80ba797ffdc6605828
SportsDataverse · team_advanced / 2025 ↗
Retrieved 2026-09-09T00:21:22.052664Z
SHA-256 63d100ace111c00e0171a958974d2f6f589c0cfcef4834a113232b0a734be77f
Publisher-stated CC BY 4.0. Silvermine normalizes the source fields, builds lagged features and fits independent correction models. Raw team-game records are also available in the efficiency desk; schedule and score-model evidence are retained in the original weekly experiment.