Research
Players, matchups and the numbers behind them.
Model lab
Experimental · not liveWe’re testing game outcomes before publishing forecasts. These are historical replays, not predictions issued before those games.
Player model testing
2025 replay · average error per eligible game. Lower is better. The comparison is a player’s last-five-game average.
| Stat | Model error | Recent average |
|---|---|---|
| Passing yards323 player-games | 65.8 | 67.5 |
| Rushing yards860 player-games | 22.7 | 22.9 |
| Receiving yards2,275 player-games | 20.5 | 20.7 |
| Receptions2,275 player-games | 1.5 | 1.5 |
Player coverage & limitations
Only recorded participants with at least five prior same-season games and an established role are evaluated. This excludes many low-volume players and missing participation records; it is not full-roster accuracy.
Passing yards: 323 of 18,522 recorded player-game rows scored. 4,461 unknown participation; 6,947 short history; 6,791 below prior-workload requirement; 0 missing stat or workload. The target 80% range covered 81.7% of 323 scored results.
Rushing yards: 860 of 18,522 recorded player-game rows scored. 4,461 unknown participation; 6,947 short history; 6,254 below prior-workload requirement; 0 missing stat or workload. The target 80% range covered 80.2% of 860 scored results.
Receiving yards: 2,275 of 18,522 recorded player-game rows scored. 4,461 unknown participation; 6,947 short history; 4,839 below prior-workload requirement; 0 missing stat or workload. The target 80% range covered 80.5% of 2,275 scored results.
Receptions: 2,275 of 18,522 recorded player-game rows scored. 4,461 unknown participation; 6,947 short history; 4,839 below prior-workload requirement; 0 missing stat or workload. The target 80% range covered 80.9% of 2,275 scored results.
No injuries, depth-chart changes, opponent adjustment or prices are included. Player-games are correlated, not independent trials. Receiving volume uses catches rather than targets. The 2025 results are now inspected historical evidence, not a future holdout.
Player model: player-volume-v0.1 · Source SHA-256: 7866246f349c626137c15352665c42e9d841b11332c10304d1eec52c20dc1394
How the game model compares
Probability error · lower is better
Shared scale: 0 (perfect) to 1 (worst). Probability error is not a measure of returns.
What is probability error?
The Brier score measures how far a predicted chance was from the actual outcome. Confident wrong forecasts receive larger penalties. The baseline uses the home-win rate from earlier games, not market prices.
Historical replay results
| Season | Correct winners | Coverage |
|---|---|---|
| 2024 | 139 of 208(66.8%) | 208 of 27264 excluded |
| 2025 | 125 of 208(60.1%) | 208 of 27264 excluded |
2025 performance was weaker than 2024. Injuries, starting-QB changes, neutral venues and market prices are not included. This is a baseline, not a validated production model.
Point estimates & uncertainty
In the 2025 replay, average error was 10.7 points for the score margin and 10.6 for the total. Ranges targeting 80% coverage reached only 76.4% and 77.9%, respectively. Uncertainty needs further calibration.
Method & data provenance
2022–2023 warmup, 2024 validation and 2025 later-season test. Each week is forecast before that week’s results enter the model. Revised historical data is not a verified archive of what was known before kickoff.
Report: · nfl-game-benchmark-v0.1
Source SHA-256: d8e11b083014020c763b2133a1de8a785cdaf1ff3ddf95bef809ad98997f8cb4