NFL: every call, graded
Winner, spread and total — locked before kickoff, graded after, with the backtest underneath. The parts that don’t work are on the page too. How the model works →
Locked before kickoff Only picks written before the game started count here. Anything scored after the fact is in the backtest section below, kept separate on purpose.
Backtest Walk-forward: every pick made from ratings built only on games before it, then graded against the closing line. Reported straight, including the parts that do not work.
| Season | Games | Winners | ATS | O/U |
|---|---|---|---|---|
| 2020 | 269 | 174/269 · 65% | 129-140 · 48% | 120-144-5 · 45% |
| 2021 | 285 | 170/285 · 60% | 136-145-4 · 48% | 144-138-3 · 51% |
| 2022 | 284 | 177/284 · 62% | 140-134-10 · 51% | 130-151-3 · 46% |
| 2023 | 285 | 177/285 · 62% | 122-149-14 · 45% | 140-141-4 · 50% |
| 2024 | 285 | 193/285 · 68% | 134-147-4 · 48% | 141-141-3 · 50% |
| 2025 | 285 | 188/285 · 66% | 145-139-1 · 51% | 153-132 · 54% |
What it says: the model picks winners at about 64% with a quarterback adjustment (the starter’s EPA value against the team’s usual, walk-forward), and its probabilities are calibrated — but the closing moneyline is better still (≈66.5%), and against closing spreads and totals the model sits at 48–49%: a coin flip, not an edge. So winner calls are a proven read; spread and total calls are leans graded in public, now against the opener as well as the close, with closing-line value tracked from the first lock. Also tested and rejected: updating the rating on EPA margin instead of points (worse at every blend), and pricing injuries beyond the quarterback from the report plus snap shares (a small positive, a tenth of the bar it has to clear). Neither is priced. The remaining gap to the market is what the market knows that public data does not.
Grades post the morning after each game day. Informational only. 21+, 1-800-GAMBLER.