Can a statistical model actually beat the NFL spread?
No — this app's own margin model went 50.4% against the spread on games it wasn't trained on, below the 52.4% needed just to break even at standard odds. Published because it's the honest answer, not because it's a good one.
See the model's current spread read in Team Bets
The evidence
This app's margin_v1 model vs. the closing spread, out-of-sample games only
| Measure | Value |
|---|---|
| Model ATS win rate | 50.4% |
| Break-even rate at -110 | 52.4% |
| Games graded (2025-2026 only) | 406 |
| Pushes | 7 |
How this was measured
This app's own regression model (predict_game_margin, trained through the 2024 season on EPA/explosiveness/pressure/red-zone/turnover/rest differentials) run against every 2025-2026 game with a known result and a closing spread — strictly out-of-sample, since grading a model on the seasons used to fit it would inflate its record. The model's picked side (home or away, whichever it favors relative to the spread) is compared to the actual result to decide a hit or a miss.
Sample size: 406 games (out-of-sample, 2025-2026). Last verified 2026-09-27.
What this doesn't tell you
A sub-break-even ATS record from one model on one style of feature set doesn't prove NFL spreads are unbeatable in general — it shows this specific approach isn't one of the ways to do it. The honest reading is that the market is hard to beat with box-score-level team efficiency stats alone, not that it's provably impossible with a different approach.
Related questions
This is a place to look, not a bet to make — every figure here describes the past, not a specific upcoming game.