NFL Model Detail
Debut season 2026. We have no NFL track record yet — and we're not going to imply otherwise. What we can point to is the methodology and the receipts from MLB, run the same way in public since March.
What the Model Predicts
Point margin, converted to a probability against any given line using the model's out-of-sample error distribution. That means one continuous prediction is priced against every spread, moneyline, and total offered rather than being trained to answer one specific number.
The Core Inputs
The model reads no odds. It never sees the market it is being asked to beat.
EPA per play
Expected points added per snap — a measure of how much each play changed a team's expected scoring, given down, distance, and field position. It's context-aware in a way that yards and points are not: a 4-yard gain on 3rd-and-3 is a conversion; the same gain on 3rd-and-8 is a punt. We carry rolling offense, defense, and special-teams EPA plus opponent-blended versions of each.
Quarterback continuity
A career-to-date EPA-per-dropback rating for each starter, shrunk toward league average by sample size so a hot handful of games doesn't masquerade as elite play. When the starter changes mid-week (injury, benching, trade), the rating swaps to the incoming QB before any pick reflects the matchup.
Snap-weighted injury depletion
Position-group availability weighted by prior-year snap share. An injured fifth receiver and an injured starting left tackle are not treated alike. The input becomes meaningful once the week's practice reports are filed — any pick published before Friday is provisional in exactly this way. We measured this precisely: across 1,342 games, lines move toward the healthier team 60% of the time between Monday and kickoff; a quarterback ruled out is worth roughly a full point of line movement on its own.
Turnover margin, regressed
Raw turnover margin is shrunk heavily toward the league mean — fumble recovery is close to a coin flip year over year. Only the persistent component (interception rate) is allowed much weight.
Special teams
Included because our own testing found special teams carry roughly 14% of the model's total feature importance — a component many public models omit entirely. Kicking, punting, return, and coverage grades all enter as opponent-blended rates.
Method
An ensemble of gradient-boosted trees, a regularized linear model, and a random forest, predicting point margin directly. The continuous margin prediction is converted into a probability against any given line using the model's out-of-sample error distribution.
Data: play-by-play from every NFL game since 2019 — roughly 342,000 plays — plus weekly injury reports, snap counts, and multi-book closing lines.
Week 1–4 Caveat
Any early-season projection rests on prior-season form. The model has no mechanism for evaluating offseason roster turnover — free agency, the draft, scheme changes. That is a genuine limitation of any early-season model, ours included, and it is why we treat Week 1 through Week 4 as the least certain stretch of the season.
Never Used As Inputs
Betting odds of any kind. Same rule as MLB. Same automated audit — the training pipeline fails CI if any market-derived column shows up in the feature list.
Validation
Strictly walk-forward. Every season is predicted using only seasons that preceded it. Feature construction is verified leak-free by an automated test that rebuilds the entire feature set on a truncated history and asserts that no earlier value changes.
We intend to run NFL the same way from Week 1 — every pick published before the game, cryptographically timestamped via Git commit + Rekor, and settled honestly in public. Judge it as it accumulates.