Methodology

Overview

The NFL model produces weekly power ratings and point spread projections for all 32 teams. Its engine is a Bayesian regression fit from scratch every week on the season's play-by-play data: no Elo chains, no manual adjustments carried between weeks. Around that engine sit four layers: a starting quarterback adjustment, a rest adjustment, an offseason prior built from mean reversion and the season win-total market, and a capped blend toward the betting market on the published line.

Every published number keeps its provenance: the pure model line is stored next to the blended line, and the backtest grades the model against the closing spread, the strongest public benchmark there is. Nothing on this site is betting advice.

Data

All football data comes from the open source nflverse project: play-by-play with expected points added (EPA) per play, schedules with closing spread and total lines, depth charts, and injuries. Charted pressure data from Pro Football Reference and quarterback time-to-throw from Next Gen Stats feed the offensive line unit ratings. Live spread and total offers with prices come from The Odds API. History runs from 2015; historical team codes are normalized to current franchises so the 32 teams are continuous.

Plays in garbage time (large leads scaled by quarter) are excluded from every rating input, so running up or conceding late points does not move a team's rating.

Rating Engine

Each team carries an offense rating and a defense rating in points per drive, plus a pace rating in drives per game. A game contributes two observations, one per side: the home offense against the away defense and vice versa, with a fitted home-field parameter (prior of two points) split between them. The model is a conjugate Gaussian ridge: a closed-form Bayesian update whose posterior mean gives the ratings and whose posterior covariance gives honest uncertainty for every rating and every projection.

The target fuses two signals: points per drive (what happened on the scoreboard) and EPA per drive (the quality of the underlying process). EPA is first rescaled onto points, then the two channels are combined in proportion to how informative each one is, measured from their residual covariance. The fusion weights are estimated, not hand-picked.

The fit is refit from scratch before every projected week using only games that started before the forecast, with a recency half-life of six weeks selected by calibration. Projections come out as a full distribution over the game margin and total, so the spread, the total, and their uncertainties all come from one place.

QB and Rest Adjustments

Team ratings absorb the quarterback play that actually happened, so the QB layer exists for one case: the week the projected starter differs from the quarterback play baked into the rating. Every quarterback carries a rolling value in points per game, built from career EPA per dropback, shrunk toward replacement level by sample size. The game adjustment is the projected starter's value minus the value embedded in the team's rating window. It is zero in the normal case and moves several points when a starter sits. Projected starters come from depth charts with a manual override file for game-day news.

A rest adjustment from schedule rest days (byes, short weeks) is wired in the same way. Calibration currently tunes its coefficient to zero: at the selected market blend the line already carries the rest information, so the model does not double count it.

Market Blend

The published line is a weighted blend of the pure model margin and the market line, with the market weight selected by calibration and hard-capped at one half so the model can never become an echo of the market. Ratings are never touched by the market: the blend is applied at the output only, and the pure model line is published alongside the blended one so the model's own opinion is always visible.

Cover and push probabilities for priced offers use a discrete margin distribution rather than a smooth curve, because NFL margins pile up on three, six, seven and ten. A continuous distribution misprices those key numbers and cannot price a push at all.

Preseason

Week one has no games to learn from, so the preseason prior blends two signals: last season's final ratings regressed halfway toward the league mean, and a rating implied by the Vegas season win-total market, which prices offseason change (quarterback moves, coaching, roster) that reversion cannot see. The blend weights and the reversion strength were selected on early-season accuracy in the development seasons. The preseason ratings also serve as prior means for the in-season fits, so September ratings start from carried-over beliefs instead of zero and let the data take over as games accumulate.

Unit Ratings

The Units view shows opponent-adjusted ratings for rushing and passing offense and defense, pass blocking, run blocking, and special teams, each in points per game above league average. They are companions for reading a team, not components of the engine: they do not sum to the offense and defense numbers, and the model does not consume them.

The line ratings deserve their caveat: offensive line play is not directly observable in public data. Pass blocking uses charted pressures and sacks allowed, corrected for the quarterback's time to throw so a quarterback who holds the ball does not tank his line's grade. Run blocking credits the line for the short-area yards of each carry. Both attribute outcomes to units without snap-level film, so treat them as informed estimates.

Backtest

Every number on the History and Performance pages comes from a walk-forward backtest: each week of each season was projected using only information available at that time, including the preseason prior built only from earlier seasons. Model hyperparameters were selected on 2016 through 2021; the seasons from 2022 onward were never touched by selection and are the honest read on accuracy.

The benchmark is the closing spread. The blended model line runs within a few tenths of a point of the closing line's mean absolute error on the holdout seasons; the pure model runs about three quarters of a point behind. Beating the close consistently is rare, and claiming to is usually a sign of leakage, which is why the gap is reported rather than hidden.

Limits

The model does not know about injuries beyond the starting quarterback, weather, coaching changes mid-season, or motivation. Its interval coverage runs slightly narrow at the selected calibration. Week one to four ratings lean heavily on the preseason prior. The unit line ratings are proxies. And the published line deliberately borrows from the market, so its accuracy partly reflects the market's; the pure line is the model alone.

Nothing here is betting advice, and the pipeline never recommends a bet.

Tech Stack

The backend is Python: pandas and numpy for the engine (the Bayesian update is a closed-form linear solve, so a full weekly refit takes milliseconds), nflreadpy for nflverse data, and a six-table publish into Postgres on Supabase. GitHub Actions runs the pipeline on a Tuesday full run and Sunday and Thursday refreshes that re-check starters and lines. This site reads the published tables directly and renders on Vercel.