How the NFL model works

Each club gets a rating built from scoring margin, not record: 1500 plus 25 points for every point of average margin per game. A 9-8 club that won six one-score games is not the club its record says, and over a season margin predicts what comes next far better than wins do.

Early in a season a few games say little, so the rating starts entirely on last season's and moves onto this season's in a straight line over the first eight games. By week nine it is this season alone. See the power ratings.

A game's projected margin is three terms added together:

• The rating gap, shrunk. Each point of rating difference is worth 0.024 points of margin (0.61 of a point per point of average margin). Taken at face value the gap is too confident: a club rated a touchdown a game better than its opponent is not a touchdown favourite. Shrinking it is the single most important thing the fit found.
• Home field. 1.49 points, plus 0.42 of a rolling estimate of how much home teams have been winning by lately — about 2.2 points this week. None at a neutral site.
• The quarterback. Each quarterback has a value — his expected points added per dropback, shrunk toward the average — and each club a baseline for its usual starter. When a club starts someone other than its usual man, the difference moves the number by 18.3 points per unit. For a game not yet played the starter is ESPN's depth-chart quarterback, skipping one listed out, doubtful, on injured reserve or suspended; a quarterback we have never seen start is rated below an average starter (-0.1), because first starts usually are.

The margin becomes a win probability through a normal curve with a spread of 13.1 points — how far real results land from our number.

The weights were fitted on every completed regular season from 2017 to 2025 and are frozen for 2026, so the whole season is scored against one version. Only the inputs move week to week: the ratings, the quarterback values and the home-field estimate. In testing on seasons the fit never saw, the quarterback term changed almost nothing in ordinary games and helped where it should — the games in which a club started someone other than its usual quarterback.

The ratings read the current standings, which include every game already played. Projecting a finished game from them would use its own result. So a pregame number is shown only for a game that has not kicked off, and past games are rebuilt from the ratings as they stood at kickoff wherever a past number is needed.

During a game the final margin is treated as uncertain around an expectation that blends the score on the board with the part of the pregame edge still to be played. The uncertainty shrinks as the clock runs — starting from 12.94 points at kickoff, fitted on NFL play-by-play rather than borrowed from college — but never below a floor, because the last minutes of an NFL game hold far more snaps than their share of the clock. Overtime is priced from how NFL overtimes have actually gone, where the team with the ball in a tied overtime wins more often than not.

Every fourth down is three choices — go, kick, punt — each priced in expected points and then in win probability. The four curves underneath (expected points by field position, conversion by distance, field-goal success by distance and punt distance) are measured from 381,814 NFL plays from 2016 to 2025, not typed in from a chart.

The playoff odds play every remaining game with the pregame model (usual quarterbacks, since who starts in December is not known in September), seed each conference — four division winners, three wild cards — and play the bracket, 10,000 times. NFL tiebreakers are approximated, and the page says how.

It is a calibrated, transparent opinion of our own: a way to see how likely each result is, why, and how that changes when a quarterback does. When we say 70%, clubs in that spot should win about seven times in ten.

It is not a betting edge and is not offered as one. The closing market line uses everything we use and more, and it is the better predictor of the final margin; our number does not beat it. Where the two differ, the market page shows the difference as information about two opinions, not as a recommendation.