How We Rank NFL Teams

The 2025 prior

We fit an offense and defense rating (in points) for every team from the entire 2025 regular season — all 272 completed games — using a margin-based model: each team's points scored and allowed, adjusted for opponent strength and home field. That gives each team a real, honest "how good were you in 2025" number.

We then regress that rating 50% toward the league mean to get the 2026 preseason prior. That 50% is a starting choice we have not backtested against real NFL outcomes — unlike our college football model, which was tuned against two full seasons of results before shipping, this number has not been checked against how NFL teams actually perform after a rating like this. We picked it because NFL rosters turn over year to year and a full season of results is a small sample to take at full face value, not because we measured it as the best number.

The published 62-99 scale

Ratings are shown on the same 62-99 scale as every other sport in this app, for one consistent number system. The curve is fit to NFL's own observed spread of team strength (not borrowed from our college football scale, which was tuned for a very different, 135-team population) — the best 2025 team lands at 99, the worst at 62, and everyone else scales linearly between. NFL's real parity (a salary cap and revenue sharing) means NFL ratings naturally cluster closer together than college football's do — that is a real difference in the sport, not a flaw in the math.

Weeks 1-2: still the labeled prior

Ratings stay pinned to the 2025-based prior through week 2 — one or two games isn't enough signal to move a number responsibly. From week 3 on, real 2026 results are blended in using a ridge regression toward that same prior, so a team that has proven itself different from its prior earns a rating that reflects that, while a team with only a game or two still leans heavily on the prior.

The ridge strength for that in-season blend is borrowed directly from our college football model. We have not independently validated that constant for a 32-team league — college football's version was walk-forward tested against two full seasons of results before it shipped; NFL's hasn't been, because there wasn't time to run that study before the season started. We'll revisit it once real 2026 results exist to check it against.

Ties

An NFL game can end in a tie. When it does, we exclude it from the model's straight-up record and accuracy check entirely, rather than inventing a half-win convention. It still counts toward each team's public win-loss-tie record; it just never enters our grading.

What this is not

This is not a betting model and carries no spread record. We don't have an NFL odds source in this pipeline, and RATINGS.md — the standing rules this whole ratings system follows — rules out that kind of claim for any sport in this app. Win probabilities shown on team-week pages and in the calibration panel come from these ratings alone.