Why our model demoted Denver

In the first version of our NFL rating, Denver tied for the top expected win total in the league, 10.65, alongside New England and Seattle. In the version that replaced it, on its Aug 18 build before week 1, Denver ranked seventh. Same team, same 2025 season on the books, a different number because we changed what the model is looking at.

The unit of analysis changed

Our first rating priced a team’s whole season at once: take last year’s results, add a few adjustments, and predict a win total. That approach kept losing to a plain baseline, a team’s prior win count carried forward with almost no adjustment at all. If a simple prior beats your model, the model isn’t earning its complexity.

So we rebuilt around the game instead of the season. The new rating comes from a ridge regression on individual game margins:

home_margin ~ home_team_effect - away_team_effect + home_field

fit walk-forward at every point in the schedule on a three-season decayed window: the current season’s games carry full weight, last season carries 0.2, and two seasons back carries 0.1 (ridge alpha 5). Every prediction for a given game uses only the games played before it. Home field advantage is not shrunk toward zero; the 32 team effects are, which keeps a hot three-game stretch from swinging a rating more than the evidence supports.

A second stage turns that rating gap into a predicted margin: 2.122 plus 1.0453 times the rating gap, plus a small roster-age term, minus 2.77 points for a team starting anyone other than its expected quarterback. That QB term is the real find in this build, worth about 2.8 points whenever a team turns to a backup. We tested rest differential and a per-team home field adjustment too. Both made the model worse on games it hadn’t seen yet, so both got cut. A layer earns its spot in the formula or it doesn’t ride along for the story.

Expected wins for 2026 come from walking each team’s real 272-game schedule and summing the probability that its predicted margin in each game is positive.

Denver, specifically

Denver’s 2025 win total was strong enough to tie for first in the old model. Its margins were not. The season-level approach rewards the win column. The game-level model prices how a team actually won, point by point, and Denver’s games were closer than its record implied. New England shows the identical pattern and falls for the identical reason.

That is the whole mechanism behind the demotion. The model is not marking down Denver’s talent or its coaching. It is separating two questions that a record-based view blurs together, how many games a team won and by how much it won them, and those questions can rank the same team very differently.

Denver’s rating under the new model, on the Aug 18 build before week 1, was plus 2.51 points against a league-average opponent on a neutral field, which worked out to 9.75 expected wins, seventh in the league. The market’s win total on the 8/17 board was 9.5. That number sat a quarter of a win above the market’s, which was not a gap we called an edge. It was context: a model that used to have Denver tied for best in football landed almost exactly where the market already sat. The live board has moved since: as of the Sep 16 refit, Denver sits 17th at 8.40 expected wins, lower still, not back up.

Detroit was the mirror image of this story and the clearest proof the mechanism works as intended. The Lions won 9 games in 2025 while outscoring opponents by about 4 points a game, the kind of team a record-based model undersells and a margin-based model catches. On that same Aug 18 build, Detroit was our number one team at 10.76 expected wins, up from 8.73 in the old model. On the Sep 16 refit, Baltimore leads at 11.28 expected wins and Detroit sits fourth at 10.65. See edgelabs.bet/nfl/power-ratings for where every team sits today.

What this number is and is not

The game-margin model beat our old season-level approach at predicting actual margins in all four seasons we held out for testing, 2022 through 2025. That is a real improvement, and it is the point of the rebuild. It does not beat the closing spread. This model fronts expected wins and margin context, not an against-the-spread pick.