How we built our NFL power rating

Detroit is number one in our NFL power rating. The Lions won nine games last year. That sentence is the reason this article exists, because the model that produced it had to be torn down and rebuilt before it could see what the scoreboard was saying.

The challenge

Our first NFL rating was simple on purpose. It predicted season win totals from prior season data, and it carried an uncomfortable finding we published anyway: last year’s wins, regressed toward the mean, beat every richer feature set we tested. Point differential, EPA per play, coaching changes, quarterback turnover. Out of sample, none of it out-predicted the plain baseline, so the plain baseline is what went live, with its error printed next to it.

Then it got challenged from two directions at once. A reader made the obvious objection and the founder made the same one internally: raw wins have a terrible memory. In 2025, Denver won fourteen games on a plus 90 point differential, about plus 5.3 a game. Seattle also won fourteen, on plus 191. A model that reads wins alone thinks those two teams are the same team. Anyone who watched them knows they were not. If the rating cannot tell a dominant fourteen wins from a lucky fourteen wins, it is grading the standings, not the football.

Fair hit. We rebuilt.

The build

The rebuild changed the unit of analysis. Instead of asking “how many games will this team win,” the new model asks a smaller question thousands of times: “by how much should this team beat that team.” Team ratings come from a ridge regression on game margins across 6,223 games from 2002 through 2025, every one of them with a closing spread attached.

Three details matter.

First, the ratings are opponent adjusted by construction. A twenty point win over a good defense moves your number more than a twenty point win over a bad one, because every margin is scored against the rating of the team on the other side.

Second, the model is fit walk-forward. At every week of every season, the ratings are rebuilt using only games played strictly before that week. No prediction ever sees the future. The window covers three seasons with decayed weights: current season games at full weight, last season at 0.2, two seasons back at 0.1. The tuning that picked those weights ran on 2004 through 2021 only, 4,624 games, and never touched the seasons we would later test on.

Third, extra features had to earn their place one layer at a time, judged only on held out data. Rest differential failed. Per team home field advantage failed. Roster age squeaked in with an improvement small enough that we treat it as decoration. The real find was quarterback continuity: a team starting someone other than its expected quarterback costs about 2.8 points, and that layer produced a genuine out of sample improvement. Everything that failed the test was cut, including features we liked.

Did it beat the old model

Held out seasons 2022 through 2025, average absolute error in points when predicting game margins. The old model’s numbers come from its own functional form refit on pre 2022 data, so both models sat the same exam.

Season	Games	v3	v1 implied	Closing spread
2022	271	9.235	9.413	8.742
2023	272	10.490	10.682	9.901
2024	272	9.953	10.785	9.610
2025	272	10.323	10.819	9.722
Average	1,087	10.000	10.425	9.494

The new model beat the old one in all four seasons, by 0.425 points of error on average. That is the win, and it is real.

Look at the third column, though. The closing spread beat both models in all four seasons. Which brings us to the part of this article most services would never write.

The number most services would hide

Grade the model against the spread. Take its predicted margin, compare it to the closing line, pick the side it implies, and count. Pushes excluded from every denominator.

Season	Record	Pct
2022	125-136	47.9%
2023	118-140	45.7%
2024	138-130	51.5%
2025	123-148	45.4%
Pooled	504-554	47.64%

47.6 percent against the spread. A model that beats our old model in four straight seasons still loses to the market. That is not a bug in our build. The closing line is the sharpest number in football. It is the compressed opinion of every syndicate, every model, and every dollar that moved it, and any shop telling you their power rating out-picks it week after week is selling you a story.

So we use the model for what it proved it can do, and we refuse to use it for what it cannot. It predicts margins better than our old model, which makes it a better engine for expected wins and for reading season long markets. It is not a spread picking machine, and we do not dress it up as one. That is exactly why a lean only appears on our board when the gap between our number and the market’s number beats the model’s own error. No gap that clears the bar, no lean. The model earns your attention by knowing its limits, and by putting them in a table instead of a footnote.

What it changes on the board

The old model had Denver sharing the top spot at 10.65 expected wins. The new one drops Denver to 9.75, seventh, because those fourteen wins came with thinner margins than the record suggested. Detroit takes over at number one: 10.76 expected wins, built on nine wins at plus 4.0 points per game. That profile, a strong scoring margin hiding under a mediocre record, is precisely the case a margin model exists to catch, and the market has Detroit’s win total hanging at 10.5.

The biggest gap on the board right now belongs to Miami. The model puts them at 7.21 expected wins against a 3.5 line, a plus 3.71 gap, and that lean survives a model that fully sees Miami’s minus 4.5 points per game last season, the new coach, and the new quarterback.

The full 32 team table, every rating, every expected win number, and every gap against the market lives at edgelabs.bet/nfl/power-ratings. Members get the leans first, plus the tools and the database behind them. Founding access is open at edgelabs.bet/join.