Every team on PredictionMarketsPicksβ NFL power ratings board has a different number today than it had yesterday. All 32 of them. Twenty-eight moved in the rankings.
Nobody played a game. Here's what happened, and what the board actually means.
Board as of Sep 1, 2026 β the tier table below reads that board. The "what moved" table is the Aug 21 rebuild this piece is about. Live board: /nfl/ratings.
The short version
Our published board was not being built from everything it was supposed to be built from. The biggest gap: the free-agency picture wasn't reaching the ratings we published. Retirements were getting through. Signings, in effect, were not.
That is a plumbing failure, not a modeling one β the method is unchanged, and nothing PredictionMarketsPicks does was re-thought. The model simply wasn't being handed its full input set before it published. We fixed it, topped the roster tracker up with signings recorded since early May, and re-ran the whole thing.
Then every number moved, because it should have.
First, what the number is
PWR is points per game above an average team on a neutral field. That's the entire definition. A team at +3.5 is expected to win by about three and a half points against a PWR 0 opponent, with home field taken out of it. A team at β3.5 loses by the same.
It splits into three pieces that all sit on that same points scale, which is what lets them add straight up:
That's deliberately a boring unit. A rating of 63.4 means nothing until you know its scale. A rating of +3.5 tells you the expected margin.
The six tiers
The tier next to each team is a plain-English bucket for that number, so the board is skimmable.
| Tier | PWR range | Teams (board as of Sep 1, 2026) |
|---|---|---|
| Elite | +7.0 and up | none |
| Contender | +4.0 to +7.0 | NE, SEA, LAR |
| Playoff | +1.0 to +4.0 | BUF, GB, JAX, DET, PHI, HOU, CHI, DAL, DEN, BAL, SF |
| Average | β2.0 to +1.0 | IND, ATL, PIT, TB, LAC, KC, CIN, NYG, WAS, CAR |
| Below Avg | β5.0 to β2.0 | NO, CLE, MIN, MIA, ARI, TEN, NYJ |
| Rebuild | below β5.0 | LV |
The tier is derived from the rating and never feeds back into it. Nothing we price β model lines, win totals, championship odds β reads the tier. It's a reading aid.
Nobody is Elite. On the Sep 1 board New England leads at +5.50 and the cutoff is +7, so the top tier sits empty. We publish it that way rather than squeezing the ladder so somebody occupies the top, because a scale that moves to guarantee a #1 isn't telling you anything.
What moved (the Aug 21 rebuild)
| Team | Was | Aug 21 | Ξ | Rank |
|---|---|---|---|---|
| NYJ | β6.28 | β4.12 | +2.16 | 32 β 31 |
| TEN | β5.73 | β3.58 | +2.15 | 31 β 29 |
| DAL | +0.59 | +1.90 | +1.31 | 15 β 10 |
| CLE | β3.22 | β2.11 | +1.11 | 29 β 25 |
| WAS | β1.99 | β1.09 | +0.90 | 25 β 23 |
| β¦ | ||||
| NYG | β0.22 | β1.08 | β0.86 | 18 β 22 |
| LAC | +0.42 | β0.65 | β1.07 | 16 β 21 |
| MIA | β2.03 | β3.19 | β1.16 | 26 β 28 |
| MIN | β1.16 | β2.46 | β1.30 | 22 β 27 |
The pattern is exactly what you'd expect from a board that had been missing free agency: the teams that moved most are the teams whose offseason mattered most. The Jets and Tennessee were being scored as if their additions hadn't happened, and both climbed out of the Rebuild tier. Dallas gained a full 1.31 points and moved into the Playoff tier. Minnesota went the other way.
At the top, Seattle stayed first on the rebuild but came down from +6.43 to +5.63 β the gap between the Seahawks and the field was overstated. New England and the Rams swapped into second and third. Las Vegas came out as the league's worst team at β6.16, a spot it had previously been sharing.
Those are the Aug 21 figures. The board was republished on Sep 1, 2026 with the final preseason inputs, and the top three reordered by hundredths: New England +5.50, Seattle +5.48, the Rams +5.13. Minnesota sits at β2.64 and Las Vegas at β5.99 on that board. The live board is always the number of record; this piece is about why everything moved on Aug 21.
The other thing we fixed
While we were in there we found the tier labels were being graded against an earlier version of the number than the one we publish β so a label could describe a draft of a rating rather than the rating printed beside it. At its worst that put 18 of 32 teams in a bucket our own published cutoffs said was wrong, and three of the six tiers could never be printed at all.
That's corrected too. Every tier on the board now comes from the same finished number sitting next to it, and the ladder in that table is the ladder on every surface we publish it.
We also found a few pages had been serving an older copy of the board than the one pricing our games. Everything now reads from one source β the hub, the ratings board, the free tool page, the embeddable widget, and the version AI agents pull β with a check that fails loudly if they ever drift apart again.
Why say any of this
Because the alternative is that 32 numbers change overnight and nobody explains it.
We publish a track record that includes the engines losing money, so quietly re-running a model and hoping nobody diffed the board would be a strange place to start being cagey. If you'd built anything off the old numbers β a shortlist, a filter, a read on a team β you should know they moved and why.
The method behind the ratings is unchanged and documented at how we build them, with the vocabulary in the glossary. What changed is that the board now gets built from everything it's supposed to.
Season starts September 9. From there the ratings update on played games rather than roster moves, which is a very different and much more interesting problem.
Educational analysis, not financial advice. Model outputs are estimates, not guarantees β trade responsibly.