NFL Power Rankings: All 32, By the Numbers

These are the model’s rankings, not ours — the exact Elo ratings behind the prediction ledger, sorted 1 to 32. No panel, no eye test, no Tuesday-morning vibes: the order changes only when a final score changes a rating. Right now that means a fully-stated walk-forward Elo carrying 2025 results plus a one-third regression toward the mean — the same opening numbers the ledger froze, and the same model that expects to get 99 games wrong this season. It knows nothing about rosters, injuries, quarterbacks, or preseason; where that blindness bites is stated plainly below, not hidden.

The board, as of Sep 15

Movement compares today’s ranks to the frozen snapshot taken opening ratings — never to yesterday’s build. Tier lines are drawn wherever adjacent teams sit 25+ Elo points apart — about half a home-field edge — so the tiers are computed, not curated.

What the tiers actually say

The board’s real tier lines this week: below #1 (48 points), below #2 (42 points), below #31 (61 points). Everything closer than 25 points is a coin flip with a lean, and treating neighbors as meaningfully ordered is how power rankings usually lie — a two-point Elo gap is noise, and the table above refuses to pretend otherwise.

Best and worst seat in each division

How to read an Elo gap

The scale is relative and the differences are what matter. A 48-point edge — exactly what the model grants a home team — converts to about a 57% win probability. The full stretch of today’s board, #1 Seattle Seahawks (1683) over #32 Tennessee Titans (1321), is 362 points: on a neutral field the model makes that a 89% game — heavy, and still nowhere near a lock, which is why the miss budget exists. The formula is the ledger’s own: 1 / (1 + 10^−gap/400).

What this page refuses to know

• No rosters, no injuries, no quarterbacks. A team that lost its starter yesterday keeps its rating until losses drag it down — and QB changes matter. The model pays for that blindness in its misses, in public.
• Preseason means nothing here. August games are not fed to the model (the signal isn’t there anyway); until week 1 grades, these ranks are 2025 plus regression, nothing more.
• Movement cannot be invented. Snapshots freeze per graded week in a committed state file; rebuilding the site twice in a day cannot manufacture an arrow.

Sources: ratings from the frozen ledger at data/predictions.json; records and schedule from the bundled nflverse game log at data/games.csv. Generated by the site’s build (build_power_rankings.py) with no hand-entered numbers anywhere.