NFL Poll of Polls: Week 3

Week 2 is in the books, and it was a strange one. Baltimore lost at home to New Orleans, Chicago managed three points against Minnesota, Cleveland won in Tampa, and Carolina beat Atlanta 34–3. Nine outlets have filed Week 3 ballots, and the panel has moved with them.

This is the same poll of polls as last week — nine rankings, standardised and pooled — with one addition: a chart of every team’s move.

Here is where the nine outlets landed after Week 2.

Seattle and Buffalo are still inseparable at the top, on 79% and 21% of the bootstrap’s No. 1 finishes. The gap between them is 0.04 rank standard deviations, which is nothing. The interesting movement is underneath.

Who went up and who went down

Every team’s week-over-week change, biggest moves first. The centre line is no change; bars reach right for a climb and left for a fall. A climb is not the same as a win and a fall is not the same as a loss — a team that won can slip because someone below it won by more.

This week’s matchups, priced

Every team plays in Week 3. All 16 games are priced below from the consensus alone, with a sandbox for building any matchup you like.

The panel

Nine outlets, and they did not agree as tightly as last week. Mean pairwise Spearman is ρ = 0.881, down from 0.926 in Week 2 — that is what a chaotic Sunday looks like in the numbers. Split-half reliability is 0.983, so the panel is still closer to one voice than nine, but noticeably less so.

Mean distance from the consensus order, in rank positions:

FOX Sports and Bleacher Report sat closest to the consensus at under one position on average. At the other end, Sports Illustrated is a genuine outlier (ρ = 0.80, mean gap 4.2 positions, one team placed 15 spots from where the panel put it). That ballot is included rather than trimmed — the shrinkage in the estimator is what stops one dissenting voice dragging the consensus around — but it is worth knowing it is there, which is why the table exists.

Sources and method

Every number above comes from one pipeline. Nine outlets publish a weekly 1–32 ballot; each ballot is standardised, pooled in a measurement-error model, shrunk toward the panel mean, and published with a range reflecting how much the outlets actually disagree. The matchups are priced from the consensus score alone, with constants fitted on 6,719 completed regular-season games.

The method, the calibration and the week-by-week replication checklist live with the code rather than in every post:

• analysis/nfl-power-rankings/SOURCES.md — one section per outlet: URL pattern, cadence, markup and the specific traps. Follow this file to reproduce a week.
• analysis/nfl-power-rankings/METHODOLOGY.md — both derivations, and what the method does not claim.
• analysis/nfl-power-rankings/README.md — the pipeline and how to run it.

This week’s panel, every ballot verified complete:

Three things a replica would have to handle, rather than smooth over:

• Conor Orr’s SI column is not in this panel, though it ran. Its list numbers 31 teams — it jumps from 29 to 31 with Pittsburgh missing — so it is not a valid 1–32 ballot and the pipeline refuses it rather than guessing where the Steelers belong. The complete SI edition is used instead.
• FOX Sports changed its markup before this week, from #7 Team to 7. Team, which broke its parser until the extractor was switched. It is also moving toward client-side rendering, so it may drop out of reach in a future week.
• Neil Reynolds’ NFL.com UK column had not published when this ran, so NFL.com supplies one voice this week instead of two.

ESPN remains absent. Every URL on the site — including the homepage — returns an empty HTTP 202 to a plain request, so it cannot be read unattended. The omission is a decision, not an oversight.