NFL Power Ratings (2026 season)

A roster-based, points-denominated power-rating system. Every team's rating is expressed in points vs. a league-average team (0.0) and is the straight sum of its components:

Team Rating = QB + Offense(non-QB) + Defense   (+ Coaching + Scheme + Edge, future)

All numbers live in data/ratings.csv — that is the one file you edit. Three scripts read it and produce the artifacts below. Edit the CSV, re-run the script(s), done.

Files

data/
  ratings.csv      THE file you edit. 32 teams: QB/Off/Def values + names + notes.
  qb_depth.csv     Backup QBs (2nd/3rd string) with values. Feeds the QB Depth tab.
  prior_2025.csv   Last season's ENDING ratings (reference prior).
  hfa.csv          Home-field advantage: base 1.5, per-team bumps. Used by spreads.py.
  config.csv       Home-field default, season label.
  writeups/        Optional per-team analysis: one <ABBR>.md per team (e.g. SEA.md).

build_ratings.py   -> NFL_Power_Ratings_2026.xlsx  (workbook: ratings, QBs, depth, projections)
generate_site.py   -> output/ratings-preview/YYYY-MM-DD/index.html (private preview)
spreads.py         -> terminal + spreads.html       (schedule + my line vs. ESPN market line)

Commands

cd ~/nfl-power-ratings

# After editing data/ratings.csv, regenerate whichever you want:
python3 build_ratings.py          # the Excel workbook
python generate_site.py
# -> output/ratings-preview/YYYY-MM-DD/index.html (dated private preview)

# Production output is explicit and is run only after preview approval:
python generate_site.py --output docs/index.html

# Spreads vs. the market (free, unofficial ESPN JSON API — no key/install):
python3 spreads.py 1              # one week
python3 spreads.py all            # full 18-week season + summary
python3 spreads.py all 2026 --html # also write spreads.html

# Open output/ratings-preview/YYYY-MM-DD/index.html in your browser.

Generation stops if any row in data/ratings.csv has needs_review=Y. Clear those flags through editorial review; never bypass the gate. The default command writes a private preview and does not replace the GitHub Pages artifact.

Independent PGO model comparison (private preview)

Matchup preview

python pgo_matchup_comparison.py 1 2026 --pgo-ratings research/pgo_v1/ratings_2026_preseason.csv

This writes a private output/pgo-matchup-preview/<date>/index.html; it does not publish. The current experimental/HOLD receipt is non-certifying.

python pgo_challenger.py --as-of 2026-07-21T12:00:00-04:00 rebuilds the locked pgo_v1 receipt. Exit 0 is validated PASS, exit 1 is an honest statistical HOLD, and exit 2 is BLOCKED. An integrity-eligible HOLD writes 32 ratings labeled EXPERIMENTAL; BLOCKED writes no ratings.

python pgo_comparison.py compares the eligible PGO snapshot with Sean McCabe's reviewed ratings and writes a dated private page under output/pgo-comparison-preview/. It never changes docs/index.html or any live service. PGO v0 remains backtest evidence only.

After editorial review and explicit publication approval, the fixed-destination release command is:

python pgo_comparison.py --publish

It writes only docs/index.html. PGO v1 remains labeled Experimental model — HOLD; the command does not modify Shopify or any rating input.

Team write-ups (click-to-expand on the site)

Every team row in the generated ratings artifact expands (click it) to show a QB/Off/Def bar breakdown plus your analysis. The analysis comes from data/writeups/<ABBR>.md — plain markdown (## headings, **bold**, - bullets, paragraphs). If a team has no write-up file yet, the row falls back to the one-line notes from ratings.csv and shows a hint with the filename to create.

# Team abbreviations match the chips (BUF, SEA, KC, LAR, ...). To add one:
$EDITOR data/writeups/BUF.md      # write markdown
python3 generate_site.py          # rebuild the dated private preview

Two examples ship already: data/writeups/SEA.md and KC.md.

Publishing to postgameoutlet.com

Shopify owns the public content and commerce shell. The approved ratings artifact remains independently hosted and is embedded in the native Power Ratings page. See SHOPIFY.md for the preview-first embed and release workflow.

Rating conventions (how the numbers were calibrated)

• QB: the dominant lever. 0.0 ≈ ~QB16-18 (a middling starter). Elite +5 to +6.5, worst starter floor ~ -2.5. Backups: 2nd string -2.5 to -4.5, 3rd string -4.5 to -6.0. QB value is expected 2026 value, so injury uncertainty is priced in (e.g. Mahomes carries an injury haircut despite being the most talented).
• Offense / Defense (non-QB units): centered at average — ~16 teams above 0.0 on each, almost all within +/-1.0. Driven by the 2026 FA/draft roster movement; the notes column in ratings.csv records the key adds/losses behind each number.
• Spread model (spreads.py): my_margin(home) = rating_home - rating_away + HFA, where HFA = per-team base (default 1.5) + 0.5 for a primetime home game. my_spread = -my_margin (negative = home favored). edge = market - my_spread; |edge| >= 1.5 is flagged.

The prior / blend (history)

The 2026 preseason numbers were sanity-checked against prior_2025.csv (last season's ending ratings) — injury-deflated finishers (KC/CIN/BAL, whose QBs were hurt) were trusted toward the roster build, hot finishers (SEA) toward the prior. The displayed Rating is now the straight component sum, NOT a blend; the prior is kept only as a reference column on the website.

PGO team model (shadow only)

python pgo_model.py runs a pinned, chronological backtest of Postgame's independent team-results model and writes its receipt under research/pgo/. It does not read Sean McCabe's QB/offense/defense inputs or any market line. A PASS makes the shadow ratings eligible for human review only; it does not publish them or add them to the ratings site.

PGO forward-looking challenger (shadow only)

Install its single dependency with python -m pip install -r requirements-pgo.txt. python pgo_challenger.py --freeze-sources --as-of <ISO-8601> explicitly freezes a research snapshot; later python pgo_challenger.py --as-of <same value> runs offline from the lock. Outputs stay in research/pgo_v1/.

PASS permits private prospective shadow tracking only. HOLD writes diagnostics and no ratings. Neither result publishes or changes McCabe ratings, Shopify, or GitHub Pages.

Notes & caveats

• ESPN endpoint is undocumented/unofficial; spreads only populate close to game week. Early-summer lines are placeholders — don't over-read specific edges yet.
• If one or two teams dominate the season-long edge list, that usually means YOUR rating on those teams is the outlier, not the market. (As of last build: GB and NYJ recurred — worth a sanity check when real lines firm up.)
• Coaching / Scheme / Edge columns exist in ratings.csv but are still 0.0 — the natural next layer.