Anatomy of a prediction

All four pick SEA. Our headline model (QBElo) is 7 percentage points more confident in SEA than the closing market. The rest of this page is how we got there.

The schedule row (selected fields)

What else the model has already seen by Sep 9

• All NFL games 2010–2025 (4,175 regular-season games) — team Elo + QBElo ratings have been updated through end of 2025 and regressed 1/3 of the way back to the mean at the off-season boundary.
• Every QB’s box-score history — QB ratings carry over un-regressed (skill is sticky); team QB baselines also carry over.
• Every play’s EPA from 2010–2025 — rolling 10-game form means for each team, shifted one game.

e. Everything merged: the one row that goes into the ML model

ml_model.assemble() stitches all of the above into a single row keyed by game_id. The 13 features below are exactly what the ML model consumes — same shape as every other game’s row.

Note: p_home_mkt (the de-vigged Vegas implied probability) is also in the row but the ML model doesn’t use it — we leave Vegas out of training so it stays an honest benchmark.

Elo — from team ratings to win probability

Standard 538-style Elo: add home-field advantage to the rating diff, then run through the Elo logistic.

The 400 in the denominator is the Elo scale (a 400-point gap = 10:1 odds). +48 of HFA means a home team beats an equally-rated road team about 56% of the time.

QBElo — Elo plus a QB adjustment

Same logistic, but the rating gap also includes qb_to_elo × (starter_rating − team_qb_baseline) for each side. With 2026 starters NaN, that adjustment is 0 here, so the math collapses to QBElo’s own team-Elo subsystem.

QBElo’s team-Elo subsystem (1685.53) is higher than plain Elo’s (1675.91) for SEA — the two rating systems update under different predictions over thousands of games and drift apart. Once 2026 starters are published, home_qb_adj and away_qb_adj become non-zero and QBElo can pull sharply if a backup is announced.

ML (Logit) — learned weights on the 13-feature row

The pipeline is a four-step sklearn.Pipeline: impute missing values with the median, standardize each feature, fit logistic regression, then Platt-scale on the 2025 season so the output is a calibrated probability.

In holdout testing the ML model gets slightly better straight-up accuracy than QBElo but worse Brier — the logistic is a touch more confident than it should be on lopsided games. Platt scaling on the most recent full season pulls it back toward calibrated.

Vegas — de-vigging the moneylines

Each American moneyline implies a raw probability that includes the book’s margin (the “vig”). The two sides sum to more than 1.0. We normalize so they sum to exactly 1.0 — the closest thing to an honest market probability we can extract for free.

All probabilities on this site labeled “Vegas” or “market” are de-vigged like this. The implementation lives in src/metrics.py (american_to_prob + market_home_prob).

a. The raw data updates (Monday/Tuesday)

• nflverse publishes the final result, home_score, away_score on the schedules row.
• Player stats for the game are added to the weekly stats parquet (used for QB value).
• Play-by-play with EPA per play hits the pbp parquet (used for rolling form).

b. Our cron runs (Tuesday 16:00 UTC)

The GitHub Action re-runs the whole pipeline and pushes a refreshed web/replay_2026.json if any prediction changed. Vercel auto-deploys.

Each model gets graded on the games that have already been played:

Brier and log loss reward calibration (a confident wrong pick hurts more); accuracy is just who-picked-the-winner.

c. The ratings update online (next time the engines run)

Both Elo and QBElo update each team’s rating after every game. Worked example: if SEA wins NE@SEA by exactly 7 (covering the −3.5):

SEA gains ~13 points by winning “as expected.” A loss would’ve cost them more like 27 (because y − p_home would have flipped sign and grown in magnitude). Online updates like this are why Elo doesn’t need a re-train: it adapts every week.

d. The site updates itself

The refreshed JSON includes new cumulative and weekly Brier / log-loss / accuracy. The live tracker reads it next page load and the running scoreboard moves. The Week 1 page shows the result chip next to each pick.