How the model works
- Data: every NFL regular-season and playoff game for the current season plus the prior 3 seasons from ESPN's public scoreboard. Past seasons are cached on disk; the current season refreshes every 15 minutes. Finished games are graded against ESPN’s closing line (in-game odds are ignored); upcoming lines come from ESPN.
- Power rating: an adjusted-margin rating in points (an SRS-style least-squares fit). Margins are capped to limit blowout inflation, home field is removed, and ratings start from last season's rating (regressed toward average) so early-season numbers are stable.
- Offense / Defense: opponent-adjusted points scored and allowed per game compared with the league average (+ is better for both).
- SOS: the average power rating of opponents played. Remaining SOS uses the opponents still to play.
- Prediction: home margin = rating(home) − rating(away) + home field (0 at neutral sites), adjusted for a starting quarterback listed out, the gap between the two starters’ EPA per play, and the time zones the visiting team crosses. Win probability = Φ(margin / σ). Edge = model margin − market spread.
- Backtest is walk-forward: each week is predicted using only games played before it, and each game is priced with what was known before kickoff (the quarterback listed on the final injury report, not the one who played). It scores the ratings model only; Deep Dive adds XGBoost and is the site’s record. This page loads only the current season plus the three before it, so its ratings start from fewer seasons, and its records come from the browser’s own copy of the model: they can differ from Deep Dive’s, most for the oldest season.
Seasons loaded
—
Games
—
Current season
—
Upcoming w/ line
—
Model SU acc.
—
Model ATS
—
Model O/U
—
Model settings
Prior regression is the share of last season's rating carried into the new season. Prior weight is how many games that prior counts for. Market weight blends the betting line into win probability and projected score (ATS edges still use the pure model line). The defaults are the rating model’s settings, with its own edge cut-offs; the live XGBoost engine behind the game cards uses its own cut-offs, listed on Deep Dive, which also has the live record. The Backtest tab shows how each setting performs; edges are leans, context only, until a record is a proven edge.
| Matchup (away @ home) · Ratings | SU pick · Win % · Fair ML |
Model · Market · Edge · ATS |
Score Projected · Actual |
Model · Market · O/U pick |
|---|
Public news headlines, social posts, injury/suspension mentions, coaching chatter and a sentiment estimate for both teams.