Game quality

What each result was worth beyond the score. Every play is graded by EPA (expected points added), then adjusted for who it came against: a defence's rating is what it allowed to everyone else that season, so a big day against a bad defence counts for less. Season views include playoff games.

Case Keenum — 2023

Adj EPA / dropback
-0.346
Raw EPA / dropback
-0.350
Dropbacks
59
Turnover EPA
-12.7
WPA
-0.30
WeekOpp DropbacksAdj EPA/dbRaw EPA/dbScheduleTotal adj EPA INT / FLTurnover EPASacksSack EPAWPACPOE
15TEN 39 -0.109 -0.065 -0.044 -4.2 1 / 0 -7.1 3 -4.3 -0.03 +0.5
16CLE 20 -0.809 -0.906 +0.096 -16.2 2 / 0 -5.6 3 -5.2 -0.27 -9.6
Opponent adjustment. A unit's rating for a game is its average over its other games that season, blended with 16 games' worth of prior (league average plus half of last season's gap from it; 24 games for pass defence, which is noisier). Early in a season that leans on last year; by midseason it is mostly this year. Chosen by how well it predicts a held-out game: it cut that error by about a quarter against no shrinkage, and adjusted numbers agree between odd and even weeks more than raw ones do (team EPA margin r 0.52 → 0.55, defence 0.32 → 0.36, QB EPA/dropback 0.47 → 0.49, 2016–2025).
Win expectancy is a logistic fit of wins on EPA margin over 6,036 completed games (it names the winner 86.5% of the time). These describe what happened; one game is not a forecast. Play-by-play credits the quarterback with every sack and dropback, so a bad line shows up in his numbers. Seasons 2016–2026, built 2026-09-14 01:25 UTC, from nflverse. Write-up: backtests/nfl_game_quality/FINDINGS.md.