College football team ratings
Every FBS team, week by week. The power rating is how many points better than an average FBS team a team should be on a neutral field. It combines two things: how efficiently the team has actually played (EPA per play, adjusted for the strength of every opponent it faced) and CFBD’s Elo rating, which carries information from past seasons. In a backtest the combination predicted next week’s margins better than either one alone (see the note below). The chart shows the efficiency half on its own: offense across, defense up. Move the week slider to watch the season unfold; ratings before week 4 rest on very few games and are faded.
Opponent adjustment. A ridge regression of every play’s EPA on the offense, the defense, and home field estimates each team’s strength; a team’s adjusted number is its own EPA/play minus the strength of the opponents it actually faced. The penalty (λ = 50) was chosen by how well the ratings predicted the next week’s margins. Garbage time removes plays when the margin is more than 38 points in the 2nd quarter, 28 in the 3rd, or 22 in the 4th.
The backtest. From week 4 on, ratings through each week predicted the next week’s games between FBS teams, 2014–2025, bowls included; each season was predicted from a model fit on the other seasons. Average miss on the same 2,917 games, in points: power rating 12.55, CFBD Elo 12.68, our efficiency rating alone 13.02, raw EPA/play 13.35, home field alone 15.08. The power rating beat Elo by 0.13 points per game (90% interval 0.07 to 0.20). The betting market did better than all of them: the median opening line missed by 12.27 and the closing line by 12.09. Lines are used only as this after-the-fact benchmark.
Against CFBD’s WEPA. WEPA (CFBD’s own opponent-adjusted EPA) is only published as season totals, so it can’t be tested inside a season without seeing the future. Instead, each season’s final numbers predicted the next season’s first 4 weeks (2,001 games): our rating missed by 14.22 points and WEPA by 14.52 (ours better by 0.30; 90% interval 0.21 to 0.39). Preseason Elo (13.50) and SP+ (13.68) beat both: they also account for returning players and regression toward the mean.
What it can’t tell you. Ratings early in the season rest on a handful of games. Games against FCS teams count for the opponent adjustment but aren’t shown. Injuries, quarterback changes, and motivation aren’t in the data. CFBD’s play-by-play has gaps in about 3% of games (data status).
Model version. CFB power rating v1.0.0 (experimental: validated before the backtest harness existed; its harness port is pending).
Data provided by CollegeFootballData.com. Sparvi is not affiliated with the NCAA, any conference, or any school.