Ranked vs. actually good
Since 2014, the power rating picked the winner in 76.8% of 2,149 games involving an AP-ranked team, against 75.5% for the poll. When the two disagreed (208 games), the power rating was right 57% of the time (90% interval 51% to 62%): a real edge, but a modest one.
- Treat a ranking far above the power rating with suspicion. Through week 11 of 2026, the numbers rate these ranked teams well below their poll spot: Kentucky (AP 13, power rating 44); Houston (AP 14, power rating 40); UCLA (AP 4, power rating 17).
- And the reverse: LSU (AP 24, power rating 7); Ohio State (AP 21, power rating 6); Arizona (AP 22, power rating 15) play better than their ranking says.
- Neither is a lock. The power rating misses about one game in four, and the betting market predicts margins better than any rating here (method).
The poll and the numbers, week by week
What the power rating is
The team ratings tool has the details. In short: every play’s expected points added (EPA), adjusted for the strength of each opponent a team actually faced, combined with CFBD’s Elo rating, which remembers past seasons. The result is in points: how much better than an average FBS team a team should be on a neutral field. Neither half is new on its own; together they predicted margins better than either (12.55 points of average miss, against 12.68 for Elo).
Polls are a different kind of thing. Voters rank résumés: who a team beat and whether it has lost. A team that wins close games against weak opponents can climb the poll while playing like a middling team, and a strong team with two early losses can sit low for weeks. The power rating doesn’t care about record, only how well a team has played, adjusted for whom it played.
How often each is right
Both were asked the same question before every game from week 5 on, 2014–2025: which team wins? For the poll, the better-ranked team (unranked counts as below everyone); for the power rating, the higher-rated team, as of the week before. Across 2,149 games with at least one ranked team, the power rating was right 76.8% of the time and the poll 75.5%. Most of the time they agree. In the 208 games where they didn’t, the power rating won 57% of the time; the 90% interval, 51% to 62%, stays above a coin flip, but not by much.
Method, data, and limits
Data. Play-by-play, games, polls, and pre-game Elo from CollegeFootballData.com for FBS seasons 2014–2026. Games between FBS teams only; bowls count as the week after the last regular-season week.
The comparison. Each game’s pick uses information from before it: the AP poll released for that week, and the power rating through the previous week. Ties in AP rank (two unranked teams) are left out. The power rating’s two weights (efficiency vs. Elo) were fit on these seasons’ games; its separate backtest, which predicts each season from the others, is in the tool’s method note.
What it can’t tell you. Picking winners is a coarse test: it ignores how confident each side was. The sample of disagreements is small (208 games), so the edge could be smaller than it looks. Injuries and quarterback changes reach the poll faster than the numbers. And the betting market, which isn’t a rating at all, beats both (12.09 points of average miss for the closing line).
Code: cfb/ · Data provided by CollegeFootballData.com. Sparvi is not affiliated with the NCAA, the AP, the College Football Playoff, any conference, or any school.