The committee punishes the loss, not the margin
After a loss by 7 points or fewer, a top-12 team fell 3.9 spots on average in the College Football Playoff committee’s next ranking; after a loss by 15 or more, 5.6. The power rating moved the same teams 1.3 and 5.7 spots. The committee does count the margin, but each point of it moved the power rating about 4 times as much (142 losses, 2014–2025).
- Expect a ranked team that loses a close game this week to drop about 3.9 spots on Tuesday, even if it played well. By the numbers, a close loss to a good team says little about how good a team is.
- A lopsided loss costs about the same in both. The committee and the power rating agree that losing by three touchdowns is bad news.
- Losing to an unranked team hurts more in the committee’s eyes (5.7 spots, against 4.1 for a loss to a ranked team).
This week’s games with ranked teams
Every top-12 loss since 2014
Two ways of reading the same loss
The committee ranks résumés as much as teams: whom a team beat, whom it lost to, and how many losses it has. The power rating is built from how well a team played, every snap adjusted for the strength of the opponent, plus CFBD’s Elo rating. A loss counts in it only as far as the team played badly.
Those two views mostly agree about a blowout. Teams ranked in the top 12 that lost by 22 or more fell 5.3 spots in the committee’s next ranking and 6.4 in the power rating (33 games). They part ways on close losses. A loss by 1 to 3 points still cost 3.7 spots with the committee but only 1.1 in the power rating (34 games), because a team that loses by a field goal to a good opponent has usually played about as well as its rating already said.
Measured per point of margin, ten extra points of defeat cost about 0.6 spots with the committee (90% interval for the per-point figure: 0.03 to 0.09) and about 2.3 in the power rating (0.18 to 0.27). So the committee is not blind to the score; it just weighs the result far more than the margin.
What it means this week
A ranked team that loses a tight game this weekend will probably fall several spots on Tuesday whatever the numbers say. If it played well, the power rating will barely move it, and that gap is one of the places the poll and the numbers disagree. A win, by contrast, moved a top-12 team up only 0.7 spots on average (456 games): most of a ranking’s movement comes from other teams losing.
Method, data, and limits
Data. The committee’s weekly rankings, game results, and play-by-play from CollegeFootballData.com, regular seasons 2014–2025 (completed seasons only). The power rating is Sparvi’s, through each week (how it’s built).
The comparison. For each pair of consecutive committee rankings (1,211 team-weeks in all), a ranked team counts if it played exactly one game in between; byes and the rare two-game weeks are left out. The fall in the committee’s ranking is the new rank minus the old; the fall in the power rating is its rank among all FBS teams after that game minus its rank before. The main sample is the 142 losses by teams ranked in the top 12, because almost none of them fell out of the 25 and the new rank is always known. Teams ranked 13th to 25th fell out 53% of the time after a loss (232 losses), so their average fall is a lower bound. Intervals come from 2,000 bootstrap resamples of the losses.
What it can’t tell you. The two scales differ: the committee ranks 25 teams and the power rating ranks every FBS team, so a spot near the top of each is comparable, but not exactly. A team’s fall also depends on what the teams around it did that week. This doesn’t say the committee is wrong: ranking résumés is a different job from rating teams, and its published selection guidelines are not modeled here. And 62 close losses and 54 lopsided ones are small samples; the intervals above show how much they could move.
Code: cfb/loss_cost/ · Data provided by CollegeFootballData.com. Sparvi is not affiliated with the NCAA, the College Football Playoff, any conference, or any school.
Model version. CFB power rating v1.0.0 (experimental: validated before the backtest harness existed; its harness port is pending).