How the College Football Ratings Work
Methods, validation and limits behind the Pascal Rating for college football, which ranks FBS teams in the Pascal Power Rankings and drives our published college football forecasts.
Published by Pascal Institute. Edited by Christopher Merry, Chief Decision Scientist.
- What the rating measures
- Scoring every play
- Opponent adjustment
- Last season as a prior
- Combining and scaling
- Weekly expectations
- Validation
- Interpretive limits
- Sources and data
What the rating measures
The rating estimates how strong an FBS team is right before kickoff. It answers a narrow question: by how much would this team beat an average FBS team on a neutral field? It is a forecast of strength, not a judgment of who has earned the best record.
The rating uses only two kinds of evidence: points scored and allowed, and play-by-play efficiency. It does not use polls, betting lines, recruiting rankings, transfers, injuries, quarterback changes or win-loss record. That independence is deliberate. It makes the rating a second opinion rather than an echo of the market or the polls.
Scoring every play
Every play receives an expected-points value from the down, distance, field position and clock, using our own 13-term model fitted on the 2020–2022 seasons. Expected points added (EPA) for a play is the change in expected points it produces. We do not use a data provider's EPA, because those models include pregame point spreads. Every eligible play counts, including plays after a game is decided.
Opponent adjustment
Three components are computed for each team: scoring margin, offensive EPA per play and defensive EPA per play. Each is adjusted for the strength of the opponents faced through an iterative fixed-point method, so a 14-point win over a strong team counts for more than a 14-point win over a weak one. All three are centered so the average FBS team is zero.
Last season as a prior
Each team begins the season with its final values from the previous season. That prior is worth two games of evidence and fades as games are played. A game against an FCS opponent counts as 0.35 of a game.
component = (Σ weight × adjusted game value + 2 × last season's value) / (Σ weight + 2)
Combining and scaling
Margin and efficiency are combined with fixed weights fitted on 2020–2022 to predict future point differential, then placed on a public scale where 100 is an average FBS team.
| Step | Definition |
|---|---|
| Latent strength | latent = 0.5230 × margin + 18.149 × (offense + defense) |
| Public rating | rating = 100 + 10 × latent / 11.263 |
| Reading the scale | Ten rating points equal about 11.3 points of margin on a neutral field. The mapping is linear and uncapped. |
| Timing | Ratings are frozen for each week before that week's games are applied. Only completed games count. |
| Exclusions | Home field and conference have no effect on a team's rating. |
Weekly expectations
Each edition turns the ratings into an expected margin for every game involving a Top 25 team. The rating itself gives home field no value, so the weekly expectations add one. From Week 7 of 2026, that edge is 2.7 points: refit on 2020–2024 games with the rating scale held fixed, it measured 2.712 points, and it improved the win probabilities on 2025 games the refit had not seen. Week 7 was re-issued with it before any of its games kicked off. Earlier editions used 1.6 points, rounded from a controlled development estimate of 1.559, and keep the expectations they were published with. Neutral-site games get none.
| Quantity | Definition |
|---|---|
| Expected margin | (home rating − away rating) × 1.126 + 2.7, from the home team's view (1.6 in editions through Week 6 of 2026) |
| Favorite | The team with the positive expected margin after home field |
| Win probability | The final margin treated as normal around the expected margin, with a standard deviation of 15.9 points, the 2024–2025 holdout error. A 3-point favorite wins about 57% of the time, and a 14-point favorite about 81%. |
| Movement | Rank and rating change from the last build before the finished week kicked off to the first build after all of its games |
| Scorecard | Each edition locks its expectations when it is published. The next edition reports how many favorites won against the number the probabilities implied, and the average distance between final and expected margins. |
Validation
Every design choice was made on the 2020–2023 seasons, and the frozen method was then tested once on 1,606 games from 2024 and 2025 that it had never seen.
| Check | Result |
|---|---|
| Holdout accuracy, 2024–2025 | Pregame rating gap predicted final margin with an RMSE of 15.9 points and a correlation of 0.62 across 1,606 games |
| Opponent adjustment, 2023 validation | Reduced prediction error from 17.5 to 16.7 points |
| Signal comparison, 2023 validation | Margin plus efficiency beat margin alone, efficiency alone and an Elo baseline |
| Rejected variants | A margin cap, garbage-time filters and an explicit home-field term inside the rating did not help, so the simpler version was kept |
| Prior | Starting from last season beat starting from zero |
| Home edge refit, 2026 | With the rating scale fixed, 3,481 games from 2020–2024 put home field at 2.712 points. On 805 games from 2025, the win-probability log loss improved from 0.5363 with 1.6 points to 0.5331 with 2.7, and the final margin missed the expected one by 15.6 points (RMSE) |
Interpretive limits
The top of the list is rarely a clear verdict. Teams near the top are often separated by one or two rating points, a few points of expected margin, while the typical prediction error is 15.9 points. A ranking difference that small is close to a tie.
Blowouts count in full. Margin is uncapped and plays after a game is decided are kept. Caps did not help on average, but a team that keeps its starters in longer still gains.
Last season counts at full strength early. The prior is not regressed toward average and does not know about coaching changes, quarterback turnover or the transfer portal. In September, a team's rating and its opponents' ratings lean heavily on the previous year.
It is a forecast, not a résumé. An unbeaten team can fall after a narrow win over a weak opponent, and a team can rise after a close loss to a strong one. Polls and committees judge wins and losses, and the rating will often disagree with them.
FCS games are handled unevenly. An FCS game counts at reduced weight for margin and offense, but tells the model nothing about an FBS team's defense.
The inputs can change. Play-by-play data come from an upstream source that is occasionally revised, which can move a rating without a game being played. Each edition uses the stored ratings as published that day.
The error estimate was re-checked, not refit. After the holdout figures were produced, we found that historical snapshots labeled bowl games as week 1. A 2026 re-check that matches games by identifier rather than by week measured a 2025 error of 15.6 points, close to the 15.9-point holdout figure, so the expectations keep 15.9.
Accuracy is moderate. A correlation of 0.62 means the rating explains about 39% of the variation in future margins. It is a disciplined baseline, not an oracle.
Live ratings
Every team's current rating, updated as games finish, is on the College Football Team Ratings page.
Track record
See how the ratings have been scored in public, season by season and in retrospective studies, in the Track Record.
Sources and data
- AnalytIQ, College Football Rating v1 specification (methodology cfb-rating-v1, frozen): the definition, prior, FCS weight and public scale.
- AnalytIQ, College Football Sprint 2 validation report: development, validation and holdout results, including the 1.559-point home effect and the 15.914-point holdout RMSE.
- AnalytIQ, CFB home-edge calibration (2026): the 2.712-point home edge refit on 2020–2024 games with the rating scale fixed, and its test on 2025 games.
- Weekly ratings, records and schedules: stored AnalytIQ rating builds and game catalog, exported with read-only queries. Each edition's data file and the queries are kept with the published pages.