The Lab weighs several ratings of who is better. Most of them are published elsewhere and quoted here. One is computed on this site, from the same results everything else on it is built from, and this page is what it does.
One sentence
Every game says the home team won by this much. The rating is the set of team numbers that comes closest to explaining all of them at once.
Formally: choose ratings and a home-field number minimizing the
squared error of margin = rating(home) − rating(away) + home
field across every result.
Whose method this is
The method is Kenneth Massey’s least-squares rating, set out in his 1997 thesis Statistical Models Applied to the Rating of Sports Teams and used in various forms ever since, including as a component of the old BCS formula. Massey publishes his own ratings at masseyratings.com; those are his and are not reproduced here. What follows is his method applied to this site’s own results.
Three things depart from the plain version:
- Ratings are shrunk toward zero. Textbook least squares believes every result completely. A penalty on large ratings means a team only travels far from average when the evidence keeps insisting, which matters most when there is least evidence.
- Non-FBS opponents are kept, pooled into one team. The usual treatment is to drop those games. Keeping them recovers roughly 126 results a season.
- Home field is measured, not assumed. It is fitted alongside the ratings rather than set to a constant beforehand.
Each is described below, with what it is worth.
What comes out
- Ratings in points. A team rated 8 higher than another is an eight-point favorite on a neutral field. No conversion and no scaling constant — the number already is the margin it predicts.
- A measured home field. It is a free parameter of the fit rather than a constant chosen in advance, so the games decide it. Over a full season it settles near three points.
- Zero is average. Ratings are centered: positive is better than a typical FBS team, negative is worse.
How it treats a season
Every game counts the same, whenever it was played. A result from September carries the same weight in December as one from last week. Recency weighting was tested and made the rating less accurate, not more.
Ratings are pulled toward the middle. A team that has won once by 40 is not four touchdowns better than the field; it has one result. The fit is charged for large ratings, so a rating only travels far from zero when the evidence keeps insisting. The effect is large in September and fades on its own as games accumulate.
Games against FCS opposition count, together. Beating an FCS side by three says something, so those results are used, with every non-FBS opponent treated as a single pooled team. Rated individually they would contribute mostly the noise of teams that played one FBS game all year.
Blowouts are not capped. A 45-point win is entered as 45. Damping them was tested and, alongside the pull toward the middle, changed nothing.
Early in a season
For the first month, no arithmetic can separate this team is better from this team was at home. The give-away is the home field number: fitted on two weeks it reads around twelve points, four times what it should be, because it is absorbing differences the games have not yet pinned on anyone. It settles once teams have played five.
Until then the rating is not published, and last season’s is shown in its place, pulled toward average to reflect that a year has passed. The Lab labels it with the season it came from, as it does for any rating still on last year’s numbers.
How accurate it is
Measured the way it is used: rate the weeks already played, predict the week ahead, repeat for every week of a finished season. Against the same games, with and without the two adjustments above:
| season | without | with | improvement | 95% interval | weeks better |
|---|---|---|---|---|---|
| 2024 | 13.96 | 12.60 | 1.24 | 0.76 to 1.72 | 10 of 12 |
| 2025 | 13.70 | 12.43 | 1.33 | 0.77 to 1.91 | 9 of 12 |
Average miss per game, in points — lower is better. The interval is a paired bootstrap over every prediction.
Roughly a point and a quarter of accuracy per game, holding across two independent seasons and winning 10 of 12 weeks in 2024 and 9 of 12 in 2025. Against the published systems on the same out-of-sample games it is ahead of some and level with the best of them.