Guide · 6 min read
How computer rankings work
Polls rank teams by vote. Computer rankings rank them by formula, using scores. This guide covers how the formula works and why the two disagree.
The problem every ranking has to solve
College football has well over 600 teams across four divisions, and each plays only about a dozen games. Most pairs of teams never meet. To compare two teams that did not play each other, you have to go through teams they both played, or teams those teams played, and so on. A ranking is a way of chaining all of that evidence together.
Human voters do this chaining in their heads, informally, with preseason expectations and reputations doing a lot of the work. Computers do it explicitly: every game becomes an equation, and the ratings are the numbers that best satisfy all of the equations at once.
Turning games into equations
The basic idea: if Team A beat Team B by 10 points, that is evidence that A is about 10 points better than B. Write that down for every game of the season and you get thousands of equations with one unknown per team. They cannot all be satisfied exactly, because football is noisy, so the formula finds the ratings that fit them best overall. This is the idea behind least-squares ratings, which Kenneth Massey popularized for college football in the 1990s and which sports-reference sites publish as the Simple Rating System.
Schedule strength is built in. Your rating is defined relative to your opponents’ ratings, which are defined relative to their opponents’, all the way through the season. Beat a strong team and your rating rises more than if you beat a weak one, without anyone having to decide in advance who is strong.
The adjustments that matter
- Home field. Home teams win more often, so a home win is slightly less impressive than the same win on the road.
- Margin of victory. A 40-point win says more than a 3-point win, but not thirteen times more. Good formulas give blowouts diminishing returns so running up the score buys little.
- Winning itself. Many formulas add a bonus for the win, so a one-point win is clearly better than a one-point loss.
- Recency. Teams change over a season. Weighting recent games a little more lets a rating follow a team that is improving or fading.
- Small samples. After one game, almost nothing is known. Regularization pulls thin résumés toward average until the evidence piles up.
Why computers and polls disagree
Polls tend to rank by record first and reputation second, and they are slow to move a team that starts high. Computers ignore reputation entirely and care about margin and opponents. So a 6-2 team that lost close games to two top-five opponents can sit above a 7-0 team that has beaten nobody, and the 6-2 team may well be better. Whether it is more deserving is a different question, which is why we also publish a résumé ranking.
How the Power Board does it
The Power Board is a least-squares rating over all 2,436 completed 2026 games, with every division solved together on one scale. It adjusts for home field, softens blowouts, rewards wins, weights recent weeks a little more, and discounts games between divisions far apart. No polls, recruiting rankings, or preseason expectations go in. The full formula and settings are on the methodology page.