Zero preseason assumptions. Every team starts at a rating of 0 with the same uncertainty, so the preseason view is a dead tie. Nothing from last year, no polls, no recruiting rankings. The only way to move is to play games.
One rating per team, in points. A rating of +20 means that team's body of work is about 20 points better than the average FBS team's on a neutral field. Ratings are solved jointly, so beating a +15 team is worth far more than beating a −15 team, and your rating depends on how your opponents did against everyone else. Strength of schedule falls out automatically.
Each game adds two pieces of evidence.
- The result, weighted heaviest. A logistic (Bradley–Terry) term penalises the model by how improbable it makes the actual winner. Beating a team you were supposed to beat moves you very little; beating a team rated well above you moves you a lot; losing to a team rated well below you costs a lot.
- The margin, soft-capped. The scoring margin runs through a saturating curve so blowouts have diminishing returns and running up the score on overmatched opponents is mostly ignored.
Shrinkage toward zero. Every rating pays a quadratic penalty for distance from 0. That keeps undefeated teams finite in September, makes a 5-0 team with a weak schedule rank below a 5-0 team with a strong one, and treats every team identically. It is the only prior, and it is the same for everyone.
Every team is rated, not just FBS. FCS, D-II and D-III teams are entities in the same solve, rated from their own games (including games against each other), so a loss to a strong FCS team costs less than a loss to a weak one. Each division has its own learned average that its teams shrink toward, with no hand-set numbers; FBS ratings are shown relative to the FBS average.
The ± is real uncertainty. The solver's curvature at the optimum gives each rating a standard error. It is about ±10 points after week 1 and ±3 by the end of the regular season. "Likely rank" in a team's detail is the range of ranks consistent with that uncertainty, which is why September rankings should be read as a wide-open race.
Home field is a fixed 2.5 points. Estimating it early in the season is badly confounded because the stronger team is almost always the one at home.
What we know now vs. what we knew then. Each week on the timeline is re-solved using only games through that week, so it is an honest replay. Within a week, every earlier result is valued by everything its opponent has done through that week.
Expected win in a team's game log is retrospective: given everyone's ratings through the selected week, how likely was that result? Low percentages on wins are quality wins; high percentages on losses are bad losses.