Models
Accuracy and log-loss by part of the season, one game-by-game back-test for every model · generated 2026-10-19T03:49:47.993409+00:00
One back-test for every model, the same one the Methodology page reports: 2014–2025 regular seasons (2020 excluded), completed FBS-vs-FBS games from week 2. Each week's games are predicted from ratings built on earlier weeks only, and anything a model fits -- the two-layer settings and home field, the classic and Elo win-probability curves -- is fitted on the other seasons. The two-layer model is graded on its power rating's own win probabilities (its predictive layer; the published ranking is WAB, built on top of it). Log-loss: lower is better.
| Model | weeks 2-6 acc · log-loss | weeks 7+ acc · log-loss | all weeks acc · log-loss |
|---|---|---|---|
| Two-layer | 69.3% · 0.5802 | 71.5% · 0.5472 | 70.6% · 0.5603 |
| Two-layer (v1) | 68.5% · 0.5928 | 71.8% · 0.5472 | 70.5% · 0.5653 |
| Resume (classic) | 64.4% · 0.6319 | 66.4% · 0.6106 | 65.6% · 0.6191 |
| Elo | 67.3% · 0.6052 | 70.7% · 0.5622 | 69.4% · 0.5793 |
Same 7,198-game set for every model in the all-weeks column.
Two-layer
Two-layer: the model behind every other page on this site. Power is a ridge regression on capped scoring margins (every team starts at average, no preseason input); the ranking is wins above bubble built on top of it. Graded here on power's own win probabilities.
Two-layer (v1)
Two-layer (v1): the same model as first published on 2026-10-18, before that day's methodology fixes -- home field fitted each week rather than held at the conference-game estimate. Kept as the prior model, the way Classic is.
Resume (classic)
Resume (classic): the site's model until October 2026, still shown on the Classic page. Rank-based resume scoring with no preseason bias and no polls -- a win or loss is valued off the opponent's current rank, scaled by margin of victory, shrunk for small samples, and iterated to a fixed point every week.
Elo
Elo: a standard margin-of-victory-weighted Elo rating, tracked purely game to game from a flat 1500 starting point for every FBS team, with no resume shape, strength-of-schedule term, or shrinkage. Included as the field's textbook baseline.