Every FBS game of the week, forecast from how each team has played against the opponents it has faced, with the luck of its results taken out, and from where it stood last season. College schedules differ enormously, so every team is rated against the schedule it has played. No betting information goes in. Tested on every FBS game from 2021 to 2025, this forecast picked 74.4% of winners.
Updated Nov 15, 6:07 pm ET. Each forecast is frozen at kickoff and scored when the game is played. How it has done
Why the curves are so wide: luck. Even with both teams’ strength known exactly, one game’s bounces, kicks, calls and sequencing move the final margin by about ±13.4 points (one standard deviation, measured from every play). These forecasts carry ±16.0, so about 70% of the width is that luck; the rest is what nobody knows yet about the teams. That is why a 7-point favorite still loses about 3 times in 10.
This season so far: 75.4% of 751 FBS games picked right (log loss 0.480). The 331 games played before the forecasts went live were forecast afterwards from what was known before each kickoff.
Every FBS game from 2021 to 2025, each season forecast by a model fitted on the others (4,546 games):
Forecast
Picks right
Log loss
Brier
Ours, every FBS game
74.4%
0.495
0.167
Log loss and Brier score the probability, not just the pick: lower is better; a coin flip scores 0.693 and 0.250. College picks run higher than the NFL’s because so many games are mismatches.
Weeks
Picks right
Log loss
Weeks 1–2
83.1%
0.351
Weeks 3–5
77.6%
0.465
Weeks 6–9
70.7%
0.564
Week 10 on
72.0%
0.528
Bowls and playoff
57.5%
0.669
Its win chances and ranges are honest: the P90 range held 91% of final margins, the P95 range 95%. Forecasts made further ahead are less sure: 74.4% picked right the week before, 72.4% four weeks ahead, 70.9% eight weeks ahead, with the curve widening from ±16.0 to ±17.1 points.
What goes in, and what doesn’t
This season, against its opponents. Each team’s margins (capped at 50 points, so a blowout of a weak team doesn’t count for too much) and its results adjusted for luck (every kick, fumble bounce, 3rd and 4th down and judgment call re-rolled), each rated against the teams it has played.
Last season. Each team starts from where it finished, pulled a fifth of the way back toward average.
Home field. None at neutral sites.
Not in it: who starts at quarterback and who is hurt, which college data doesn’t carry the way the NFL’s does, and the betting line.