How the number is made.
A gradient-boosted classifier over fourteen game-state features, fit on FBS play-by-play from 2014-2022 and calibrated so a stated 70% actually wins about 70% of the time. No power ratings, no live market lines, no human adjustment.
01 · Inputs
Fourteen features, all derivable from the broadcast. Nothing about team quality enters the model directly except a single pregame market-implied prior - a 2-win team down 3 in the fourth gets close to the same in-game number as a playoff team in the same spot.
| Feature | Encoding | Rel. gain* |
|---|---|---|
| Score differential | Signed, home − away | 0.290 |
| Seconds remaining (regulation-normalized) | Log-scaled | 0.190 |
| Pregame prior | Spread-implied win probability | 0.140 |
| Yards to goal | 1-99 | 0.100 |
| Garbage time factor | Continuous, 0-1 | 0.080 |
| Down | Ordinal 1-4 | 0.060 |
| Distance | Yards, capped 30 | 0.050 |
| Is home offense | Boolean | 0.030 |
| Is overtime | Boolean | 0.020 |
| Overtime period | Ordinal, 0 if regulation | 0.010 |
| Offense timeouts | 0-3 | 0.010 |
| Defense timeouts | 0-3 | 0.010 |
| Rule era | 0 = pre-2021, 1 = 2021-22, 2 = 2023+ | 0.005 |
| Has pregame prior | Boolean | 0.005 |
*Illustrative - no per-feature gain export has been published yet.
02 · Training and validation
Fit on 2014-2022, validated on 2023, held out on 2024 - split by season rather than at random, so no play from a game the model scores was ever in its training data. Overtime plays are included in the same unified model rather than a separate one, to avoid a probability discontinuity at the start of overtime; the model reports lower confidence there rather than switching curves.
03 · Calibration
Predicted probability against observed win rate, 2024 test season, decile buckets. On the diagonal means honest.