How the number is made.
A gradient-boosted classifier over fifty-two 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
Fifty-two features, all derivable from the broadcast - seventeen game-state basics plus thirty-five hand-authored situational indicators (field-goal range, down-and-distance leverage, score-margin context). Nothing about team quality enters the model directly except a single pregame prior, blending market odds with recruiting/returning-production/roster signal - 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. influence* |
|---|---|---|
| Pregame prior | Spread + roster-blended, decays to 0 as the clock runs out | 0.309 |
| Is home offense | Boolean | 0.281 |
| Score differential | Signed, home − away | 0.137 |
| Garbage time factor | Continuous, 0-1 | 0.073 |
| Score differential ÷ offense timeouts | Continuous | 0.054 |
| Blowout margin | Boolean, margin > 24 | 0.034 |
| Score differential ÷ seconds remaining | Continuous, undamped | 0.028 |
| Seconds remaining (regulation-normalized) | Log-scaled | 0.013 |
| Time per score still needed | Continuous, seconds | 0.011 |
| One-score game | Boolean, margin ≤ 8 | 0.009 |
| Scoring-chance index | Continuous, 0-1 | 0.007 |
| Timeout advantage | Signed, offense − defense | 0.006 |
| Yards to goal | 1-99 | 0.005 |
| Defense timeouts | 0-3 | 0.005 |
| Offense timeouts | 0-3 | 0.004 |
| Rule era | 0 = pre-2021, 1 = 2021-22, 2 = 2023+ | 0.004 |
| Two-score game | Boolean, 8 < margin ≤ 16 | 0.004 |
| Down | Ordinal 1-4 | 0.004 |
| Two-minute warning | Boolean, either half | 0.003 |
| Scores needed to tie or lead | Integer, 0+ | 0.003 |
| Favorable scoring situation | Boolean, in range & down ≤ 2 | 0.002 |
| Trailing with time to spare | Boolean, trailing & ≥ 10 min | 0.001 |
| Distance | Yards, capped 30 | 0.001 |
| Out of field-goal range | Boolean, yards to goal > 50 | 0.001 |
| Hurry-up, no timeouts | Boolean | 0.001 |
| Is overtime | Boolean | 0.001 |
| Both sides low on timeouts | Boolean, both ≤ 1 | 0.000 |
| 4th down | Boolean | 0.000 |
| One-score game, final 2 min | Boolean | 0.000 |
| One-score game, final 30 sec | Boolean | 0.000 |
| Offense out of timeouts | Boolean | 0.000 |
| 4th-down field-goal decision | Boolean, 4th down & in range | 0.000 |
| Desperation trailing | Boolean, down 9+ & ≤ 5 min | 0.000 |
| Money down | Boolean, 3rd/4th & ≤ 3 | 0.000 |
| Overtime period | Ordinal, 0 if regulation | 0.000 |
| Late-half scoring urgency | Boolean | 0.000 |
| Clock-killing situation | Boolean, leading & ≤ 5 min & down ≤ 2 | 0.000 |
| Goal-to-go | Boolean, distance ≥ yards to goal | 0.000 |
| 4th & short | Boolean, 4th & ≤ 2 | 0.000 |
| 3rd down in the red zone | Boolean | 0.000 |
| Two-score game, final 5 min | Boolean | 0.000 |
| 3rd down | Boolean | 0.000 |
| Long yardage | Boolean, distance ≥ 8 | 0.000 |
| Defense out of timeouts | Boolean | 0.000 |
| In long field-goal range | Boolean, 35 < yards to goal ≤ 50 | 0.000 |
| Short yardage | Boolean, distance ≤ 2 | 0.000 |
| Two-score game, final 2 min | Boolean | 0.000 |
| Backed up in own territory | Boolean, yards to goal ≥ 90 | 0.000 |
| In red zone | Boolean, yards to goal ≤ 20 | 0.000 |
| In field-goal range | Boolean, yards to goal ≤ 35 | 0.000 |
| Plus territory | Boolean, yards to goal ≤ 50 | 0.000 |
| Has pregame prior | Boolean | 0.000 |
*Mean absolute SHAP contribution across the 2024 test season, normalized to sum to 1 across all features.
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.
04 · What it does not know
Season outlook
Each team page's projected record starts from the pregame points model, the same projection the game page shows before kickoff, for every game left in the regular season. Each game's chance of a win is the model's own. The whole league's remaining schedule is then played out 50,000 times, and the share of those seasons ending at each record is the chart. Conference championship games are left out of the record: nobody knows who plays in them until the end. The playoff odds carry the same simulated seasons on through the title games and the playoff. The points model only covers games between FBS teams, so a game against a lower-division opponent uses how often FBS teams have won those since 2014: about 97% for Power conference teams and 91% for the rest.
Results within a season aren't independent. A team the model misjudges one week tends to be misjudged the next, so each simulated season gives every team one strength surprise that carries through all its games. Its size was measured on 2018-2025, scoring each season's final records from projections made at weeks 4, 7 and 10. It turned out small, because the model re-reads every result each week: it widens the range of records a little and keeps every game's chance exactly the model's. Every game is worth one win, so the games that decide a record are the ones closest to a coin flip. Bowl eligibility is six wins, at most one against a lower-division opponent. The outlook updates on Tuesdays, after the week's games are scored, and games six weeks away are projected with what's known today.
Expected points
A second, separate model values every snap by the next score in the half: from its down, distance, field position and clock, the chance the next points are a touchdown, field goal or safety for either team, weighted by their value. First and ten at your own 25 is worth under half a point; first and goal at the opponent's 10, about five. It uses no team strength and doesn't feed the win-probability model.
Team pages judge each drive against where it started: its result (seven for a touchdown, three for a field goal, or the value it handed the other side) minus its first snap's expected points. Averaged over a season, that is points over expected per drive. Finishing does the same from a drive's first snap inside the opponent's 40, and field position is the average starting value. Those are raw; the adjusted per-drive number fits every team's offense and defense against the opponents it actually played, with home field, and reports what each would do against an average FBS opponent on a neutral field.
Fourth downs
Game pages grade every fourth down after the fact. Each option plays out from the situation before the snap. Going for it either converts or turns the ball over where it stands. A field goal is made or missed. A punt lands where punts from that spot have historically landed. The chances come from past college plays: conversion by distance and field position, field goals by kick distance. Each option's number is the win probability it leaves the team with the ball, averaged over those outcomes.
For most of the game the options are compared on expected points, which sees field position clearly. When a team trails in the fourth quarter, the comparison shifts toward win probability as the clock runs down, because only the scoreboard can say a team has to go. Gaps under one point are called close: the curves behind them are not that precise. The weakest spot is trailing by two scores late, where it still punts more than it should.
Team pages add up a season of these calls. A call that differs from the favored option gives up the gap between the two, and the season's total is win probability left on the table, in expected wins. Against the model a call can only cost, so a team is ranked per game against the rest of FBS. A field goal from a spot the model puts out of range is left ungraded: it means the play-by-play logged the wrong field position.
The same card grades every touchdown's try: kick the extra point or go for two. Since 2014, college kickers have made about 97% of extra points and offenses have converted about 45% of two-point tries. Either way the other team gets a kickoff next, so the only difference between the two choices is the points and the odds of getting them.
Before the fourth quarter a point is just a point, so the comparison is expected points, and the kick (0.97 points on average) beats going for two (0.90). Over the fourth quarter it shifts to win probability, whether the team leads or trails, since late in a game the margin a point creates can matter more than the point. That is where going for two pays: down two to tie, up one to lead by three, up five to lead by seven.
Team pages add up a season of tries too. Most are kicks the numbers agree with, and a typical team gives up less than a point of win probability on tries over a whole season, so tries are shown against the FBS median rather than ranked.
Players: wins above replacement
Team pages value players the way baseball's WAR does, adapted from the NFL version (nflWAR). Every play's expected-points value is shared among the players named in it - the passer, the target, the runner, the kicker - with the opposing defense accounted for, and a player's season is compared with what a typical backup would have produced on the same plays. About 53 points of season scoring margin make one win. Each number comes with a 90% interval from resampling the season's drives.
Only snaps while the game was in doubt count (each team between a 5% and 95% chance to win), so a backup mopping up a blowout doesn't set the bar. That also means a starter on a team that blows opponents out plays fewer counted snaps, so value per play is shown beside the total. Linemen and defenders aren't named in public play-by-play, so the offensive line, front seven and secondary are valued as units, against the season's median FBS unit rather than a backup.
Talent and roster
Roster talent is 247Sports' team talent composite, which adds up the recruiting ratings of everyone on the roster, transfers included. Recruiting averages a team's last four class scores, and returning production is the share of last season's offensive value (expected points added) from players who are back. All three come from CollegeFootballData and are ranked among that season's FBS teams. Talent ratings start in 2015.
Talent conversion compares a team's SP+ with the SP+ its talent predicts. The prediction is one curve fitted on every completed FBS season since 2015, with talent measured against that season's FBS average. Talent alone explains about half the spread in SP+, so the other half is conversion: coaching, development, scheme, luck, and the recruiting services' own rating errors. A season in progress is scored against the same curve but isn't part of it, and its SP+ still leans on preseason expectations early on.
Advanced stats
Team pages show CollegeFootballData's season advanced stats as they publish them, not numbers YardLine computes. Most are built on EPA (CFBD calls it PPA): how much a play changed the expected points of the drive, so a 4-yard run on 3rd and 3 counts for more than the same run on 3rd and 10. Success rate and explosiveness split that into how often a team stays on schedule and how big its good plays are.
Garbage time, the stretches of a lopsided game where the result is no longer in doubt, is left out. Each number is placed as a percentile among that season's FBS teams, flipped for defense so 100th is always best. The header's SP+, FPI and SRS are outside power ratings, shown as published for context.