Early Downs › Methodology
Methodology
On this page: Postgame win expectancy · Misleading Final · Projected snaps
Postgame win expectancy
Postgame win expectancy answers one question: if these two teams produced this game's stats 1,000 times, how often would each win? The stats that reflect how a team played (success rate, yards per play, field position, finishing drives and turnover chances) are held fixed; the bounces are re-rolled. It measures how this game was played, not which team is better, and it ignores the final score on purpose, so it can disagree with it. Early Downs uses it in recaps and for expected wins. This page documents exactly how it is calculated and how well it holds up.
One model, both leagues
Early Downs calculates its own postgame win expectancy for the NFL and college football with the same model, built on Bill Connelly's five factors, the ingredients he describes for his SP+ postgame win expectancy. For each team, compared with its opponent:
- Efficiency: success rate. NFL: share of scrimmage plays with positive EPA (passes, sacks, scrambles and runs; no kneels or spikes). College: CollegeFootballData.com success rate, garbage time excluded.
- Explosiveness: yards per play.
- Field position: average starting spot of each drive.
- Finishing drives: points per trip past the opponent's 40 (touchdown 7, field goal 3).
- Turnovers, luck-neutral: expected turnovers rather than actual ones. Half of all fumbles (recoveries are close to a coin flip); NFL interceptions as thrown; college interceptions estimated as 22% of interceptions plus passes broken up.
EPA per play and the score itself are deliberately left out. EPA credits touchdowns directly, so a model built on it mostly re-reads the scoreboard and comes out overconfident.
Each league's weights come from a linear fit of final margins on those five differences, using the 2022–2025 regular seasons. That gives a played-like margin in points. Win expectancy turns that margin into a probability with a logistic curve, whose slightly heavier tails fit held-out games better than a normal curve at the extremes. Even lopsided games are shown as ">99%", never 100%. The two leagues were fit separately and landed on very similar weights:
| Points of played-like margin | NFL | College |
|---|---|---|
| Success rate, per 10 percentage points | +3.22 | +4.43 |
| Yards per play, per yard | +3.13 | +3.40 |
| Expected turnovers, per turnover | +2.08 | +2.11 |
| Field position, per yard of average drive start | +0.37 | +0.43 |
| Finishing drives, per point per trip past the opponent's 40 | +2.21 | +2.33 |
Checked against games it never saw
NFL: 1,084 games; curve scale 3.55 points. Predicting each season from a model trained only on the other seasons, the played-like margin picked the actual winner 85% of the time, with a typical miss of 6.2 points.
| Win exp. | Games | Avg. | Won |
|---|---|---|---|
| 0–10% | 250 | 3% | 3% |
| 10–25% | 102 | 17% | 20% |
| 25–50% | 150 | 36% | 41% |
| 50–75% | 137 | 63% | 59% |
| 75–90% | 120 | 84% | 87% |
| 90–100% | 325 | 97% | 98% |
Held-out seasons, home team's win expectancy. A well-calibrated model wins about as often as it says.
College: 2,991 games; curve scale 4.50 points. Predicting each season from a model trained only on the other seasons, the played-like margin picked the actual winner 87% of the time, with a typical miss of 8.0 points.
| Win exp. | Games | Avg. | Won |
|---|---|---|---|
| 0–10% | 660 | 3% | 3% |
| 10–25% | 288 | 17% | 19% |
| 25–50% | 305 | 38% | 42% |
| 50–75% | 341 | 63% | 62% |
| 75–90% | 322 | 83% | 86% |
| 90–100% | 1075 | 98% | 98% |
Held-out seasons, home team's win expectancy. A well-calibrated model wins about as often as it says.
NFL this season so far (120 games):
| Win exp. | Games | Avg. | Won |
|---|---|---|---|
| 0–10% | 32 | 4% | 3% |
| 10–25% | 16 | 18% | 6% |
| 25–50% | 8 | 41% | 38% |
| 50–75% | 17 | 63% | 71% |
| 75–90% | 15 | 82% | 93% |
| 90–100% | 32 | 97% | 97% |
Updated with every data refresh. Log loss 0.276; lower is better (a coin flip scores 0.693).
College this season so far (505 FBS games):
| Win exp. | Games | Avg. | Won |
|---|---|---|---|
| 0–10% | 96 | 2% | 2% |
| 10–25% | 40 | 18% | 15% |
| 25–50% | 50 | 37% | 46% |
| 50–75% | 60 | 65% | 63% |
| 75–90% | 52 | 84% | 83% |
| 90–100% | 207 | 98% | 99% |
Log loss 0.263.
Why an NFL game can reach >99% despite parity
NFL parity is about how evenly matched teams are before kickoff. Postgame win expectancy asks something else: given how this game was actually played, how often does that team win? Teams that clearly win the five factors almost always win the game, even in the NFL. In 2022–26 games the model was not trained on, teams that played like 9–12-point winners won 95% of the time, 15–20-point winners 137 of 138, and 20+-point winners 132 of 132.
Compared with SP+ postgame win expectancy
Bill Connelly publishes his SP+ postgame win expectancy for every FBS game. Early Downs' college model never sees those numbers, but on 500 FBS games through 2026 week 9 the two agree closely: played-like margins correlate at 0.97, win expectancies differ by 6.2 percentage points on average, and they disagree about which team played better in 6.8% of games. For comparison, CollegeFootballData.com's own postgame win probability differs from SP+ by 12.5 points on average and disagrees on the better team 13.8% of the time, which is why Early Downs no longer uses it. The NFL model uses the same method; no comparable public NFL postgame number exists to check it against.
What it doesn't include
Kicking and return game beyond their effect on field position, penalties, and anything about the opponent's strength: this measures one game's performance, not team quality. The models are retrained once a season; the code is versioned with the site.
Misleading Final
A finished game is flagged when the final score and the played-like margin disagree a lot, from the winner's point of view:
- Flipped result: the winner's postgame win expectancy was 35% or lower. The loser played well enough to win about 2 times in 3 or more.
- Closer than the score: the winner won by 10+ points (NFL) or 14+ points (college) more than its played-like margin.
- Bigger than the score: the winner played like a team 10+ (NFL) or 14+ (college) points better than the final margin.
The two margin flags only count when the final and the played-like margin tell different stories: a one-score game (1–8 points), a two-score game (9–16) or a blowout (17+). A 45-point win that played like a 23-point win is a blowout either way, so it isn't flagged. College thresholds are higher because college scores swing further from how games were played. In a team's own game log, a flagged game is green when that team played better than the score says and red when the score made it look better than it played; game tiles and recaps stay neutral, since the same game is good news for one team and bad news for the other.
Projected snaps
Each matchup shows how many offensive snaps to expect in regulation (overtime removed), compared with the NFL average. It describes how a game is likely to be played. It is not a pick, and Early Downs doesn't make picks.
For each offense, using only games already played this season (pulled toward the league average early in the season):
- Volume: the offense's plays per game averaged with the plays the opposing defense faces per game.
- Neutral pace: seconds per snap when the game is competitive (win probability 20–80%, outside the last two minutes of a half), for the offense, for the opposing defense, and for the opposing offense, since both offenses share the clock.
- Script pace: the offense's pace when leading or trailing by 8+.
A linear fit on 1,087 regular-season games (2022–2025) weights those inputs. Pass rate over expected and the betting spread and total were also tested; neither improved the projection once pace was included, so they are left out.
How accurate it is. Predicting each season from a model trained only on the other seasons, the projected game total missed by 5.7 snaps on average, versus 6.0 for simply guessing the league average. Games projected faster than average did run more snaps, and slower ones fewer, but single games vary a lot: turnovers, long drives and quick scores move snap counts in ways no pregame number can see. Most games (80%) land within 9 snaps of the projection, which is the range shown on each matchup.
Pace is already reflected in betting totals: in the back-test the projection had no relationship with how games finished against the closing total. Treat it as one more data point about the game, not a signal.
College
College tempo differs far more between teams than NFL tempo does, so the simple version works well: each offense's plays per game (rushes plus pass attempts) averaged with the plays the opposing defense faces, pulled toward the FBS average early in the season and calibrated on 2,848 FBS-vs-FBS games from 2022–2025. Overtime games are left out, both from the calibration and from each team's season averages, because college box scores don't separate overtime plays.
Predicting each season from the others, the projected game total missed by 7.8 snaps on average, versus 9.0 for guessing the FBS average. Opponent tempo, drives per game and plays per drive added nothing. The betting spread and total helped slightly, but they are left out so the projection stays independent of the market. Most games (80%) land within 12 snaps of the projection.