Learn · deserved to win
What is post-game win expectancy?
Post-game win expectancy asks how often a team would win a game, given how it played. Bill Connelly built it for college football from a game’s key stats, and FTN publishes an NFL version. We answer the same question by replaying each game from its own plays, 4,000 times.
The classic version
Take the stats of one game (success rate, yards per play, expected points added per play, turnovers) and fit a model of who won. Plug in a game, and the model says how often a team with those numbers wins. If a team lost a game its numbers usually win, it was unlucky.
Why it looks better than it is
Those stats contain the result. Expected points added counts touchdowns as big plays, so a team that scored more almost always “played better” by that measure. A model built on them predicts the winner of the same game very well (log loss 0.287 on our 2024–2025 holdout, where lower is closer) because it partly re-describes the score. Asked a question it can’t see the answer to, whether its numbers predict the rest of a season, it does barely better than actual wins (0.396 against 0.378).
The replay instead
We keep each team’s plays and throw away the order they came in. The plays are drawn again, by down and distance, into new drives and new games, with kicks and turnovers at the rates the teams earned rather than the ones that happened. The share of replays a team wins is what its play was worth that day, and it is calibrated: when it says 90%, the team wins about 90% of the time (chart).
The biggest heist since 2019 by that measure: the Browns beat the Ravens 29–24 (2024, Week 8) in a game they win 1% of the time.