How it works

One number: the chance to win

Everything reduces to a team's win probability from a given game state: score, time, down, distance, field position, timeouts and which team gets the second-half kickoff, plus how strong the two teams are. A model trained on 1999–2025 NFL play-by-play answers that question for any moment. It is checked on seasons it never saw: of all the moments it calls 30%, about 30% end in a win (calibration error 0.006 across 1,029,188 held-out plays).

Judging a call

For a fourth down or a try, we list what can happen under each option — a conversion, a stop, a made or missed kick — and how likely each is, then average the win probability of what follows. The option with the highest average is the model's pick. The verdict never looks at what happened next: good call, bad luck is a real outcome.

• Good call — the coach picked the best option.
• Shoulda — another option was clearly better.
• Coulda — a toss-up: the options are within about 1.5 points, or 20 copies of the model trained on resampled seasons don't agree. Many famous calls land here, and we say so.

The breakevenis the success rate at which two options tie — “going for two is right if you convert at least 43% of the time” — the clearest way to explain a call.

Rewriting a play

Flip a kick, an interception, a fumble or a fourth-down stop and the rules engine produces the state that would have followed. The swing is the change in win probability right there. Then Play it out: a simulator plays the rest of the game thousands of times, drawing real plays from similar situations, tilted toward the stronger team, under the current kickoff and overtime rules. Adding the points to the final score is the wrong answer — the clock, the play calls and the other team's urgency would all have changed — so we don't.

Flags

For a flag we measure impact, not correctness: the state with the penalty enforced against the state without it. Whether a call was right comes only from official sources. Officials are never named.

What we never claim

That a coach or an official cost a team the game; certainty (chances are whole percentages, with a toss-up band); anything about real players inside simulated endings — those are told at the team level.

Data and models

Play-by-play comes from nflverse (CC BY 4.0). The models are ours, built on the structure of the open nflfastR and nfl4th projects: a win-probability model with 20 bootstrap copies for ranges, a fourth-down conversion model, a two-point model, a field-goal model, and outcome tables for punts and kickoffs. On the 2025 season, held out from training, the win-probability model's log loss was 0.4754 over 40,371 plays. Model version v1, trained 2026-10-03.

Team strength enters as a single internal input that is never shown. There is no betting content here and never will be.