Win probability model

A win probability model estimates, at any point during a game, the likelihood that a given team will go on to win, based on the score, time remaining, and sport-specific situational factors such as possession, down and distance in football or base-runner state in baseball. The output is a probability that updates play by play or possession by possession as the game unfolds.

Models are typically trained on large samples of historical play-by-play data using logistic regression or gradient-boosted trees, learning the empirical relationship between a game state and the eventual outcome. The change in win probability caused by a single play, win probability added, is used to rank the most impactful plays and players in a game, distinct from expected points added, which values a play in points rather than win probability.

Win probability curves are a staple of broadcast graphics and drive fourth-down, two-point-conversion and pulling-the-goalie decisions across sports, and they relate closely to game-level models like pythagorean expectation and team-strength ratings such as elo rating. The main pitfall is treating a model's output as a precise, guaranteed number: models are less reliable in rare or extreme game states underrepresented in training data, and a stated probability is an estimate, not a certainty, however confidently it is displayed.