This model estimates each team’s chance of winning as the game goes on. A trained model reads the score gap, the time left, and the game situation, and updates after every tracked event.
P(win) = calibrated score(state inputs)The inputs are Δscore (home minus away goals), t (seconds remaining), and sit (game situation: 5v5, PP, PK, EN), together with the other features the model was trained on. We don’t publish the model’s exact form or coefficients; the number shown is the calibrated output.
Every tracked NHL event — faceoff, shot, goal, penalty, stoppage — arrives with a timestamp, score state, and strength code.
At each event, the model extracts score differential, seconds remaining, game situation, and period to form the feature vector.
The trained model scores each event, then a calibration step adjusts the raw score, so the number you see is a calibrated estimate rather than the model’s raw output.
The probability updates the game view live through the games data feed, shown as the winding chart line.
Near-certain
Only a historic collapse overturns this.
Strong favorite
Firmly in control of the outcome.
Slight lean
Meaningful edge but still very live.
Coin flip
Game is effectively level at this moment.
In-game win probability
Tied-state pressure and pregame context model
Promoted · validated
Win probability release 10Win probability reflects the current game state, not a forecast of future performance. A team at 20% can absolutely win — it simply means that teams in an equivalent state historically win about 20% of the time. The model does not account for goalie matchups, injuries, or other contextual factors not encoded in the game-state inputs.
Choose a game and open Game Story to see how goals, penalties, time, and score state moved the win-probability line.