How our expected goals (xG) model works

Each unblocked attempt (a goal, a shot on goal or a missed shot) gets a probability between 0 and 1. A team's or player's xG is the sum over their shots.

• Four models: even strength, power play and shorthanded are gradient-boosted. Empty net is a small logistic model, because there are only about 1,000 empty-net attempts a season. Power-play shots behave differently (the extra attacker changes the geometry), so they get their own model.
• Blocked shots get no xG. The feed records where the block happened, not where the shot came from. NST, MoneyPuck and Evolving-Hockey follow the same convention.
• Penalty shots use a constant 0.30: 110 goals on 373 attempts over nine seasons, shrunk toward the long-run rate of about 0.32. The shootout is excluded.
• Shots with missing coordinates get a constant 0.03 and are flagged as imputed. Only 7 of 1.01 million attempts lacked coordinates.

We train in Python (LightGBM) on every unblocked attempt from 2017-18 to 2025-26: 1,014,425 attempts over 11,739 games. Hyperparameters are tuned with five-fold cross-validation grouped by game, so shots from one game never sit on both sides of a split. The second half of 2025-26, every game from January 25 on including the playoffs, is held out. It is never used for tuning, and all the numbers below come from it. The shipped model is then refit on every season.

Older seasons need one more step. A model centred on recent hockey overrates 2017-18 and 2018-19 shots by 13-16% when it rescores them. So for each completed season and strength state we store one intercept that makes total xG match total goals for that season. For example, even strength in 2017-18 is shifted down by 0.25 on the log-odds scale. The current season gets no adjustment. The intercepts fix the totals but not the ranking of shots: before 2023-24 the model separates goals from non-goals less well (AUC about 0.70-0.74, against 0.80 or more since), because some of its signals only exist in the newer data.

What matters most: distance (26% of the even-strength model's gain), lateral position (15%), angle, puck speed from the previous event, rebound timing, time since the previous event and tip-ins.

Features are computed once, in TypeScript, from the archived play-by-play. The training set is exported by that same code, so the trainer never reimplements a feature. A parity test requires the site's scorer to reproduce the trainer's probabilities: on 2,000 sampled shots the largest difference was 3 × 10⁻¹⁶. That is how we avoid the common bug where training and serving compute features differently.

Every scored game stores the model version that scored it (currently xg-v2). We retrain once a year in August. During the season the weights stay frozen; if goals and xG drift apart by more than 4% after 300 games, we recalibrate the intercept only and bump the patch number. A new version rescores every season, and the previous one is kept for rollback.