Expected Goals (xG) measures the probability of a shot becoming a goal, based on where and how it was taken. We built our own model because the EIHL's built-in xG sometimes over- or underestimates shot quality.
What is xG?
Every shot has some probability of going in. A shot from right in front of the net might have a 25% chance; a speculative blue-line shot, 2%. Sum the per-shot xG and you get Expected Goals — a quality measure of a team's chances regardless of whether the puck went in.
Why we built our own
On the full 2025 season as a holdout test set (261 games), 78.5% of games were won by the team with higher Tracker xG, vs 72.0% using the league's xG.
METRIC
EIHL
TRACKER
Log Loss
0.1987
0.1929
Brier Score
0.0521
0.0513
ROC AUC
0.7749
0.7936
Game-Level MAE
1.400
1.230
Winner Prediction
72.0%
78.5%
Winner (regulation)
76.5%
81.6%
Winner (OT/SO)
58.5%
69.2%
What goes into it
26 features in three tiers, processed by an XGBoost gradient-boosted tree model. A goalie quality adjustment is applied on top.
WHERE THE SHOT WAS TAKEN
· Distance to goal
· Angle to goal
· Off-centre distance
· Behind goal line
· In the slot
· Inner slot
WHAT WAS HAPPENING
· Score differential
· Period & time
· Rebound (within 3s)
· Rebound off save
· Shot pressure (30/60s)
· Rush / transition
· Home/away
· Possible empty net
WHO TOOK THE SHOT
· Power-play context
· Shooter season SH%
· Rolling 10-game form
· PP specialist ratio
· Opponent defensive quality
· Opposing goalie quality
Technical details
Trained on 116,882 shots from 2021–2024 seasons, tested on a held-out 2025 season (26,222 shots). Hyperparameters tuned via randomised search with time-series cross-validation. A goalie adjustment layer scales base xG by the opposing goalie's SV% relative to league average — shots against strong goalies get a slight reduction.