Our xG model

Expected goals, built in-house. A gradient-boosted model reads each shot's angle to the goal, its distance, the situation and whether it's a header, and returns the probability it goes in. Trained on 535,000 shots and cross-validated, so the number is honest, and it's ours, so we can price shots on any feed we hold, not a single vendor's number.

0.781AUC

0.2704Log loss (ours)

0.2516Log loss (vendor)

52,751xG total

52,716Actual goals

Shot value by location

An open-play foot shot from each spot. Brighter = higher chance. The value collapses fast as the angle to the goal tightens, which is exactly why the model leans on angle first.

Calibration

Each dot is a decile of shots: our predicted xG against how often those shots actually scored. They sit on the diagonal, the model is well-calibrated, its numbers mean what they say.

What drives the number

Angle to goal 67%

Distance 16%

Header 11%

Open play 3%

From corner 2%

Free kick 1%

Set piece 0%

LightGBM, 5-fold cross-validated on 535,522 non-penalty shots (penalties get their 0.779 empirical rate). Vendor log loss is Understat's own xG on the same shots, for reference. Information only.