How accurate is xG?
An xG model is a set of probabilities. A good one is right on average: shots it calls 0.20 go in about one time in five. Here is how that is tested, and what we know.
Calibration: the key test
Group all shots by their xG and count how many went in. If shots valued around 0.20 were scored 20% of the time, around 0.50 half the time, the model is well calibrated.
Scores analysts use
Brier score
The average squared gap between each shot's xG and what happened (1 for a goal, 0 for none). Lower is better; good xG models score about 0.078 to 0.081.
AUROC
How well a model ranks goals above non-goals. Researchers warn it can flatter a poorly calibrated model, so calibration and Brier come first.
Data needed
A simple distance-and-angle model stops improving after about 6,000 shots; richer models kept improving after five seasons of data (KU Leuven).
A Brier score, worked through
What we know, and what we don't
- No commercial provider publishes its Brier score, log loss or calibration.
- Between 79% and 93% of team seasons fall within a 95% band around their xG, depending on the model.
- Over seasons, xG predicts future results better than goals do, though its lead has narrowed recently.
- For a single shot, no model is "right": the shot goes in or not. Accuracy is a property of many shots.
Questions
Is xG accurate?
Good models are well calibrated over many shots: shots valued at 0.20 go in about 20% of the time. No provider publishes its exact figures.
What is a Brier score?
The average squared difference between predicted probabilities and outcomes. For xG, good models score about 0.08; lower is better.
Sources
- KU Leuven DTAI: How data availability affects learning good xG models · 2020
- Hudl StatsBomb: Expected goals (xG) explained
- American Soccer Analysis: is xG the best predictor of future results? · July 2022
Checked 2026-10-28. Methods change: every figure here is dated to its source.
xG explains football; it is not a betting tip. 18+ · Please gamble responsibly.