International Edition
AnalysisEditorial analysis

Expected goals measures the chance, not the finish

xG is a model output, not a fact about a match. What it does well, what it cannot see, and why post-shot models exist for goalkeepers.

By Meridian XI Data Desk · Published 9 Aug 2026, 14:00 UTC · 4 min read

Expected goals is a probability estimate. A model is trained on a large historical set of shots, each labelled with features the data provider recorded — distance, angle, body part, pattern of play, sometimes defender positions — and it outputs, for a new shot, the share of historically similar attempts that were scored.

What follows from that definition

Three things follow immediately. First, xG describes the chance created, not the execution: a 0.08 shot that goes in was still an 0.08 chance. Second, the model only knows what was recorded, so anything absent from the feature set — the goalkeeper's starting position, a defender's last-second block, a wet ball — is invisible to it. Third, because providers train on different data with different features, two published xG figures for the same match are not necessarily in conflict; they are different models.

Sample size is the real constraint

A single match's xG total is a weak signal. Over a season it becomes informative, because the differences between finishing luck and chance quality wash out at scale. This is why "underperforming xG" is a meaningful observation across 30 matches and close to meaningless across three. Anyone quoting a five-match xG gap as evidence of a striker's decline is quoting noise.

Post-shot xG and goalkeepers

For goalkeepers, plain xG is the wrong tool, because it prices the chance before the shot exists. Post-shot xG conditions on where the shot actually went — its placement and trajectory — and therefore estimates how likely that specific attempt was to beat a goalkeeper. Comparing goals conceded against post-shot xG is the standard way of asking whether a goalkeeper saved more than the shots he faced implied.

Even then, the metric ignores most of the job. Positioning that deters a shot, a claimed cross, a decision to sweep or hold, and everything in the build-up phase are outside the model. A goalkeeper assessment that consists only of a post-shot xG differential is not an assessment; it is one column of one table.

Our rule

Meridian XI names the provider or model behind any figure we publish, states the sample it covers, and does not present a modelled value as a match fact.

Sources

  • Editorial analysisMeridian XI Data Desk methodological note

Related articles