Expected goals (xG): how the metric is built

Expected goals is the metric that pulled football analytics into the mainstream, and it is still the one most often quoted and least often explained. At its core it is simple: give every shot a probability of becoming a goal, then add the probabilities together. What makes it useful, and what makes it arguable, is how that probability is produced.

This guide takes the xG model apart. It looks at the inputs a model learns from, at why two providers disagree, and at the reading habits that keep the metric honest. We describe mechanics only — the model is the subject, not any particular player or match.

From one shot to one probability

A model is trained on a large library of past shots. For each of them the outcome is known: it either became a goal or it did not. The model looks for the features that separate the goals from the misses and learns how strongly each feature pushes the probability up or down. A new shot is then scored by the weights the model learned.

Those weights are not opinions; they are the patterns in the data the model was trained on. But they are also not universal truths. Change the training set, change the features, and the same shot gets a slightly different value. That is the honest limit of xG, and it is why the definition travels with the number.

The features that shape a chance

Distance to goal is the strongest single input. A shot from six metres and a shot from thirty metres are not the same event, and the model reflects that steeply. Angle matters too, because a wide position shrinks the target the shooter can see.

Around those two core inputs, models layer more context. The body part used, the type of assist that preceded the shot, whether the shot came from a set piece or open play, and how many defenders stood between the ball and the goal all adjust the value. Each additional feature makes the number finer and the model harder to compare with a rival.

Why two sites disagree

If you ever see two xG totals for the same match and they differ, nothing is broken. The providers defined the question slightly differently. One may include the shot before the shot; another may model defensive pressure; a third may treat penalties in a way the others do not.

For the reader the practical rule is simple: compare xG within one provider across matches, not across providers for the same match. The metric is consistent with itself, and that consistency is what makes trend-reading possible. Mixing sources produces noise, not insight.

Reading xG across a run of games

A single match of xG is a small sample and it swings widely. The metric becomes far more informative when you read it over a run of games, where a team that consistently creates high-value chances separates itself from one that scored from two long shots in a single afternoon.

The classic use is comparing xG with actual goals. A persistent gap in either direction is a signal worth investigating — clinical finishing, a hot goalkeeper, or simply variance evening out. The gap is a question, not an answer, and that is exactly what a good metric should produce.

Common xG misreadings to avoid

Most arguments about expected goals come from a handful of recurring mistakes. Naming them makes them easier to catch in your own reasoning.

• Treating xG as a prediction of the final score.
• Assuming a higher xG automatically means a team deserved to win.
• Comparing xG values produced by different providers.
• Reading a single match as if it were a season-long trend.
• Forgetting that penalties are often handled separately from open play.
• Quoting xG without saying which model produced it.

How shot features push the probability

A qualitative picture of how the main inputs influence an xG value. The bars show relative influence in a typical model, not any specific match.

What xG includes and what it leaves out

A model is defined as much by its omissions as by its inputs. This table separates the two sides.

Sample size decides how much a metric deserves to be believed. A single match is noise, a handful of matches is a hint, and a full season is the point at which patterns start to separate from luck.

Game state distorts almost every rate metric. A team protecting a lead presses less, concedes more territory and posts numbers that describe its situation as much as its style.

It is worth separating process metrics from outcome metrics. Process describes how a team played; outcome describes what the match produced. Both matter, and confusing one for the other produces false lessons.

A question, not a verdict

Expected goals is best understood as a well-defined question about chance quality. It tells you how good the chances were, in the terms of the model that computed them. Everything beyond that — whether the team was good, whether the result was fair — is a separate argument you have to make yourself.