Expected Goals (xG) Explained: What the Number Measures and What It Cannot

Expected goals (xG) is the probability that a shot will be scored, estimated from how often similar shots were scored in the past. Each shot gets a value between 0 and 1; add them up and you get a team's or a player's xG for a match or a season — the number of goals those chances would produce, on average, for an average finisher. In soccer (association football), where goals are rare, that makes xG a way to describe how good a team's chances were without waiting for the ball to go in.

What it is not is a forecast of the result. Opta's own explainer puts it directly: xG "is only measuring chance quality and not the expected outcome of the game".

From one shot to a number

An xG model is trained on a large set of past shots, each labelled goal or no goal, and learns how the circumstances of a shot change the chance of scoring. In a 2023 PLOS ONE study, Mead, O'Hare and McMenemy list the usual approaches — logistic regression, gradient boosting, neural networks, support vector machines and tree-based classifiers — and identify the predominant inputs as distance, angle and shot type. Their own dataset, a full season of top-five-league event data, had roughly one goal for every tenth shot, which is the base rate every model starts from.

Commercial models add more context. Opta describes a gradient-boosting model ("XGBoost") trained on nearly one million shots from 40 competitions between 2018-19 and 2021-22, weighing more than 20 variables, among them:

• distance and angle to goal;
• the goalkeeper's position;
• how clear the shooter's view of the goal is, given other players' positions;
• pressure from defenders;
• shot type — which foot, a header, a volley, a one-on-one;
• pattern of play — open play, fast break, free kick, corner, throw-in;
• the previous action, such as a through ball or a cross.

Penalties are handled separately: Opta gives every penalty a constant 0.79 xG, reflecting historical conversion.

What the "formula" looks like

There is no single xG formula, because each provider trains its own model and few publish the fitted weights. The simplest common form, logistic regression, looks like this:

P(goal) = 1 / (1 + e^−(b₀ + b₁·distance + b₂·angle + b₃·header + …))

The model learns the weights b₀, b₁ and so on from historical shots; the logistic function squeezes the total into a probability between 0 and 1. A tree-based model such as gradient boosting replaces the weighted sum with many small decision rules, but the output means the same thing: the share of similar shots that went in.

A worked example: adding up a match

Assume two teams take the following shots. The values are invented for illustration, except the penalty, which uses Opta's constant.

Team A created about twice Team B's xG. Now treat every shot as an independent event with the probability shown — a simplification, since shots in a match are not truly independent. On those assumptions:

• Team A fails to score at all about 9% of the time (0.21 × 0.65 × 0.70 × 0.97³ ≈ 0.087);
• Team B scores at least once about 66% of the time (1 − 0.40 × 0.96⁴ ≈ 0.660);
• combining the two, Team A wins about 60% of the time, draws about 30% and loses about 10%.

So the side with twice the chances fails to win roughly four times in ten. That is not a flaw in xG; it is what probabilities do over a handful of events. Understanding probability in sports simulations walks through the same independence logic in a game setting.

Expected goals against actual goals

The gap between a player's goals and their xG is a popular way to talk about finishing. It is also easy to over-read.

Davis and Robberechts, in a paper on finishing ability, argue three things. The gap between actual and expected goals is a weak measure on its own because shot outcomes vary a great deal and samples are small. Sustained overperformance requires both a high volume of shots and exceptional finishing. And standard models carry a persistent bias that makes actual and expected goals look closer for excellent finishers than they really are — so the best finishers can be underrated rather than flattered. Their proposed fix borrows a calibration technique from AI fairness research.

Opta adds a caution about the opposite mistake: a team that has outscored its xG does not then have to underperform to "even out". Assuming it must is the gambler's fallacy; future shots carry their own probabilities.