What is expected goals (xG)? A complete beginner’s guide

Expected goals, almost always written xG, is the most widely quoted statistic in modern football analytics. You will hear it on broadcasts, see it under every shot map on social media, and find it in every serious post-match report. But what does it actually measure, where does the number come from, and what does a value such as “0.23 xG” really mean? This guide walks through the idea from first principles, drawing on the literature review from my 2024 undergraduate project at UFRJ on xG models.

The problem xG was invented to solve

Football is a low-scoring sport. A single match can turn on one deflection, a mistimed clearance or a shot rattling off the crossbar. Because of that, the scoreline often fails to capture which team actually played better. A side can dominate possession, take 25 shots, hit the woodwork twice, and still lose 1–0 to a team that converted its only real chance.

Traditional statistics — shots, shots on target, possession percentage — treat every attempt as equal. A scuffed effort from 35 yards counts the same as a point-blank header from six. Expected goals sits on top of that raw data and answers a sharper question: given where this shot was taken and under what conditions, what is the probability of it becoming a goal?

A plain-English definition

Expected goals is a statistical metric that estimates the probability of a shot ending in a goal, expressed as a number between 0 and 1. An xG of 0.30 means that, historically, shots taken in very similar circumstances went in roughly thirty percent of the time. Add up the xG of every attempt in a match and you get a team’s xG total — a measure of how many goals the quality of their chances would typically produce.

Two consequences follow:

• xG is about process, not outcome. A team with 2.3 xG that scores zero has still created very good chances — they simply did not convert that day.
• xG is descriptive, not deterministic. A specific 0.30 shot will either go in or not; the number expresses the long-run frequency.

Where the numbers come from

Building an xG model is a supervised machine-learning problem: researchers take a large historical dataset of shots for which they know whether each one became a goal, describe each shot through a set of features (x/y coordinates, distance, angle, body part used, defender positions, shot technique, whether it was taken under pressure or first time) and then train a model to predict the probability of a goal from those features.

The simplest approach, used by Rathke in his 2017 paper, divides the opposing half into a handful of zones and reports the historical scoring rate for each zone. Zone 1, a tiny area right in front of the six-yard box, has a conversion rate close to 40%; zone 6, the part of the pitch near the sideline outside the penalty area, sits below 3%. It is easy to build intuition from those numbers alone: a shot from the edge of the six-yard box is not “ten times” better than a shot from outside the box, it can be thirteen times better.

Modern xG models are far more sophisticated. They use logistic regression, decision trees, random forests, gradient boosting, AdaBoost or convolutional neural networks, and they incorporate defender positions, goalkeeper location, player density and even biographical data about the shooter. Evaluation is usually reported through a ROC-AUC score, the Brier score, Precision and Recall, and correlation with a reference xG source such as StatsBomb’s own model.

Reading xG: some worked examples

Imagine three shots. A tap-in from two yards with an empty net might be worth roughly 0.85 xG: very hard to miss, but not impossible. A header from the penalty spot, defender in front and goalkeeper set, might score 0.15 xG. A speculative curler from 30 yards with three defenders in the way may be worth only 0.02 xG.

Over a season, the gap between a striker’s goals and their xG tells you whether they are over-performing (finishing above their chance quality, which tends to regress) or creating high-quality chances without converting.

What xG does not tell you

xG is powerful but limited. It cannot, on its own, tell you whether the shooter had a clear line of sight, whether the pass preceding the shot was in the air or on the deck, or whether a defender was lunging in at the exact moment of contact. Different models capture different amounts of context, which is exactly why this site lets you compare several of them at once. No single xG number is the “truth” — each is an opinion about shot quality, filtered through the features a particular model can observe.

Where to go next

Play with the simulator on the homepage: place an attacker in different zones, add and remove defenders, move the goalkeeper, and watch the xG from each model change in real time. When you are ready, read how xG models differ for a side-by-side tour of the academic models used in this calculator, or the xG glossary for the rest of the vocabulary.