Twenty years ago, a shot was a shot. Now every televised chance flashes an expected-goals value before the ball has stopped rolling, and "but the xG…" has become the way arguments about football are won and lost. It is the most important number in the modern game, and also the most misunderstood. Here is what it actually says.

What expected goals actually measures

Expected goals, or xG, is a single number attached to a single shot: the probability that a shot taken from that position, in that situation, becomes a goal. It runs from 0 to 1. A tap-in from two yards might be worth 0.9 xG: nine times out of ten, it's a goal. A hopeful dig from thirty yards might be 0.03: you'd expect it to go in about once every thirty-three attempts. Add up every shot a team takes in a match and you get the team's xG: the number of goals an average finisher would have scored from those chances.

That reframing is the whole point. A 1–0 scoreline tells you who won. The xG tells you whether the winner deserved to: whether they manufactured better chances or simply had a better afternoon in front of goal. Take the 2022 World Cup final, a 3–3 epic that Argentina won on penalties. Strip out the shootout and the underlying numbers were less of a coin-flip than the scoreline suggested.

2.76
Argentina xG
2.27
France xG
20–10
Shots, Argentina–France

Argentina out-created France 2.76 to 2.27 and out-shot them two to one. France's evening was a monument to finishing: three goals (two of them Kylian Mbappé penalties and a half-volley) from 2.27 xG and ten shots. The single biggest chance of the match belonged to Lionel Messi, a penalty worth 0.78 xG, which he scored. That is xG in one game: not a verdict, but a much fuller account of the story than the final score.

Shot map of the 2022 World Cup final, Argentina versus France, with each shot drawn as a circle whose area is proportional to its xG and goals marked as stars.
Every non-shootout shot in the 2022 World Cup final, sized by xG. Argentina's volume of good chances is the story the 3–3 hides.Data: StatsBomb open data, retrieved June 2026.

How an xG model is actually trained

An xG model is not a panel of pundits assigning gut-feel difficulty scores. It is a machine-learning model trained on hundreds of thousands of historical shots, each one labelled with whether it was scored. The model learns, from that mountain of evidence, how the features of a shot map to the probability of a goal.

The features that matter most are geometric, and they are exactly the ones your eyes already use:

  • Distance to goal. The single strongest predictor. Closer is better, and the relationship is steep.
  • Angle to goal. How much of the goal mouth the shooter can actually see. A shot from the byline has almost nothing to aim at, even from close range.
  • Body part. Headers convert worse than the equivalent shot with the feet; a header from the penalty spot is a different proposition from a volley there.
  • Type of chance. Was it a fast break with the defence scrambling, a set-piece scramble, a penalty (a near-constant 0.76-ish), a one-on-one with the keeper?
  • Defenders and the goalkeeper. The richest datasets, StatsBomb's among them, record the exact position of every defender and the keeper at the instant of the shot, so the model knows whether the shooter was crowded out or clean through.

Feed those features into a model (historically a logistic regression, now usually a gradient-boosted ensemble) and out comes a probability. Crucially, the model is fit on what actually happened to similar shots, not on anyone's opinion. When it says 0.107, it means: across all the shots that have ever looked like this one, about 11 in 100 were scored.

Does it actually work? The calibration test

A probability model is only as good as its calibration: when it says 0.20, do those shots really go in about a fifth of the time? The honest way to check is to bin a big pile of shots by some feature, then compare the model's average xG in each bin against how often those shots were actually scored. If the model is any good, the two lines should sit on top of each other.

Here is that test run on every non-penalty shot at the 2022 World Cup (1,425 of them) grouped by distance from goal.

Non-penalty shots at the 2022 World Cup, by distance from goal. Penalties excluded. Data: StatsBomb open data, retrieved June 2026.
Distance bandShotsMean xG (model)Actual conversion
0–6 yards500.4344%
6–12 yards3140.1619%
12–18 yards3660.1113%
18–24 yards3140.055%
24–40 yards3810.022%
Line chart comparing mean xG to actual goal conversion rate across five distance bands for all non-penalty shots at the 2022 World Cup; the two lines track each other closely.
The model's average xG (green) against how often those shots were really scored (clay). They track each other band for band: the mark of a calibrated model.Data: StatsBomb open data, retrieved June 2026.

Band by band, the model's average probability lands within a couple of points of reality. That is what "xG works" means in practice: not that it predicts any single shot (it can't, and doesn't try), but that across many shots its probabilities come true. The average World Cup shot was worth 0.107 xG, and the tournament's 1,453 shots produced 169 goals. The model knew that before a ball was kicked.

What xG is good for, and what it isn't

xG is a sample-size instrument. Over a season it is one of the most predictive numbers in football: teams that consistently create more than they concede tend to climb, and a side massively outscoring its xG is usually living on borrowed time. Over ninety minutes it is noisier: a single fluke screamer or a missed sitter can flip the xG narrative, which is why "they lost the xG battle" is a thin argument about one game and a strong one about ten.

The most common abuse is treating xG as a measure of finishing. It isn't; it's a measure of chance quality. A striker who repeatedly outscores his xG might be an elite finisher, or might be on a hot streak that regression will quietly end. The honest stance is humility: xG tells you how good the chances were, and the gap between xG and goals is a question to investigate, not an answer to brandish. Used that way, it really did change how we watch the game.

Reading one match of xG without overreacting

This section condenses two separate articles first published on 10 June and 15 June 2026.

The scoreline hides the most in a 1-0. The 2022 World Cup had 10 of them, and the two ends of that list tell opposite stories. Switzerland beat Cameroon 1-0 having created 1.43 xG to Cameroon’s 0.61: on the balance of chances, the better side by a distance. Cameroon beat Brazil 1-0 with 0.38 xG against 2.38, and 7 shots against 21, the goal arriving in second-half added time (91:48 on the match clock): they got away with one. Same result, opposite stories. The habit worth building is to write the score and the xG side by side after every match (“1-0, xG 1.43 to 0.61” or “1-0, xG 0.38 to 2.38”) and decide which kind of result it was: deserved, smash-and-grab, or genuinely even. Of the tournament’s 49 matches that had a winner after extra time, the side with less xG won 17.

Two traps catch people reading a running xG total live. The first is the single big chance: xG is cumulative, so one penalty or one point-blank header can swing a team’s total from 0.4 to 1.2 in a single moment. Always ask whether a team’s xG is built from many decent chances or one outlier. The second is game state: a team that goes 2–0 up after twenty minutes will often sit deep and concede territory, and its opponent then piles up shots without being the better side. Late xG against a parked bus is cheaper than early xG against an open game, so read the live total alongside the scoreline and the clock.

Game state can invert a whole match. Australia beat Tunisia 1-0 at the 2022 World Cup while “losing the xG” 0.36 to 1.05. Mitchell Duke scored at 22:50, and from then on Tunisia needed a goal and Australia did not. Tunisia took 13 of its 14 shots while behind, and those 13 carried 1.02 of its 1.05 xG (97.2%); while the score was level it managed one shot, worth 0.03. Morocco’s 2-1 win over Canada has the same shape: Morocco led from 3:29, and 1.06 of Canada’s 1.10 xG came while it trailed (4 of its 5 shots). The xG measured the shots faithfully; read without the scoreline, the raw total describes a match in which only one side still needed to attack. That distortion, and how to correct for it, is the subject of game state and score effects.

The underlying problem is sample size. A match xG total is the sum of roughly a dozen shot probabilities, so it swings widely game to game even when a team’s underlying quality is unchanged. One match cannot settle how good a team really is, whether its finishing is lucky or skilled, what the result looks like corrected for who was chasing the game, or how many chances a defence or keeper prevented before they ever became shots. In a tournament, that means accumulating matches before concluding, splitting totals by game state, and reaching for post-shot xG and goalkeeper metrics before blaming the finishing. After one group game a team’s xG is a whisper; by the quarter-finals, five or six matches in, the cumulative picture starts to describe something repeatable. The number worth tracking most closely is expected goals conceded: defensive structure is harder to fake across multiple rounds than chance creation is to sustain against elite defences.

Sources, notes & further reading

  • StatsBomb open data: the shot-level data and xG values used throughout this article.
  • The four match examples in “Reading one match of xG without overreacting” (added November 7, 2026): Switzerland 1-0 Cameroon, Cameroon 1-0 Brazil, Australia 1-0 Tunisia and Morocco 2-1 Canada, from the 2022 World Cup event files in StatsBomb’s open-data repository, StatsBomb Open Data (StatsBomb Public Data User Agreement), retrieved June 4, 2026. Team xG includes penalties and leaves out shoot-out kicks; a shot’s game state is the score when it was taken. Computed by our analysis script; the match totals are in data/stats/expected-goals-explained-examples.json.
  • StatsBomb: primers on their xG model and what its freeze-frame features add.
  • Understat: season-long xG tables for the major European leagues.
  • FBref: xG and advanced data (via Opta) for a wide range of competitions.

Published by SoccerAnalytics from public data and the sources named in this piece. How this site is made.