xG Guide · Part 1 · Chapter 4

Why xG differs between sites

There is no official xG. FIFA, UEFA and the leagues have no standard, and every data company runs its own model on its own data. The idea is the same; the numbers are not.

Who measures xG

Provider	Model (as published)	Main inputs	Penalty
Opta / Stats Perform	Gradient-boosted trees (XGBoost), about 1 million shots from 40 competitions; a separate model for women's football	20+: distance, angle, keeper position, pressure, clarity of view, shot type, pattern of play, the action before	0.79
Hudl StatsBomb	Gradient-boosted trees with sense-keeping rules	Every visible player at the shot ("freeze frame"), defenders in the way, share of goal covered, keeper stance	0.76 / 0.78
Wyscout (Hudl)	Its own machine-learning model, trained on all shots	Location, shot type, situation, assist type	0.76
Understat	A neural network, 100,000+ shots	Not published	—
Bundesliga (Sportec Solutions)	The league's own model	Not verified by us	—
FBref	Not its own: StatsBomb until October 2022, then Opta	As its provider	0.79 (Opta era)

What each provider shows publicly, as of 2026-10-28. Broadcasters show their data partner's model: Opta's for the Premier League and UK television.

Even the penalty differs

The value each provider gives a penalty: the only near-fixed number in xG, still not the same everywhere.

How big the differences are

13%of 2018/19 top-five-league matches: two providers disagreed on which team had more xG (KU Leuven)

20.48 · 17.77one striker's season xG (Aubameyang) from two providers, same shots

≈ 2 mhow far apart two providers placed the same shot, at most

The KU Leuven study compared two providers over a full season. Shot locations differed by up to a couple of metres, one provider left some shots out (such as those blocked at the shooter's feet), and values differed most close to goal. Rankings of the best players stayed broadly similar, but match totals often did not.

A model change moves the numbers too

Premier League open-goal chances: total xG before and after Opta's 2022 update, against the goals actually scored (Opta Analyst).

Four providers, nine thousand matches

An independent comparison published in November 2025 (MartinOnData, The Python Football Review) set Opta, StatsBomb, Wyscout and Understat side by side over about 9,000 top-five-league matches from 2017/18 to 2021/22, using each match's team totals.

1 in 4matches where at least one of the four providers named a different team as having had more xG

0.86–0.96how closely providers' match totals moved together (correlation; 1 would be identical)

1.33–1.44average xG per team per match in this sample: StatsBomb lowest, Wyscout highest

7 placesthe widest gap in one team's season xG ranking between two providers, even though season totals agreed closely (above 0.97)

Average xG per team per match, 2017/18 to 2021/22, top five leagues (MartinOnData, November 2025). Same matches, four models.

Together with the KU Leuven study, the message is consistent: over a season the providers broadly agree; for a single match, the verdict on who had the better chances flips in roughly one match in eight to one in four, depending on how many sources you compare. The comparison is a practitioner's study, not peer-reviewed, and uses match totals only.

People feed the model

Event data is collected by trained operators, live and from video. They agree very well on whether something happened; the only peer-reviewed reliability study of Opta data found very high agreement between two teams of operators, though for a single match. The judgement calls are where differences come from: the exact spot, foot or head, whether a pass was a cross or a through ball, how much pressure there was. Each of these is a model input, so a different call means a different xG. Analysts are warned never to mix providers' data.

What it means for you

Compare like with like: one provider, one model version.
Read small differences between sites as noise, not as a disagreement about the match.
Look for the source: a site that does not name its xG provider leaves you guessing.

Questions

Why does Understat show different xG from FotMob?

They use different models: Understat its own neural network, FotMob Opta's. Different inputs and training data give different values for the same shots.

Which xG is the most accurate?

No commercial provider publishes its accuracy figures, so no one can say for sure. Models with more information about each shot (keeper and defender positions) should do better in principle.

Is there an official xG?

No. There is no FIFA, UEFA or league standard. Leagues and broadcasters show their data partner's model.

Sources

Checked 2026-10-28. Methods change: every figure here is dated to its source.

xG explains football; it is not a betting tip. 18+ · Please gamble responsibly.