Expected Goals (xG) Explained: What the Number Really Means
A team loses 1β0 despite taking 18 shots. The next week they win 2β1 from three attempts. If you only look at the scoreline, the first performance was terrible and the second was clinical. Expected goals β xG β says the opposite may be true: the first team probably played well and got unlucky, the second probably got away with one. It is the single most useful number in modern football analysis, and also one of the most misunderstood.
What xG actually is
Every shot is assigned a value between 0 and 1: the probability that a shot like it becomes a goal, based on how often similar shots have gone in before. A penalty is worth about 0.76β0.79 depending on the provider β Opta uses 0.79, StatsBomb 0.76 (raised to 0.78 in 2022) β because roughly three in four penalties are scored. A header from a corner is worth about 0.06. A speculative effort from 25 yards sits around 0.03. A tap-in from six yards out, central, is worth roughly 0.40β0.50.
Add up every shot a team takes and you get its xG for the match: how many goals an average team would have scored from those chances. A team that wins 2β1 with an xG of 0.6 against 1.8 created the worse chances and probably rode its luck. That is the whole idea β separate the quality of the chances from the randomness of whether they went in.
A short history
The term "expected goals" dates back to a 1993 academic paper by Vic Bennett and Sarah Hilditch, but the modern metric arrived in April 2012, when Opta analyst Sam Green published a blog post assessing Premier League goalscorers. Green asked a simple question: if two strikers score the same number of goals, but one needed twice as many shots, who is the better finisher β and who just had the better chances? He built a model from Opta's shot data that valued each attempt by its location, angle, body part and the type of pass that created it.
In 2011β12, Darren Bent accumulated 9.9 xG from 43 shots while Adel Taarabt managed just 3.0 from 73 β same league, wildly different chance quality. Green also flagged Luis SuΓ‘rez as "especially unfortunate" after 11 league goals in 2011β12; SuΓ‘rez then scored 23 the next season and 31 the one after. The chances were always there; the goals caught up.
It took five more years for xG to reach the mainstream: the BBC's Match of the Day started showing it in the 2017β18 season, and today it is on every broadcast graphic and stat sheet.
How the models are built
An xG model is a classification algorithm β usually logistic regression or a gradient-boosted tree β trained on hundreds of thousands of historical shots, each labelled goal or no goal. Opta's current model was trained on nearly a million shots from 40 competitions between 2018β19 and 2021β22, using more than 20 variables.
The inputs, in rough order of importance:
- Distance and angle to goal. The dominant factor. Value falls off steeply with distance and toward the touchlines, where the scoring angle narrows.
- Body part. A header from the penalty spot is worth far less than a shot with the foot from the same place.
- Assist type and phase of play. A through ball, a cross, a rebound, a counter-attack and a set piece all convert at different rates from the same spot.
- Defenders and goalkeeper. Advanced models use freeze-frame data: how many defenders sit between shooter and goal, and where the keeper is standing. A shot with two defenders in the way is worth less than the same shot with a clear sightline.
Different providers train on different data with different inputs, so the same shot gets slightly different values β Opta might call it 0.25, StatsBomb 0.21. Treat any xG figure as an estimate, not a measurement to two decimal places.
How to read it β and how not to
Over a season, trust it. The best single use of xG is the difference between a team's xG for and against (xG difference). Across a full season it predicts future results better than the league table does, because the table is polluted by the luck xG strips out. A mid-table side with a top-six xG difference is usually a team about to climb.
In a single match, don't over-read it. One game is about 25 shots total, and the randomness in a handful of 0.10 chances swamps the signal. A 2.4β0.8 xG win is meaningful; a 1.1β0.9 draw tells you almost nothing.
It doesn't judge the shooter. Standard xG asks what an average player would do with the chance. It does not credit a great finisher for beating the model or blame a poor one for missing sitters β that gap, goals minus xG, is itself the measure of finishing, and it is mostly noise for everyone except the very best and worst finishers over very large samples.
It ignores game state. A team leading 1β0 late often stops attacking, and a team chasing the game takes desperate shots; xG counts the chances the same either way.
How xG connects to our predictor
GoalMath's match predictor is built on the same idea at team level: Ξ» (lambda) is simply a team's expected goals for the match, derived from attack and defense strengths relative to league averages, and fed into the Dixon-Coles model to produce 1X2, over/under and BTTS probabilities. xG looks backward at the chances a team created; Ξ» looks forward at the chances it should create. Both are the same question β how good were the chances? β asked in different directions.
Used that way, xG is the honest fan's best friend: it tells you whether your team played well on a night the scoreboard says otherwise, and whether a winning run is built on something real. The score tells you what happened. xG tells you what usually happens from chances like those β and over a season, "usually" wins.
Related reading: How the Dixon-Coles Model Predicts Soccer Scores Β· How to Read Our Match Predictions