xG in football: what expected goals are and how they are calculated

Created in 2012, xG arrived in football through the analysts' door and now shows up on the broadcast graphic, in the manager's press conference and in the pub argument. It arrived with a stubborn misunderstanding attached: plenty of people think the number predicts the result. It does not. xG is a quality score given to a shot after it happened, and it exists to separate the side that created good chances from the side that shot a lot from distance. Opta itself records, in describing the metric, its adoption by broadcasters such as Sky Sports, the BBC's Match of the Day, beIN Sports and NBC.

What the number measures, and the scale it uses

Expected goals, or xG, measure how likely a shot was to be scored by comparing it with similar shots from the past. The scale runs from zero to one: zero is a chance nobody converts, one is a chance a player would be expected to score every single time. Opta, which introduced the metric in 2012 through the analyst Sam Green, calibrates its model on nearly one million shots from its own historical database.

With an average of 25 shots per game, counting by eye does not hold up, and the idea is older than the acronym anyway. When somebody says a striker misses that nine times out of ten, they are describing a shot with a high xG. What the metric adds is a number you can add up, compare and follow across a full season instead of a feeling you argue about.

Opta's model runs on a machine-learning technique called XGBoost, and it does not look at the shot in isolation. It weighs distance to goal, angle, the type of finish, the action that came before and the goalkeeper's position. The keeper's distance stands in for reaction time, and the model also accounts for whether he was inside the box and able to use his hands.

The penalty is the declared exception. Because it is the most consistent shot in football, it does not go through the model at all: it is given a constant value, and Opta uses 0.79 xG, reflecting the historical conversion rate. That is the most useful anchor a viewer has. A team that finishes on a little over 0.79 xG created the equivalent of one penalty and almost nothing else.

What xG does not tell you

None of this forecasts who wins. A side can pile up good chances and lose, and that does not break the number: it describes precisely what happened, which was creating a lot and converting little. Over a single match luck rules; the metric only becomes trustworthy once dozens of games are stacked together.

It does not judge the player's decision either. A shot from outside the box with a low xG can be the right choice when no pass exists, and a spectacular goal remains spectacular even when it came from a chance the model rated as unlikely. xG measures the chance, not the merit of whoever resolved it.

Why two sites show different xG

There is no official xG. Every provider trains its own model, on its own data and its own variables, so the same shot comes out slightly differently depending on who does the sums. Opta even keeps a separate model for women's football, because measurement showed distance behaves differently there: applying the men's model would underestimate close-range shots and overestimate long-range ones.

For a viewer, the useful reading is simple. Compare xG with goals for the same team across many matches rather than one; distrust any conclusion drawn from ninety minutes; and remember the number describes the chance that existed, not the one that might have existed had the pass come off.