From counts to estimates
Traditional sports statistics are counts: goals, shots, kills, assists. Model-based statistics are different. They are estimates produced by a mathematical model that has been fitted to past data. Expected goals, usually written xG, is the best-known example from football, but the idea appears across sport and esports, from win probability charts to player rating systems.
The central question a model answers is: given the situation, how likely was this outcome, judging by what has happened in similar situations before?
How expected goals works
In football, an expected goals model assigns each shot a number between 0 and 1. That number is the estimated probability that a shot of that kind becomes a goal. The model is typically built from a large collection of past shots and what happened to them. Inputs often include the distance from goal, the angle, the body part used and the type of play that led to the shot. Different providers use different inputs and different methods, so their numbers can disagree.
To get a team’s expected goals for a match, you add up the values for its shots. An illustrative example with three shots:
- Shot A, a long-range effort: 0.05
- Shot B, a close-range chance from a cross: 0.30
- Shot C, a clear chance in the box: 0.65
The total is 0.05 + 0.30 + 0.65 = 1.00. On those probabilities, a team creating exactly these chances would be expected to score about one goal, on average, if the same situation were repeated many times. In a single match it might score none or two. The expected figure describes the average of many repetitions, not what must happen.
What it is good for
Because goals are rare in football, final scores are noisy. A team can play well and lose, or play poorly and win by a fluke. Expected goals tries to describe the quality of chances created and allowed, which can be a steadier guide to performance than the scoreline over a few games. If a team repeatedly scores far more or far fewer goals than its chances would suggest, that gap is often a mix of finishing luck and the limits of the model. Over a longer time, results often drift back towards what the chances suggest, which is the same effect as regression to the mean.
What it does not capture
Models are only as good as the data and assumptions behind them. Some typical limits:
- Missing context. A model may not know how closely a shot was defended, the goalkeeper’s position, or how a player is feeling.
- Average assumptions. A model usually treats a shot as an average shot of its type. A particularly accurate or poor finisher may differ from that average.
- Different providers, different numbers. There is no single official value, so comparisons across sources need care.
- Not a prediction of the next match. It summarises past chances, and says little with certainty about what comes next.
- Small samples. One match of expected goals is a thin base. The caution in the article on small sample sizes applies strongly.
Other model-based numbers
Expected goals is one member of a family. Win probability charts estimate each side’s chance of winning at every moment, based on similar past situations. Rating systems such as Elo ratings turn results into a single strength estimate. Player value metrics try to combine many actions into one number. Esports analytics uses the same approach with data such as gold, objectives and positioning, and sports probability models cover wider uses. They share the same structure: a simplified description of reality, calibrated on history.
How to read them
Treat a model-based statistic as an informed estimate, not a measurement. A few habits help:
- Ask what the model includes and, just as important, what it ignores.
- Compare over many games rather than one.
- Remember that the output is a probability. A shot rated 0.30 is expected to fail about seven times out of ten, and a miss does not mean the model was wrong.
- Be wary of precise-looking decimals, which can create an impression of certainty the data does not support.
- Look for how the creator checked it. A well-built model is usually tested on data it was not fitted to.
This article is about understanding what the numbers mean for analysis and viewing. It is not guidance for predicting results or placing money on them, and no model removes the uncertainty in games where chance plays a part.
Why the idea matters beyond football
Understanding model-based statistics builds an intuition that carries across fields. A model gives a probability, the real outcome is one draw from that distribution, and judging the model requires many such draws. Once that clicks, much of the confusion about “the stats said we should have won” falls away: a good chance that does not convert is not a contradiction, it is how probabilities work.