Against the Tyranny of Expected Goals

I want to be clear from the outset that expected goals is a good metric. It is better than shot counts, dramatically better than possession percentage, and it has been genuinely useful in dragging football analysis out of a period where the sport was discussed almost entirely through anecdote. None of what follows is an argument that xG is wrong.

It is an argument that we have started using it for jobs it was never designed to do, and that the results are making us worse at watching football rather than better.

What the Number Is Actually For

Expected goals estimates the probability that a given shot becomes a goal, based on the characteristics of similar historical shots. Distance, angle, body part, and the nature of the assist go in, and a number between zero and one comes out. Sum them across a match and you have an estimate of how many goals a set of chances would typically produce.

The critical word is typically. The metric is a statement about a distribution, and it becomes reliable as the sample grows. Across a season, a team's xG total tells you something meaningful about the quality of chances they are generating. Across ninety minutes, it tells you considerably less than the confidence of its users suggests.

The Single-Match Problem

Yet the single-match xG figure is now the dominant unit of post-match discussion. A side loses 1-0 and the graphic appears: 2.4 to 0.7, an injustice, a performance that deserved more.

Sometimes that reading is correct. Often it is not, and the reason is that a match-level xG total can be assembled in wildly different ways that the summary number flattens completely. Ten shots of 0.24 each and one shot of 2.4 are not remotely the same performance, and the second is not even possible, which is precisely the point. A team that racks up xG through repeated speculative efforts from the edge of the box has not created the same thing as a team that engineered three clear openings.

The total tells you how much. It refuses to tell you how, and how is usually what you wanted to know.

What Gets Lost

Here is my actual complaint. Expected goals measures shots, and football is mostly not shots. The metric has nothing to say about the pass that broke a line and produced nothing, the run that dragged two defenders and created space for a teammate who chose not to use it, or the twenty minutes of territorial suffocation that ended without a chance because the final ball was poor.

These things are the substance of most matches. A metric that ignores them is not lying, but a discourse that treats that metric as a summary of the match is losing most of the information. We have swapped one kind of blindness for another and congratulated ourselves on the upgrade.

The Deserving Fallacy

The phrase that irritates me most is that a team deserved to win on xG. Deserving is a moral claim, and expected goals is not a moral instrument. A side that creates four excellent chances and misses all four has not been robbed. They have failed to finish, and finishing is a skill the sport has always valued, quite reasonably, because putting the ball in the net is the entire object of the exercise.

There is a defensible version of this argument, which is that finishing is less repeatable than chance creation and therefore a side with high xG and low goals is likely to improve. That is a forecast, and a decent one. It is not a claim about justice, and collapsing the two has made a useful predictive tool into a rhetorical weapon.

The Defensive Blind Spot

There is also an asymmetry that rarely gets acknowledged. Expected goals conceded is a much noisier signal than expected goals created, because good defending frequently manifests as the shot never being taken at all.

A defence that forces a striker into a poor decision and a hopeless effort from twenty-five yards is recorded as having conceded a small xG value. A defence that is carved open but recovers to block the shot at source is recorded as having conceded nothing. The metric handles the second case badly and the first case backwards.

What I Would Rather We Did

Use xG over seasons, where it is strong. Use it to evaluate recruitment, where the sample supports it. Use it as one input into a judgement about whether a run of results is sustainable, which is genuinely what it is best at.

Stop putting it on screen four minutes after full time as though it settles anything. Stop treating a match-level figure as a verdict. And stop assuming that a number's existence makes it the most important thing available.

Football was analysed badly for a long time. The corrective was necessary. But a single metric, however good, was never going to be able to carry the whole conversation, and the conversation is noticeably poorer for asking it to try.