Why expected goals proves it is a weaker argument than it appears

What the model is built to do

A shot-quality model assigns each attempt a likelihood of being scored based on features such as distance, angle, body part and the pattern that created it. It is fitted on a very large collection of past attempts, which means its output describes what a typical attempt from that position tends to produce. That is a genuinely useful thing to know, because it separates the volume of chances from the outcome of the finishing.

The utility depends on the reader understanding that the estimate is about a category of shot rather than about the specific shot taken. Almost every misuse of the number comes from forgetting that distinction.

The average finisher assumption

Because the model is fitted across many players, it implicitly assumes the shooter is an average member of the population it was trained on. That assumption is reasonable in aggregate and can be badly wrong for a single attempt taken by an unusually good or poor finisher. It is also wrong in a systematic direction for teams whose attacking players are consistently drawn from one end of that distribution.

The model cannot correct for this without becoming a model of players as well as positions, which changes what it is measuring. Using the output as a statement about what should have happened therefore imports an assumption the user has not examined.

Sample size and the illusion of precision

A single match contains a small number of attempts, so the total for that match is an estimate built from very few observations. Estimates from small samples move around a great deal, which means two matches with different totals may not differ in any meaningful way. The number is nonetheless quoted to two decimal places, and that presentation conveys a precision the underlying data does not support.

Over a season the totals become considerably more informative, because the noise averages out as the count of attempts grows. The same statistic can therefore be near-worthless over one match and genuinely revealing over many, which its users rarely signal.

What the model does not see at all

The model observes shots, so a passage of sustained territorial control that produces no attempt contributes nothing to the total. A side that defends deep and concedes possession without conceding attempts can look dominant in the metric while never threatening. Equally, the value of forcing an opponent into a low-quality attempt rather than no attempt at all is recorded as a positive for the attacking side.

These are not errors in the model; they are consequences of choosing shots as the unit of analysis. Anyone using the output to summarise a match is using a shot statistic to answer a question about everything that is not a shot.

Using it as an instrument rather than a judge

The defensible use is comparative and cumulative: tracking whether a side is generating better positions over a run of matches than it was before. The indefensible use is settling an argument about a single result by quoting a total as though it recorded what deserved to happen. Deserve is a moral idea and the model has no access to it, since it only estimates frequencies from historical attempts.

A tool that answers a narrow question well becomes actively misleading when asked a broad one it was never designed for. The strength of the method is the assumptions it makes explicit, and quoting the output alone throws that strength away.