Expected goals has moved from analytics departments into broadcast graphics, and the way it is used publicly frequently misses what it is for.

What it is

An estimate of the probability that a given shot results in a goal, based on historical outcomes of similar shots.

Which means it is a measure of chance quality rather than of performance.

Summing it across a match gives an indication of how many goals the chances created would typically produce.

The inputs

Distance, angle, body part, type of assist, and whether it followed a set piece or a counter-attack.

Which are the variables that historical data shows matter most.

Models differ in what they include, which is why published figures for the same match vary between providers.

What it does well

Over many matches, it predicts future goal output better than past goal output does.

Which is a genuinely useful property and is the main reason clubs use it.

A team consistently outperforming or underperforming its expected total tends to regress.

What it does poorly

Single matches, where the sample is far too small.

Which is exactly where broadcasts use it most.

A team can win a match having lost the expected goals count, and that says almost nothing.

What it excludes

Defender positioning at the moment of the shot, goalkeeper position and shot placement.

Which are substantial omissions, addressed partly by models incorporating tracking data.

Post-shot models that include where the ball ended up measure something different — finishing quality rather than chance quality.

The finishing question

Whether players have durable finishing skill above the model's expectation is genuinely contested.

Which is because the sample sizes needed to demonstrate it are large and careers are short.

The honest position is that most apparent finishing differences are within the range of chance.

Chances not taken

The metric only counts shots, so a team that works into good positions without shooting registers nothing.

Which is why possession-value models measuring the whole sequence have developed alongside it.

Using it sensibly

Over a season, informative. Over a match, an interesting detail. As a definitive verdict on who deserved to win, misused.

Model differences

Different providers include different variables and train on different data.

Which is why the same match produces different published figures.

Comparing figures across sources without knowing the models is a common error in public discussion.

Penalties

A penalty carries a fixed high value that dominates a match total.

Which means a match with one penalty and little else can show a misleading picture.

Non-penalty figures are frequently reported separately for exactly this reason.

Recruitment use

Clubs use underlying numbers to identify players whose output may improve or decline.

Which is where the metric has genuine practical value.

Buying a player on the basis of an unusually high conversion rate is precisely the mistake these models help avoid.

Defensive application

Expected goals conceded assesses defensive performance independently of goalkeeping.

Which separates the quality of chances allowed from whether they were saved.

Communicating it

Presented as a probability rather than as a verdict, it is a useful addition to watching a match.

Sample size requirements

Meaningful conclusions need many matches rather than a few.

Which is the single most important caveat and the one most frequently ignored.

A season is a reasonable minimum for team-level conclusions and considerably more for individual finishing.

Possession value models

Frameworks assigning value to every action rather than only to shots.

Which addresses the main structural limitation of shot-based metrics.

These are used internally by clubs more than they appear in public discussion.

Goalkeeper assessment

Comparing goals conceded against post-shot expected values isolates shot-stopping.

Which is one of the clearer applications of the framework.

It excludes distribution and command of the area, which are substantial parts of the role.

Public reception

The metric attracted resistance partly because it was presented as a verdict rather than as a probability.

Which is a communication failure rather than a problem with the tool.

Using it well

Over a season, in combination with watching, and with the model's limitations understood.

Set piece separation

Chances from dead balls have different conversion profiles from open play.

Which better models account for separately.

A team dependent on set piece chances has a different underlying picture from one creating the same total in open play.

Closing

It is a useful probability estimate over sufficient samples and a poor arbiter of a single result.

Where it goes next

Tracking data incorporating defender and goalkeeper positions produces more accurate estimates.

Which is already in use at clubs and appears publicly less often.

The public version will probably improve as tracking data becomes more widely licensed.

A reasonable position

Use it as one input, over sufficient samples, alongside watching matches.

Which is how the people who built these models describe using them.

The confident public verdicts based on a single match are not how anyone working with the data actually operates.