The messages arrive in the same shape every weekend. A team wins 2-0 and loses the expected goals count, and somebody wants to know how a number can contradict a scoreline. Underneath the annoyance there is almost always a reasonable question about what the model was ever claiming.
Most of the confusion comes from one thing. Expected goals is a statement about a chance, made without reference to the person taking it. Once that sinks in, the arguments about whether a particular striker is undervalued mostly dissolve, and the remaining objections turn out to be the interesting ones.
What follows are the questions readers send most often, answered in the order they usually arrive, with the limits stated where they exist. The last section deals with the practical problem: how to put the number in a match report without overclaiming.
The questions readers keep sending
- What is it counting? The historical conversion rate of chances that looked like this one.
- Why is my striker not in it? Because most public models are built without the shooter’s identity, deliberately.
- Why is a header worth less? Because headers convert worse than feet from the same place, and the model has seen a great many of both.
- Why do two sites disagree? Different training data, different inputs, no shared standard. Differences of a couple of tenths in a match are normal.
- Can I use it after one match? For describing chances, yes. For judging finishing, no, and not for months.
- The honest use. As a description of what a team created, cited with its shot count, never as a verdict on a result.
What is an expected goals model counting?
Every shot in the historical database is described by a handful of measurements: how far from goal, at what angle, with which body part, from what kind of pass, under roughly what kind of pressure. Group together all the shots that share those characteristics, look at what proportion of them were scored, and that proportion becomes the value assigned to the next shot that matches.
So a value of 0.12 is not a prediction that a twelfth of a goal is about to happen. It is a report that historically, about twelve in a hundred shots of this description went in. It is a rate, borrowed from the past and pinned to the present.
This is also why the number has no opinion about what happened next. A chance rated 0.9 that is missed was still a chance that goes in nine times out of ten. The model is not wrong when the ball goes over the bar; it was never making a claim about that particular attempt.
Why does the model not know who is shooting?
Because including the shooter creates a problem the data cannot easily solve. Finishing skill and chance quality are entangled: a player who scores a lot may be a superb finisher, or may be the one who gets on the end of the best chances, or may have had a good year. Separating those requires far more shots than most careers contain.
The rough scale of the problem is worth stating plainly. Individual shot outcomes are extremely noisy, and the number of attempts needed before a finishing difference emerges from that noise runs into the hundreds. Most forwards take fewer than a hundred shots in a league season, which means a single season tells you very little about finishing and quite a lot about the chances a player is getting.
Some models do adjust for the shooter, and a few include the goalkeeper as well. They are legitimate, and they answer a different question. When someone quotes an unadorned xG figure from a public source, they are quoting a chance-quality number that treats every shooter as the league average, which is exactly what makes it comparable between teams.
Why is a header worth less than a shot with the foot?
Because headers from the same position have converted at a lower rate over very large samples. The mechanism is not mysterious: less power is available, placement is coarser, the ball is usually arriving from a cross with defenders converging, and contact happens in the air with no chance to adjust.
The consequence for reading a match is specific. A team that generates six headed chances in the six-yard box and a team that generates four shots with the foot from the same area can finish the match with similar shot counts and very different expected goals. Neither number is lying; they are describing different kinds of opportunity.
The same logic covers the other body-part adjustments. A shot with the weaker foot converts slightly worse in most datasets, a deflected or scuffed contact is not something a pre-shot model can see at all, and a first-time strike from a low cross has its own conversion history separate from a controlled shot.
Why is a penalty always the same number?
Because it is the one shot in the sport taken from a fixed spot, at a fixed distance, with the defence held out of the way. There is nothing left for a model to distinguish. Every penalty gets the same value, a little over three quarters of a goal, drawn from the long-run conversion rate of spot kicks.
This has a side effect that catches people out in recaps. A team that scores one penalty and creates nothing else can post a respectable expected goals total for the match, and reporting that total without saying it is almost all a penalty is misleading by omission.
The convention among analysts is to quote non-penalty expected goals when describing how a team played, and to mention the penalty separately. It is a small habit that removes most of the arguments before they start.
Why do two websites give different numbers for the same match?
There is no governing standard for these models. Each provider trains on its own historical data, chooses its own inputs, decides how to treat defensive pressure, and draws its own boundary around what counts as a shot. Two competent models can look at the same match and disagree by a few tenths of a goal.
The disagreements cluster in predictable places: big chances where one model sees a defender in the line and another does not, headers from crosses, and shots that follow a rebound. On a quiet match with ten ordinary attempts, models tend to converge.
The practical rule for anyone quoting it is to name the source and never mix providers in the same comparison. A season total from one model set against a match figure from another is not a comparison at all.
Is one match too small a sample?
It depends on the claim. Using a match total to describe the chances created is fine, because that is a description of things that happened rather than an estimate of an underlying quality. Using it to judge whether a team is good, or whether a striker can finish, is not.
A team takes something in the region of a dozen shots in a match, and most of those are worth less than a tenth of a goal each. That means a single high-value chance can account for a third of the total. One header cleared off the line, counted or not counted, moves the match figure more than the other twenty attempts combined.
The reasonable horizon for team-level conclusions is a stretch of matches rather than one, which is the same problem that turns up with early-season claims of every kind. Nothing about the model changes that; it is a property of how few shots a football match contains.
Does a high xG loss mean a team was unlucky?
Sometimes, and less often than the phrase implies. There are at least three other explanations, and separating them is most of the analytical work.
The first is finishing, which does vary between players even if it is hard to measure over short spans. The second is goalkeeping: a keeper having an exceptional night suppresses conversion, and a pre-shot model cannot see him. The third is chance structure, where a team accumulates a large total from many mediocre attempts rather than a few good ones, which is a strategic difference rather than misfortune.
The honest sentence acknowledges the ambiguity. A team out-created its opponent and did not convert; whether that is variance, finishing or a goalkeeper is not answerable from one match, and saying so costs nothing.
What does the number miss entirely?
Quite a lot, and knowing the list is what stops it being used badly. A standard pre-shot model does not see the shot after it leaves the foot, so placement, power and the keeper’s position are invisible. It does not see the chances that were never taken, which means a side that walks the ball into the six-yard box and passes when it should shoot is credited with nothing.
It also has no memory within a possession. A shot blocked on the line and the rebound that follows are counted as two separate chances, which can inflate a sequence that produced one real opportunity. Some analysts adjust for this; most published totals do not.
Finally it is blind to the state of the match. Chasing a two-goal deficit against a deep block produces a particular kind of shot volume, and comparing that total against a side protecting a lead compares two different games. Possession-value models, which score every action rather than only shots, exist partly to fill these gaps, and they carry their own assumptions.
How to cite the number without overclaiming
Three formulations survive scrutiny, and they cover almost every situation a match report needs.
- Pair it with the shot count. “Sixteen shots for 1.1 expected goals” tells a reader immediately that the volume was empty. Either number alone tells them less than half of that.
- Name the biggest chance. If one opportunity is worth a third of the total, say which one. It converts an abstract figure into a moment the reader watched.
- Use it to describe, not to overturn. The number describes what was created. It does not decide who deserved to win, and a report that says a team deserved better on xG has stopped reporting and started arguing.
The table below sets out the version to avoid against the version that holds.
| Claim | Why it fails | Version that holds |
|---|---|---|
| They deserved to win on expected goals | Turns a description into a verdict | They created the better chances and did not take them |
| He should have scored, it was 0.4 | 0.4 means it is missed more often than scored | It was the best chance of the half and went begging |
| Their xG proves the manager is wrong | One match cannot support a claim about method | Over this run the chance quality has not improved |
| They posted 1.8 xG | Hides a penalty and the shot count | 1.8 including a penalty, from fourteen attempts |
What xG is for, in the end
It exists because shot counts were a poor summary and scorelines are a small sample. Between those two, expected goals gives a description of what a team built that does not swing on whether one ball crossed the line. That is a genuine improvement, and it is also the entire claim.
Treated that way it settles arguments rather than starting them. Treated as a verdict on who should have won, it becomes the same rhetorical device as possession share was a decade ago, which is a fate that has met most new measurements in sport at some point in their adoption.
The best test of whether you are using it well is whether the sentence would survive being read out to the coach of the losing side. If it describes chances, it survives. If it awards a moral result, it does not, and the same standard is worth applying to the louder end of sports argument generally.
Frequently Asked Questions
Does a shot from outside the box ever get a high value?
Rarely. Long-range attempts convert at a low rate across enormous samples, so almost all of them land somewhere between two and five in a hundred. A spectacular goal from thirty metres is a low-value shot that went in, and the model is not wrong about it.
Is expected goals used inside clubs or only by the public?
Both, but not in the same form. Club analysis departments generally work with richer models that include tracking data, defender positions and post-shot information, and they use them for recruitment and opposition work rather than for arguing about a weekend result.
What is post-shot expected goals?
A second model applied only to shots on target, which uses where the ball ended up in the goal. It measures the quality of the strike and is mainly used to evaluate goalkeepers, since the difference between goals conceded and post-shot expectation is a reasonable shot-stopping metric.
Should a low-scoring sport use this at all?
The idea generalises wherever chances can be described and outcomes counted, and versions exist in hockey and basketball. The scarcity of goals is precisely why football needed it, because with two or three scoring events a match the result carries very little information about the play.
How many matches before a season total means something?
Team-level chance-quality numbers begin to be informative faster than most people expect, but a single figure is still worth reading alongside the fixtures behind it. Judging a squad on ten matches without checking who those matches were against remains the most common error in the genre.
The number is a description of chances, nothing more and nothing less. Quote it next to the shot count, say where the penalty is, and most of the weekend’s arguments never get started.
