Why expected goals outgrew the sport that invented it
Fifteen years ago, expected goals was a private metric used by a small handful of analytics consultancies in English soccer. Today it's printed on the bottom of broadcast graphics, dropped into post-match interviews by managers, used by recruitment departments to value players, and copied — in spirit if not in name — by basketball, hockey, and lacrosse. It's not the most rigorous stat ever invented and it isn't the most accurate. It is, almost certainly, the most important publicly available sports stat of the last twenty years. It's worth understanding what it is, what it does well, and where the limits actually are.
What it does
Expected goals — written xG — assigns each shot a probability of becoming a goal based on the situation in which it was taken. A toe-poke from twenty-five yards out gets a low number, somewhere around 0.02 (a 2% chance of going in). A clean header from six yards out off a cross gets a high number, somewhere around 0.4. A penalty kick gets roughly 0.78, the league-average penalty conversion rate. Sum every shot in a match and you get a team's xG for that match. Sum every shot a player takes over a season and you get their season xG.
That single quantity — how many goals you should have scored, given the chances you actually generated — turns out to do a remarkable amount of work. It's a much better predictor of next-match performance than goals are, because goals fluctuate around xG with a lot of noise; the underlying shot-quality stream is the more stable signal. It's a much better evaluation tool for strikers than goals alone, because two strikers with the same conversion total can be producing wildly different shot quality, and the one producing higher xG is more likely to keep scoring.
What it travels
The reason xG escaped soccer is that the underlying idea travels. Any sport where you can assign a meaningful "probability that this scoring attempt becomes a score" can build the same metric, and every analytics-mature sport eventually has. Hockey runs an expected-goals model on every shot, sometimes called xG and sometimes goals saved above expected when applied to the goalie. Basketball runs a closely related model on shot location and defender distance, usually called expected effective field-goal percentage. Lacrosse and water polo have their own versions.
The traveling happened for two reasons. The first is methodological: the math is the same. You collect a large database of shots, annotate them with the relevant situational features, and fit a model that predicts conversion. The model itself is not exotic — a modest gradient-boosted tree will do — and the data is increasingly available. The second reason is rhetorical: xG gives you a way to say "your team is overperforming" or "your striker is unlucky" with quantitative weight behind it. Once one sport had a vocabulary for that, the others wanted one too.
What xG does well
Three things, mostly. First, it separates shot generation from finishing. Two teams can take the same number of shots and produce very different xG totals because the shots had very different quality; xG calls that out. Second, it stabilizes evaluation across small samples. Goals are noisy and arrive in clusters; xG accumulates much more smoothly, which means a striker with twelve goals in eighteen matches can be correctly diagnosed as either a sustainable scorer or a regressor based on their shot stream. Third, it gives front offices a defensible language for talking about process versus outcome — important when the outcome is dictated by a goalkeeper's reflexes, a deflection, or pure luck.
What it doesn't capture
xG is honest about what it is, but it's worth being explicit about what it isn't. The standard model knows about shot location, shot type, body part, and the type of pass that preceded the shot. It does not know about defensive pressure, except indirectly. It does not know whether the goalkeeper was wrong-footed. It does not know how good the striker was at creating the half-yard of space that made the shot possible in the first place. There is more to scoring than the moment of the shot, and xG by definition only sees the shot.
That's why advanced versions of the metric exist. Defensive xG adjusts for opponents and pressure. Post-shot xG (the model that recalculates after the ball is struck, given how it was struck) is used to evaluate goalkeepers. Possession-value frameworks try to credit the build-up play, not just the shooter. Each of these is a more sophisticated answer to the question "what is xG missing?" and each is incrementally better, at the cost of being less transparent.
What it broke
xG broke the football media's relationship with goals, and that's mostly been good. Before xG, an over-performing forward was treated as a hot striker; an under-performing one was treated as a guy having a bad season. After xG, both became potentially mean- reverting cases — and the front offices that bet on regression to xG generally beat the front offices that bet on goals. The smart clubs sold high on overperformers and bought low on underperformers before the market caught up.
The market has now caught up. xG is on the broadcast, fans quote it, agents reference it in negotiations, and the cleanest regression-to-mean trades are no longer cheap. The cycle is familiar: the metric becomes information, the information becomes consensus, the edge gets priced in. Front offices respond by going deeper — possession value, off-ball runs, defensive contribution models — and the public eventually catches up there too.
The reason it stuck
xG didn't take over because it was the most sophisticated metric anyone had built. There were and are more sophisticated ones. It took over because it was simple enough to put on a graphic, honest enough to actually mean something, and useful enough to change how decisions got made. That's a rare combination. Most stats are either too crude to be useful or too complicated to be cited. xG is in the narrow band where a number is interesting enough to argue about and clear enough to print.
The next generation of metrics — possession value, defensive contribution, on-ball pressure — are mostly in the "too complicated to be cited" bucket right now. One of them will probably make the same jump xG did sometime in the next decade, when somebody figures out how to render it as a single number that a viewer can absorb in two seconds. Until then, xG is the compromise. It's a flawed metric. It's also the most useful flawed metric in sports.