Two scenarios. Hold them side by side; they will do most of the work in this article.
The first is a striker. Over a full league season, he takes 140 shots. The cumulative xG of those shots is 18.4. He scores 26 — outperforming his xG by nearly eight goals. Across his three previous seasons, the same gap appears, smaller in magnitude but consistent in direction. The model says he is being lucky. The eye says he is finishing better than the average finisher would from those positions. The analyst’s job is to decide which is true, and how much.
The second is a single match. Team A wins 1-0. Their xG is 0.4. Team B’s is 1.8. The losing manager goes on television and says they were “the better team”, pointing at the xG number. The winning manager says football is about goals. A pundit says xG is “made up”. None of the three statements is correct, but each contains a partial truth the other two are ignoring.
Both scenarios involve the same metric. In the first, xG is doing what it was designed to do — separating chance quality from finishing outcome over a sample large enough to mean something. In the second, it is being asked to do something it cannot do, by people who do not know that it cannot.
Most public arguments about Expected Goals are arguments of the second type. This piece is an attempt to put that right.
The Origins Story
No metric arrives fully formed. xG has a longer prehistory than most of its current advocates realise, and understanding that history matters — not for nostalgia, but because it explains why the metric looks the way it does, where the gaps in it came from, and why it took so long to become a thing an ordinary football supporter might encounter on a Saturday evening broadcast.
The first person to attempt something like systematic shot analysis in football was Charles Reep, a Royal Air Force wing commander who attended a Swindon Town match in 1950 and decided, with the particular conviction of a man who had never done something before, to write down everything that happened. Over the following decades, Reep logged roughly 2,000 matches by hand, producing the largest individual football dataset that had ever existed and drawing from it a series of conclusions that were wrong in almost every direction. His central finding — that the majority of goals came from sequences of three passes or fewer, and that teams should therefore pursue rapid direct football and avoid the build-up — was a piece of survivorship thinking so complete it might serve as a textbook example of the error. He had measured where goals came from without adjusting for how many attacks of each length occurred in the first place. Most attacks are short. Most goals therefore come from short attacks. This tells you approximately nothing about the efficiency of different styles of play.
Reep was not doing what xG does. He was not assigning probabilities to individual shots. But the impulse — to replace intuition about football with counted, logged, structured data — is the same impulse, and the errors he made are errors the field is still working to correct. His influence on British coaching was extensive and, in retrospect, largely harmful; Graham Taylor and Charles Hughes both drew on his work when arguing that direct football was statistically superior. The irony is that a flawed attempt at empiricism produced some of the most anti-empirical coaching orthodoxy the game has seen.
The modern xG story begins in earnest around 2012, at a considerable remove from Reep’s notepads. Sam Green, working at Opta, published a paper at the MIT Sloan Sports Analytics Conference that year presenting a model for predicting shots on a probabilistic basis using location, body part, and assist type. Green was not the only person thinking along these lines — similar work was happening at academic sports science departments in the UK and Scandinavia — but Opta’s paper marked the point at which the methodology had commercial infrastructure behind it. The data company had the match logs. The model gave those logs a new use.
What made it a public phenomenon rather than an internal tool was the football analytics blogosphere, which in 2012-14 was at the height of its creative energy. Michael Caley, writing at Cartilage Free Captain and later his own site, published accessible explanations of how xG models worked and, crucially, the team-level outputs of his own model run on publicly available data. Caley’s work was precise without being exclusionary — he showed his assumptions, acknowledged the limitations, and distinguished clearly between what the numbers could and could not support. That combination of intellectual honesty and accessibility set a standard the field has been trying to maintain since, with varying success.
The phrase “expected goals” appeared in academic sports science literature before it landed in football analytics. Researchers studying basketball and ice hockey had been using “expected value” frameworks for offensive possessions from at least the mid-2000s, and the language filtered across. What football analytics did was adapt the framework to a sport where the discrete shooting event was relatively rare and the game state between events was continuous, complicated and poorly understood. That adaptation required patience with small samples. A basketball team might take 80 shots per game. A football team takes 12. The statistical patience required is an order of magnitude longer.
The cultural readiness for this kind of metric owed something to baseball. Moneyball — Michael Lewis’s 2003 book about the Oakland Athletics’ use of sabermetric analysis — had done for quantitative sports analysis what a successful product launch does for a market category: it proved there was an audience. The book argued, somewhat simplistically but not wrongly, that traditional scouting had systematic blind spots and that careful attention to undervalued statistics could give an underfunded team competitive advantages against richer rivals. When Moneyball became a film in 2011, with Brad Pitt playing Billy Beane and Jonah Hill playing a character who is, essentially, an analyst with spreadsheets, the message reached an audience that had never read the book. Football was slower to adopt the approach than baseball, partly because of the game’s greater complexity, partly because the commercial data infrastructure took longer to mature, and partly because football’s traditional scouting culture was — and in places remains — more resistant to quantification than its baseball equivalent.
The shift from simple location-based models to the multi-variable models in use today was driven by data availability rather than theoretical breakthrough. The early public models — including Caley’s — worked with distance, angle, and body part because those were the features the public data contained. As event data became richer and as companies like StatsBomb began collecting information that competitors were not — freeze-frame snapshots, goalkeeper positioning, defensive pressure recorded at the moment of the shot — the models gained new variables and the probabilities became correspondingly more reliable. The current generation of models is not a different concept from Green’s 2012 paper; it is the same concept with orders of magnitude more training data and considerably more variables.
Why football specifically, rather than rugby or cricket or another field sport? Several reasons converge. Football is the world’s most watched sport, which means the commercial incentive to understand it quantitatively is immense. The discrete shot event is cleanly identifiable in event data, unlike the continuous movement that comprises most of a match. And football’s low-scoring nature makes goal probability an unusually interesting quantity — in a sport where most matches are decided by one or two goals, the difference between a 0.12 xG chance and a 0.35 xG chance is often the whole story. In rugby or basketball, any individual scoring opportunity is a smaller fraction of the total score; the marginal value of a precise probability estimate per action is lower. In football, xG per shot is practically the unit of competitive advantage.
What xG Actually Is
The formal definition is unromantic. Expected Goals is a number between 0 and 1 assigned to a shot, representing the probability that a shot of that type, from that position, in that situation, would historically have been scored. A shot worth 0.30 xG is one that, in the data the model was trained on, was converted thirty per cent of the time.
The number is produced by a statistical model — usually a logistic regression or a gradient-boosted tree, depending on the provider — trained on hundreds of thousands or millions of historical shots, each labelled with whether or not it became a goal. The model learns the relationship between the features of a shot and the probability of it being scored. When a new shot is taken, the same features are passed through the model and a probability comes out. That probability is the xG of the shot.
The implication that follows from this is the one most readers fail to internalise. xG is not a measure of how good a chance was. It is a measure of how good chances of that type have been, on average, in the past. It is a historical average dressed up as a current judgement. If the data was good and the sample was large, that average is informative. If either is not, it is not.
This is also why xG is a probability, not a verdict. A 0.7 xG shot is a very good chance, but it does not become a goal seven times out of ten because some specific player is meant to score it seven times out of ten. It becomes a goal seven times out of ten because, in the training data, the population of shots looking roughly like this one had that conversion rate.
How It Is Calculated
The features that go into a modern xG model are the inputs an analyst can extract from event data — the structured logs that companies like Opta, StatsBomb, Wyscout and others produce from every professional match.
The standard inputs are well-established and broadly common across providers. Distance from goal and shot angle — the angular size of the goal as seen from the shooter’s position — do most of the work. Body part matters: a header is worth less than a foot shot from the same location, because headers are harder to direct and harder to hit cleanly. The type of assist matters — a shot from a cutback or pull-back is worth more than a shot from the same location after a cross, because cutbacks tend to leave the goalkeeper out of position. Whether the shot followed a through-ball, a set piece, a counter-attack or open play is included for the same reason.
The differences between providers begin where the inputs end. Opta’s published methodology describes a model trained on roughly one million shots, considering more than twenty variables. StatsBomb’s model is trained on a smaller volume of data, but on richer data, because StatsBomb collects information other providers do not.
The most important of those is the freeze-frame. From 2018 onwards, StatsBomb began recording the position of every visible player at the moment a shot is taken, including the goalkeeper. That allows their xG model to know how many defenders are between the shooter and the goal, where the goalkeeper is, whether the shot is under pressure, and — uniquely — the height at which the shot is struck. A daisy-cutter from the edge of the box is a different shot from a half-volley that bounces up to chest height; freeze-frame data lets the model know which it is.
This is also why public xG numbers from different providers do not match. A shot logged by Opta at 0.18 xG might be 0.14 in the data feed’s model and 0.22 in StatsBomb’s. The shot is the same. The training data, the features and the model are not. A reader who quotes one provider’s xG against another’s is comparing two related but distinct measurements.
There is also npxG — non-penalty Expected Goals — which strips penalties out of the total. Penalties are scored at a rate of around 75–80 per cent and carry an xG of roughly 0.76 to 0.79 depending on the provider. Including them in a striker’s seasonal xG flatters anyone who takes them, and tells you very little about open-play creation. For player-level analysis, npxG is almost always the more useful number.
What xG Does Well
The first thing to understand about a probabilistic model is that its outputs are reliable in aggregate and noisy in the individual. xG is no exception. Over a single match, it can be wildly misleading. Over a season, it is one of the most stable signals in football analytics.
The squad-level use is the cleanest. A team that consistently generates 1.6 xG per match and concedes 0.9 is, with high probability, a better attacking and defensive side than one generating 1.1 and conceding 1.4, regardless of where they sit in the table on a given Saturday. Tables lie because of fixture order, refereeing, injuries and finishing variance. xG-for and xG-against, averaged over thirty-eight matches, lie much less. Ben Torvaney’s work suggests a rolling window of around ten matches is the point at which xG becomes a usefully predictive signal; below that, sample noise dominates.
The second clean use is finishing assessment, with caveats. Over a single season, the gap between a player’s goals and his npxG is mostly noise. Over three or four seasons, it is partially signal. A striker who consistently outperforms his expected goals by a meaningful margin — Erling Haaland and Mohamed Salah are the standard contemporary examples — is doing something the model is not capturing. Some of that is shot placement (Haaland’s tendency to put shots into the bottom corners is well-documented in Opta’s analytical writing), some shot selection, some composure under pressure that no event-level model can quantify. The model is wrong about him in a consistent direction, and that consistency is the evidence that the wrongness is not random.
The third use is defensive shot suppression. Teams that concede few high-xG shots are, with very high reliability, well-organised defensive sides. xG-against quietly identifies sides whose defensive numbers look better than they should because of goalkeeping form, and sides whose numbers look worse than they should because a goalkeeper is bailing them out of structural problems. xG against does not care about saves. That is a feature, not a bug.
Post-Shot xG: After the Ball Leaves the Boot
Everything described above concerns pre-shot xG: the probability assigned to a chance based on who is taking the shot, from where, and under what conditions, before the ball is actually struck. It is calculated at the moment of the shot, and frozen there. What happens in the next fraction of a second — where exactly the ball goes, how hard it is hit, the flight path that results — is not included in the pre-shot number. Post-shot xG, often written PSxG, is an attempt to close that gap.
The distinction is conceptually clean. Pre-shot xG tells you how valuable the situation was before the shooter made contact. Post-shot xG tells you how likely a goal was given the specific shot that was actually produced — including its placement, trajectory, and speed. A striker standing six yards out in the centre of the goal has a pre-shot xG of roughly 0.50. If he guides the ball into the bottom-left corner at pace, the post-shot xG of the resulting shot is higher, because bottom-corner shots at that pace convert at a higher rate than the average shot from that position. If he hits it directly at the goalkeeper’s chest from the same spot, the post-shot xG collapses toward zero. The location has not changed. The shot has.
The practical application that has done most to bring PSxG to public attention is goalkeeper evaluation. Traditional shot-stopping metrics — save percentage, saves per game, clean sheets — suffer from a fundamental problem: they make no allowance for how difficult the shots being faced actually were. A goalkeeper who faces twelve easy shots and saves eleven of them has a better save percentage than one who faces twelve genuinely dangerous shots and saves nine. The first goalkeeper may actually be the inferior one. PSxG provides the correction. If a goalkeeper concedes 0.9 goals in a match where the opposition’s post-shot expected goals total is 1.4, he has saved 0.5 goals above what a league-average goalkeeper would have expected to concede — a meaningful positive contribution. The inverse identifies goalkeepers who are performing below the difficulty of the shots they face: conceding 1.6 from a PSxG of 1.4 is being saved by nothing.
StatsBomb formalised this into a metric called PSxG+/-, where the sign indicates whether the goalkeeper is saving more or fewer goals than expected from the quality of shots faced. A positive PSxG+/- means goals saved above expectation; a negative means the opposite. Over a season, a goalkeeper with a PSxG+/- of +8 has, on the model’s best estimate, prevented eight goals that an average goalkeeper at that club would have conceded. That is an enormously useful number for a recruitment department trying to decide whether an expensive goalkeeper is worth his transfer fee, or whether a cheap one is disguising structural defensive problems that will recur regardless of who stands between the posts.
The metric also illuminates why different elite goalkeepers rate differently under different systems. Alisson Becker’s PSxG numbers have historically reflected exceptional shot-stopping on genuine saves — the saves he makes tend to be from shots the model considers more dangerous than average. Manuel Neuer’s PSxG analysis, particularly across his mid-career seasons, showed something slightly different: elite positioning that meant he faced fewer genuinely dangerous shots in the first place, suppressing the raw PSxG total he needed to beat. Ter Stegen, at his peak, showed strong PSxG+/- driven partly by his ability to hold or parry cleanly on shots the model expected to be converted by rebound. None of these distinctions appears in a save percentage column.
The shot-placement research behind PSxG also produces findings that intuitively make sense once you see the numbers. Bottom corners are worth substantially more than location-only models suggest. A shot aimed at either bottom corner from inside the box converts at a meaningfully higher rate than the position alone predicts, because goalkeepers can close off straight-ahead shots with their starting position but must move laterally to cover the corners, and that movement has a time cost. Top corners are more variable: the difficulty of reaching them cuts both ways, making them riskier for the shooter but also harder for the keeper. Shots at goalkeeper height — the area the model traditionally treats simply as “on target” — convert least often despite sometimes appearing to offer the most of the goal. The placement data reveals that the frame of the goal is not a uniform target; it has topology, and PSxG is the metric that begins to describe it.
What xG Misses
The list of things xG does not measure is longer than the list of things it does, and the misuses follow directly from misunderstanding the omissions.
It does not measure the run that created the chance. xG is calculated at the moment the shot is taken; the build-up that produced the shooting position is invisible to it. A striker who consistently arrives in 0.4 xG positions is not measured by xG as a creator of those positions, even though arriving there is the skill.
It does not measure pressure on the shot, except where freeze-frame data is available. Most public xG numbers — including those on broadcast graphics and most aggregator sites — come from providers without freeze-frame data. A free header in the six-yard box and a header with a defender’s shoulder in the shooter’s face score identically in those models if the location and assist are the same. They are obviously not identical shots.
It does not measure game state. A team trailing 3-0 in the eighty-fifth minute against a low block will accumulate shots — frequently bad shots, hopeful shots — that inflate xG without representing real attacking threat. A 1.8 xG produced by twenty-two shots, most from outside the box and most after the result is settled, is a different number from a 1.8 xG produced by six clear chances. The post-match graphic does not correct for any of this.
It does not handle rebounds cleanly. xG is calculated per shot. A shot that produces a rebound which is then converted is two shots, scored separately. The causal link between them is not in the data.
It does not, except in the most data-rich models, account for defender positioning between shot and goal. A shot through a packed six-yard box and a shot through clean air score similarly in basic models. They are not similar shots.
And it does not, by definition, capture anything about the carrier of the ball — dribbling, line-breaking passes, off-ball runs that pull defenders out of position. Metrics that try to fill these gaps exist. xG does not pretend to be them.
Reading It as a Reader
Three heuristics, in descending order of usefulness.
First: the larger the sample, the more the number is worth. xG over a season is meaningful. xG over ten matches is suggestive. xG over a single match is, in most cases, decoration. If a single-match xG line is the centre of an argument, the argument is probably weaker than its proponent thinks.
Second: separate xG from npxG. Player-level discussion almost always wants npxG. Team-level discussion can use either, but consistency matters.
Third: notice which provider produced the number. Discrepancies between Opta, the data feed, StatsBomb and FBref are not large enough to overturn most conclusions, but they are large enough that quoting one provider’s xG to refute another’s is mildly incoherent.
A useful habit, when reading a single-match xG line, is to ask how many shots produced the total. A team with 1.5 xG from three shots is in a different state from one with 1.5 xG from twenty. The first has had clear chances; the second has been kept at arm’s length. Most public graphics do not distinguish.
How Clubs Actually Use xG: The Dressing Room Reality
There is a version of xG that exists in public — on broadcast graphics, in podcasts, on aggregator dashboards — and there is a rather different version that exists inside professional clubs. The two are related, but the gap between them is wider than most outside the industry appreciate. Understanding how serious clubs actually deploy xG is useful both for evaluating the metric’s real capabilities and for calibrating scepticism when you hear a club executive dismissing analytics.
The most established professional application is striker recruitment, specifically using npxG evaluated over rolling 38-match windows. A recruitment department that looks only at a striker’s goal tally is, in effect, accepting the noise of finishing variance as signal. A striker who scored fifteen goals on 8.2 npxG is almost certainly a better forward than one who scored the same number on 14.8 npxG — the first has demonstrated clinical conversion above model expectation; the second is performing at roughly the level any league-average finisher would from his positions, and may regress. Viewing player performance over 38 rolling matches, roughly one league season’s equivalent, smooths most of the single-season variance. The best recruitment departments look at 76 matches or more where the data exists — two full seasons — because the multi-season consistency signal is the one that actually predicts transfer value. A forward who consistently outperforms his npxG by 20 per cent over two or three years is carrying genuine finishing skill that the model cannot quantify; that skill has a price, and a good recruitment team knows roughly what it is.
Defensive analysis uses xG-against not as a summary verdict on a back four but as a diagnostic map. If a club’s xG-against is concentrated in a particular zone — say, consistently generating high-quality chances from the right channel — the coaching staff has a structural problem locatable in space, not just a vague sense that defending is poor. Analysts will break the xG-against down by origin: how many high-xG chances are arriving from crosses? From transitions? From set pieces? The answers point to specific tactical vulnerabilities that can be addressed in the training week. A manager looking at a team’s xG-against profile across twelve matches is, in effect, looking at a heat-map of where their defensive system is failing, zone by zone.
In-game use has grown significantly since tablet technology became routine on benches. Several clubs now receive live xG accumulation data during matches — an update at fifteen-minute intervals showing not just what has happened but how the game’s chance-quality balance is trending. If a club has conceded 0.8 xG in the first forty-five minutes from just three shots, their analysts know they are conceding high-quality positions even if the scoreline looks manageable. That information can directly influence half-time tactical adjustments or substitution timing in the second half. The limitation, which any honest analyst acknowledges, is that in-game sample sizes are tiny; a 0.8 xG half could mean structural danger or could be the product of two set-piece deflections and one goalkeeper error of positioning. The in-game number is a prompt to think harder about what you are watching, not a verdict on it.
The commercial application with the highest leverage is the regression-to-mean trade. The basic logic: a striker who has scored fifteen goals on 10 npxG is likely to score fewer goals next season, not because he will become a worse player, but because some of his overperformance was variance. A striker who scored eight goals on 13 npxG is likely to score more — the positions he is reaching are good ones, and his finishing will, with reasonable probability, normalise upward. Clubs that understand this can buy the second type of striker cheaply, because the market looks at goal tallies. They can sell the first type at a premium, for the same reason. This is not a reliable enough signal to execute in every case, but applied across a portfolio of recruitment decisions, it shifts the expected value in the buying club’s favour.
Brighton and Brentford were the clubs that first made this strategy legible to English football’s mainstream audience. Brighton’s recruitment operation under Tony Bloom and the club’s data team produced a consistent pattern: forwards acquired with output below their npxG, who then normalised upward under Graham Potter and Roberto De Zerbi’s systems that consistently generated high-quality shooting positions. Brentford’s approach under Matthew Benham went further — the Smartodds methodology that underpinned the club’s recruitment was among the earliest to treat npxG-over-expected as a primary transfer filter rather than a supporting variable. Both clubs operated near the bottom of Premier League budgets while consistently punching above their financial weight in recruitment efficiency.
The tension that thoughtful people inside clubs describe is not between xG and traditional scouting, but between the two used badly. Coaches who ignore the data entirely and scouts who substitute it for watching the player — both produce worse decisions than the analyst and scout working in the same room, with the same brief, and a shared understanding of what each tool is for. xG tells you what a striker has been doing with the chances he has reached; it does not tell you whether his movement to create those chances is replicable against different defensive systems, or whether his character will hold up in a relegation fight. The scout’s eye addresses those questions. Neither is dispensable.
The dangerous middle ground — and it is populated by clubs at every level — is the half-informed application. A club that has heard of xG, has purchased access to a data platform, and has assigned a young analyst to produce weekly xG reports without the infrastructure, the football knowledge, or the seniority to have those reports taken seriously in a recruitment meeting is not using xG; it is producing a performance of analytics that gives the club the self-image of a data-driven operation without any of the structural advantages one actually produces. This pattern is common enough that analysts who move between clubs have a term for it: dashboard theatre.
What Comes Next
The frontier of public xG is moving in three directions, all of them sensible.
The first is freeze-frame and tracking-derived xG, where StatsBomb is furthest ahead but not alone. By 2026 it is no longer controversial to claim that defender positions, goalkeeper positions and shot height should be in the model; the only question is how to acquire the data. StatsBomb 360, launched in 2021, formalised the freeze-frame product into a continuous tracking-style dataset, and the underlying xG model has been upgraded accordingly.
The second is the family of derivative metrics that try to credit the build-up. xG Chains credit every player involved in a possession that ends in a shot with the xG of that shot. xG Buildup does the same but excludes the assist and the shot itself, surfacing the contributions of midfielders and defenders whose work disappears under traditional goals-and-assists accounting. They are useful tools, with the caveat that they distribute credit by participation, not by causal contribution.
The third, less mature but coming, is contextual xG that adjusts for game state, scoreline, and the defensive shape against which a chance is created. The early work is encouraging; the direction is clear. The next generation of public xG will know that a shot in the eighty-fifth minute at 3-0 down is not the same shot it would have been at 0-0.
Beyond those three directions lie the metrics that move away from xG’s shot-centric framework entirely. Expected Threat, developed publicly by Karun Singh and formalised in academic work thereafter, is the most accessible of these. Where xG measures the probability of a shot becoming a goal, xT measures the value of moving the ball to any given position on the pitch — a unit derived from how often ball possession in each zone of the field leads to a goal within the subsequent few actions. A midfielder who regularly carries the ball from the midfield third into the opposition half in a way that increases the team’s threat is, under xT, receiving credit for those carries. Under xG, he is invisible until someone shoots. The distinction sounds simple and its implications are significant: xT begins to quantify the progression and transition work that creates the situations xG then measures.
Further from the mainstream is the VAEP framework — Valuing Actions by Estimating Probabilities — developed by Tom Decroos, Lotte Bransen, Jan Van Haaren and Jesse Davis. VAEP attempts something considerably more ambitious than xG or xT: it assigns a value to every on-ball action in a match, not just shots or carries. A tackle, a header, a short pass that moves the ball five metres — each is given an expected contribution to scoring and an expected contribution to conceding, and the difference between these is the action’s value. The theoretical appeal is obvious. Every action in football has a consequence for the game state, and a framework that values all of them rather than just shooting events is closer in principle to a complete description of what is happening. The practical limitations are also obvious: VAEP requires rich tracking data to produce reliable values for low-frequency actions, and the model complexity means that interpreting outputs requires more statistical sophistication than most broadcast contexts can accommodate.
Physical and GPS tracking data is beginning to be integrated with xG in ways that will eventually change the concept significantly. Running load, pressing intensity measured in metres per second per player, defensive line speed — these physical metrics, now routinely collected by clubs using wearable technology and optical tracking systems, can in principle be used to contextualise the xG numbers produced in a given match. A team that has run fifteen per cent more than their season average and is conceding high-xG chances in the seventieth minute may be conceding them partly because of accumulated physical fatigue, not purely structural organisation. That contextualisation is available to clubs with the infrastructure to combine the data feeds; it is not yet available in the public-facing models.
The concept of expected goals added for non-shooting actions — sometimes framed as expected goals contribution for individual players — is the direction several data companies are moving their public-facing products toward. The logic is to express every significant on-ball action in the same unit as xG, making the number intuitively legible to audiences who have learned to think in those terms. A dribble that moves the ball from outside the box to inside it represents an increase in the team’s probability of scoring a goal on this possession; that increase can, with a sufficient model, be expressed as an xG addition. A key pass that creates a chance worth 0.22 xG is credited to the assisting player in that unit. A defensive intervention that reduces the opponent’s probability of scoring from 0.30 to 0.05 represents a 0.25 xG subtraction from the threat — credited to the defender. Whether the underlying models are reliable enough yet to support that level of granularity per action is a live methodological debate; the direction is clear enough that the debate will eventually resolve in practice rather than in theory.
The furthest horizon, and the one most likely to obsolete large portions of the current xG ecosystem within a decade, is the application of AI vision systems to match footage. Current event data is collected by human taggers or semi-automated systems watching footage and logging events: this was a shot, this was a tackle, this was a pass. The freeze-frame data that gives StatsBomb’s xG model its advantage is still, in the end, a snapshot — a still image of player positions at one discrete moment. Computer vision systems trained on full match footage can in principle track every player continuously throughout a match, producing not snapshots but trajectories: a complete description of where every player was at every moment of the game, at resolution no human tagger can achieve. When that data is cheap enough and accurate enough to underpin xG models at scale, the distinction between pre-shot and post-shot xG may collapse; the model will know not just where the goalkeeper was at the moment of the shot, but how he had been moving in the preceding seconds, and what that movement implied about where he could reach. The models will know not just that there were three defenders between the shooter and the goal, but how fast each was moving, in which direction, and what that implied about the probability of their intervention. That version of xG has not been built yet. But the components exist, and several companies are assembling them.
A Closing Note
xG is a single number, and the single-number summary is a particular kind of intellectual object. It is a compression. It throws away most of what happened in order to produce something portable, comparable, and quickly readable. That is its strength and, inevitably, its weakness.
The mistake is not to use xG. It is to forget you are using a compressed representation, and to argue with someone else’s compressed representation as if it were the thing itself. xG is the best widely-available single-number summary of chance quality football has ever had. It is a long way from being a complete description of any match it appears in, and a longer way still from being a substitute for watching the game.
The honest position is the boring one. Use xG as it was designed: at scale, as a sanity check on what you think you have seen, with awareness of its provider and its sample size. Distrust anyone who deploys it as a single-match weapon, and distrust equally anyone who dismisses it because they once saw it disagree with the scoreline. Both errors are the same error, from opposite directions.
The model is not the match. But the match, watched without the model, is also a less informative object than its viewers tend to assume. The job is to hold both at once.