Four Summers of Reading xG: The Limits of Data at Major Tournaments

In the summer of 2026 I sat in front of a computer screen in a small rented room in Shenzhen, just turned 18 and in my first year at university. My left hand held a page of notes; my right hand scrolled through a shot dataset I had assembled myself from public statistics pages. The World Cup semi-final between France and Belgium was played in Saint Petersburg on July 10, 2026. I calculated France's xG at about 1.6 and Belgium's at about 0.8. The final score was 1-0 to France, the only goal a header from Samuel Umtiti in the 51st minute after a corner.

I stayed two more hours after the final whistle, the keyboard clicking steadily in the quiet room. The spreadsheet was not wrong in its numbers, but it did not forecast the decisive goal. A set-piece had broken the chain of logic I had spent weeks building.

xG, expected goals, estimates the probability that a shot becomes a goal based on location, angle, shot type, the number of defenders and the situation leading to the shot. Event data providers collect hundreds of thousands of shots each season and assign weights through regression models. As a student I used free data and built my own weights by hand, which meant almost every situation was treated as roughly equal.

The largest error lay in the fact that Umtiti's headed corner was credited with only a very small addition, even though on the training ground it was a pattern rehearsed again and again. I spent a full month rewatching the footage, pausing at every phase, re-measuring distances and the angle of the run. I adjusted the model by adding weight to set pieces, especially corners in which an attacker runs across his marker before jumping.

The later version was more accurate, but I no longer believed a single number could close the story of a match. A model only describes what has been fed into the model. The rest of the match, including the decisive moment, sits outside it.

In 2026, when the pandemic emptied stadiums across China, I was a data analysis intern at a sports company in Shenzhen. I collected figures from 240 Chinese Super League matches. The home win rate fell from 47% to 39% when no spectators were present. The PPDA index, the number of passes a team allows its opponent per defensive action, dropped from 11.2 to 10.5.

My reading at the time: teams pressed harder because they no longer had to keep their emotional rhythm with the stands, yet their scoring efficiency fell. The sound that echoed out of empty stands through the public address system created an odd tempo, and I stood in the middle of an empty stadium and heard the background noise of football.

My internal report was soon published on the company's news page and drew attention from several local analysts. Since then I never separate numbers from match context. Every article carries a separate section describing the crowd, the weather, the travel schedule and the fixture density, because a number without context easily turns into a deliberate lie.

Whether a stadium has spectators or not, the match still needs someone to tell it.

In November 2026 I worked as a data assistant for an online sports outlet covering the World Cup in Qatar. On November 22, 2026, Saudi Arabia beat Argentina 2-1 in a match the winners barely controlled. I calculated Saudi Arabia's xG at 0.35 and Argentina's at 1.9.

The article was quickly attacked by a section of readers as an insult to the underdog's victory. I did not take it down. I wrote a follow-up using tracking and positional data, showing that Argentina held the ball but defended loosely in two decisive transitions, exposing the gap between centre-back and full-back exactly when it needed to be closed. That insistence earned me an invitation to work as an independent data expert for a European football magazine.

0.35 is a number, but the battle over what it names is the real story. The same data: one side calls it a victory of character, the other calls it a statistical accident.

At Euro 2026 I followed the Georgia national team for two weeks as the data reporter for that magazine. From qualifying data I calculated their average xGA at roughly 0.9 goals per match, among the lowest in the tournament, even though they rarely controlled possession. I wrote that Georgia would surprise Portugal. On June 26, 2026, they won 2-0 with two sharp counter-attacks. The post-match analysis was shared thousands of times, and a club in China approached me about a part-time data consultancy.

Based on my experience watching matches across four major tournaments, every analysis I write now follows a three-part structure: raw numbers, contextual analysis, and an anticipated rebuttal. The third part matters most, because it forces me to reread my own model through the eyes of someone who does not believe it.

I also learned to simplify complex metrics with hand-drawn graphics and to add a “why this number matters” section so that general readers are not left behind. A metric without an explanation is just a handsome line in an internal report.

One point strikes me as a tactical consequence more troubling than any technical error. Modern models are trained mainly on shots created from movements cutting inside, so they unintentionally undervalue wide dribbles and crosses. The result is that the inverted winger has become the default option in most academies, while the traditional touchline winger is pushed out of the development system before he can prove his value in situations the model does not measure. Data is not tactically neutral, because it is built from what people chose to record.

The follow-on effect sits in youth development. The academies of big clubs are talent warehouses rather than pathways; most of the young players inside them never reach the first team, and the share who genuinely make the step up stays below 10%. When data is used to rank 15-year-olds on the same set of metrics, the system tightens the very uniformity it created.

The rebuttal deserves to be stated plainly. My 240-match result from 2026 is a correlation, not proof of causation. The absence of crowds coincided with a compressed schedule, quarantine camps and a long pre-season break; any one of those could explain most of the eight-percentage-point drop in home advantage. Concluding quickly from a single piece of statistics is the fastest way to turn data analysis into clickbait.

Data analysis is moving into the dressing room with charts and models, while its conclusions often sit apart from the actual rhythm of a team. A model running on tens of thousands of shots knows nothing about a holding midfielder who has lost two nights of sleep to a feverish child, or a full-back playing through anxiety over a contract about to expire. xG does not lie; it simply never tells the whole truth.

I once sat in a meeting where every metric pointed one way and the head coach shook his head, because he had just rewatched the footage and seen a player mistime his run in three consecutive phases. Both sides had grounds. The dispute was not between data and instinct; it was between two different ways of observing the same event.

For the next round of major tournaments I will watch set-piece goal share first, because that is where models perform worst and where many knockout ties are decided. Alongside it, the gap between xG and actual goals for underrated teams; if that gap repeats across two consecutive tournaments, the problem most likely lies in the model rather than in luck. And finally, the quality of transition defence in midfield, something positional data can measure before the ball reaches the box.

Football does not live inside the spreadsheet; it lives between the cells. Every transfer figure converts a life into a number, and the person writing with data has a duty to remember that before touching the keyboard.

When the tournament closes, I keep one thing: data cannot measure the moment that silences an entire stadium.