0.35 Beats 1.90: Re-reading Three Matches the xG Model Left Behind

The 53rd minute at Lusail, 22 November 2026. Salem Al-Dawsari collects the ball at the edge of Argentina's box, turns through two blue-and-white shirts and curls it into the far corner past Emiliano Martínez. In Shenzhen it is 3 a.m., and the spreadsheet in front of me has just produced a number I read three times: Saudi Arabia's xG at that moment, 0.35.

Argentina finished on 1.90. The scoreline finished 1-2.

I deleted my opening line four times that night. Publish the number alone and I turn a team's victory into a technical error of an algorithm. Hide the number and I betray the trade that has fed me. The only honest route left is to tell both stories at once, and to state plainly what the model saw and what it missed.

A model that never claims to be right

I built my first xG model in the summer of 2026, freshly 18 and a first-year student in Shenzhen. I pulled shot data from public stat sites, calculated it myself, got it wrong myself, corrected it myself. The France–Belgium semi-final at the 2026 World Cup was the first lesson. My model gave France 1.6 and Belgium 0.8; France won 1-0 through Samuel Umtiti's header from Antoine Griezmann's corner in the 51st minute. I spent a month rewatching footage, separating set pieces into their own bucket and adding weight to them. The error dropped. The bigger lesson sat elsewhere: my model measured the number of shots correctly and the value of a rehearsed routine incorrectly.

A short account of how I read xG. Every shot gets a probability of becoming a goal based on distance, angle, body part, how many defenders and the goalkeeper stand in the way, the pressure from the nearest marker, and where the ball came from: open play, corner, free kick or counter. Add it all up and you get an expectation of goals. That number depends on the model. For the same match, different providers typically differ by 10 to 30 percent because they define pressure and clear chances differently. I always print the model, the version and the margin of error next to any number I publish. That is why I open here instead of opening with a conclusion.

In 2026, when the pandemic emptied stadiums in China, I was a data analysis intern at a sports company in Shenzhen, collecting figures from 240 Chinese Super League matches. Home win rate fell from 47 percent to 39 percent. PPDA — the passes a team allows its opponent per defensive action — rose on average from 11.2 to 10.5, meaning sides pressed harder and scored less efficiently. Whether a stadium has a crowd or not, the match still needs someone to retell it. My internal report was later published on the company's news site, and I learned something that still holds: a number stripped of the crowd, the weather and the travel schedule turns very easily into a deliberate lie.

Three matches, three numbers, three readings

Match one, 10 July 2026 in Saint Petersburg. France 1-0 Belgium. Look only at xG and France edged it and won, which sounds reasonable. Separate the set pieces and the picture changes: Griezmann's corner was a prepared routine, Umtiti attacking the near post on a diagonal, Belgium's back line losing its man. In my recalibrated model, that header accounts for most of the gap in chance quality between the two sides. What decided the match sat in the corner column, not in the attacking line.

Match two, 22 November 2026 in Lusail. Argentina dominated possession, fired several times as many shots as Saudi Arabia, and had three goals struck off for offside in the first half. Saudi Arabia chose a high line and an offside trap they had drilled. Saleh Al-Shehri equalised in the 48th minute with a finish from a narrow angle, then Al-Dawsari scored the winner in the 53rd with a piece of individual skill. The 0.35 is the number, but the fight over naming it is the reality. In the same match, another provider put Saudi Arabia's xG at nearly double my model's figure because it weights blocked shots differently.

Match three, 26 June 2026 in Gelsenkirchen. Georgia 2-0 Portugal. In qualifying I calculated Georgia's average xGA as far lower than their standing as a first-time finalist suggested, even though they rarely held the ball. I wrote before the match that Georgia could spring a surprise, at roughly 70 percent confidence because their sample was small. They did. Khvicha Kvaratskhelia opened the scoring in the second minute from a counter, Georges Mikautadze doubled it from the penalty spot in the 57th. Portugal had already secured top spot and rotated, so I do not read this as an absolute measure of the two teams. I read it as evidence that a disciplined low block and a goalkeeper playing well can produce a larger swing than any gap in squad value.

Valuation deserves a look too. At Euro 2026 Portugal's squad value ran several times higher than Georgia's, and Georgia still won. In January 2026, Kvaratskhelia moved from Napoli to Paris Saint-Germain for a fee European media put around 70 million euros. Every transfer figure is a life converted into a price, and also a moment when the market quietly re-prices itself after watching a match.

Three matches, three shared traits. First, most of the xG gap comes from a small cluster of moments rather than spreading evenly across the game. Second, the value of a chance depends on whether it arrived from open play or from a rehearsed routine. Third, the winning side is not necessarily the side that created more quality in the model, but it is always the side that converted the highest-weighted moments. A model does not predict the winner; it measures who created better chances, and those are two different questions.

The counter-intuitive angle

I have to say plainly what many readers of my numbers do not want to hear. An xG of 0.35 does not mean Saudi Arabia won by luck, and an xG of 1.90 does not mean Argentina deserved to win. A model only answers the question it was built to answer. I built mine to measure chance quality. I did not build it to measure Mohammed Al-Owais's calm in the instant the ball left Al-Dawsari's boot, or to measure the shout from a back line that holds its offside line in the 80th minute. Data is a monastery, but I choose to walk out of the gate and look for football.

There is a second problem, and it is professional. The moment a number is published, the right to name it belongs to several parties at once: the data provider wants it to sell subscriptions, the broadcaster wants it to spark argument, the club wants it to defend them, and the fan wants it to confirm what they already believed. I have been accused of insulting a weaker team's victory, and I chose not to take the piece down. I chose to write a further analysis using tracking data to show why the favourite controlled the ball yet left gaps in two decisive moments. Defending a point with emotion is easy; defending it with data means accepting that the data itself can overturn you.

What I am tracking next

I keep circling the biggest blind spot in all three matches. Conversion from set pieces is the least properly modelled variable in the game, and it is exactly where underdogs collect most of their points. Next round I will watch three signals: first, how many counter-attacks each possession-conceding side converts per match; second, the average PPDA of the sides rated lower, since it reveals whether they dare to push up; third, each team's share of goals from set pieces, because the corner column usually answers before the attacking line does.

I do not build a spreadsheet for the match; I build a spreadsheet for the doubt. Every time my model is right, I write it down. Every time it is wrong, I write it down more carefully. And every time someone tells me I am insulting their team, I read it closely before replying.

xG does not lie, it simply never tells the whole truth. What remains belongs to the person sitting in front of a screen at three in the morning, knowing the number just calculated may not survive the second half.