The 25-meter shot and the first misaligned brick of the xG model

On July 6, 2026, in Kazan, I sat in a rented apartment in Shanghai with three monitors open. The left monitor ran a model I had spent eleven months building. The center monitor displayed a moving heat map. The right monitor showed an odds board refreshing every second. Brazil had a cumulative xG of 2.19. Belgium had 0.87. My model said Brazil would win with a 68% probability. Twelve hours earlier, I had gone on air on a Chinese sports channel and declared Brazil would advance. The final score was 2-1 to Belgium. De Bruyne scored from a long-range shot outside the high-xG zone. Renato Augusto pulled one back with a header inside a zone my model labeled "low risk." After the final whistle, I deleted that model file. It was the second time in my career I had to rewrite from zero. Let me set the context before going into the details. World Cup 2026 was the first time a personal model of mine was commercialized. I was no longer an analyst sitting in a studio. I was paid to make predictions, and people bet according to those predictions. That changed the weight of every number on the spreadsheet. When a single data line could cost a stranger in Hanoi a month's salary, that data line was no longer data. It was a moral decision. My model at the time had three main pillars. The first was cumulative xG over the last ten games — a metric measuring the quality of scoring chances based on position, shot angle, and the number of defenders blocking. The second was PPDA, the number of passes allowed per defensive action, used to measure pressing intensity. The third was the average height of the defensive line. I once believed that a defensive line under 1.82 meters would struggle against teams that were good at heading from corners. That third pillar had helped me predict South Korea beating Germany 2-0 in the group stage less than two weeks earlier. I remember tweeting that prediction before the match kicked off, and someone called me a madman. When Germany actually lost, I said nothing. I just increased the weight of the "defensive height" variable by another 8%. That was the first mistake, and I only realized it three weeks after the Brazil-Belgium match. The problem was not the variable. The problem was that I believed I was predicting the result, when in reality I was only describing possibility. Look back at the decisive shot. De Bruyne struck from about 25 meters. According to the most modern model I re-ran for comparison, the xG of that shot sits around 0.04. That means in 100 shots from that position, only 4 go in. A model that only looks at the number would conclude: that was luck. Belgium won through a random moment. But I watched the replay of that shot about forty times. And what I realized did not lie in the shot itself. It lay in the three seconds before it. Before De Bruyne swung his leg, the Brazilian defense had done two things. First, they dropped back into a line 22 meters from goal and let a Belgian midfielder hold the ball in central midfield. Second, nobody pressed De Bruyne from behind for the entire three seconds. Three seconds. For a midfielder of his caliber, three seconds is a long afternoon. That is a type of error xG cannot see. xG sees shot position, shot angle, number of blockers. It does not see "time-space" — the interval a player is allowed to stand still and set himself before striking. My model had no variable measuring that. And because no variable measured it, I had assigned that shot a low probability. The truth is that the low probability lay in the shot, but the probability of the shot occurring was high. Those are two different stories, and my model merged them into one. The same thing happened with Renato Augusto's equalizer. The goal came from a header inside a zone my model classified as "low risk" because the density of Belgian defenders in the box was high. But when I reviewed it, I counted three Belgian defenders dropping back half a step late. That half-step was in none of the models I had ever run. That half-step is the difference between 0.12 xG and a real goal. I once thought xG was an objective measure. After the Kazan night, I understood that it is only objective in describing chances, not objective in describing people. A defender dropping back half a step late does not do it because he is poor. He drops back late because he is tired, because he is worrying about his partner on the flank, because he has already run 68 minutes, because inside his head there is a fear the spreadsheet cannot measure. That is the layer of data that gets dropped, and I had dropped it for eleven months while building the model. After the match, I argued fiercely on social media with a colleague. He said my model was useless. I said he did not understand statistics. Both of us were half right, and both of us were half wrong. What I did not say at the time was that I read the match's data log exactly 41 times over the following three weeks. I did it not to prove myself right. I did it to find which brick had shifted before the wall collapsed. The shifted brick was not the defensive-height variable. The shifted brick was the belief that past data can predict a moment that has never existed. The World Cup is a stage where pressure is compressed to the point that each player behaves differently from who he is in an ordinary match. A midfielder stands still for three seconds in the group stage. But in the round of 16, at Kazan stadium, under the pressure of a nation, he can stand still for seven seconds if nobody presses. My model was trained on group-stage data. It did not know that patience is also a variable of the tournament. This is where I want to stop and say plainly what many data analysts dare not say. Every model is wrong, but a few are wrong in a useful way. My model was wrong in the Brazil-Belgium match. But it was wrong in a very specific way, and that specific wrongness taught me more than the correctness of the South Korea-Germany match. People only read the result. Analysts read the deviation. xG does not score goals, but it makes people argue more than the actual ball. That is precisely its value, and precisely its risk. When a metric becomes a common language, people forget it is only a slice of reality. I had forgotten that for eleven months. I remembered it in three weeks. During those three weeks, I rewrote the source code. I added a "tournament factor" — a variable describing the degree of psychological compression at each stage. Group stage coefficient 1.0. Round of 16 coefficient 1.15. Quarterfinal coefficient 1.3. Semifinal coefficient 1.45. Final coefficient 1.6. The idea is crude, but a crude idea that is right beats a sophisticated model that is wrong. I also added another variable, which I called "defensive silence gap" — the average number of seconds an opposing midfielder is allowed to hold the ball without being pressed within the last 10 meters. That variable is not in any public API. I had to count it by eye. To be fair, I am not sure that variable would have helped me predict the Brazil-Belgium match correctly. I never re-ran it on that match, because I do not want to know the answer. There are tests where you do not need the result, only the knowledge that you asked the right question. Football stopped rolling in 2026, but randomness has never taken a lunch break. That is true of any model. Every spreadsheet is a meditation, except that after the meditation you lose money. And in this sport, people pay you to say out loud things you should only say to yourself at three in the morning. There is one question I still cannot answer after seven years. Whether the half-step late drop of that Belgian defender can be measured by modern skeletal-tracking technology. Technically, the answer is yes. But once you measure it, you must decide whether it is an independent or dependent variable. And that decision is not made by a computer. It is made by you, based on what you believe about people. That is why I do not believe in the era of pure data. I believe in the era of analysts who know they are guessing, and guess with discipline. The Brazil-Belgium match was not the last time I was wrong. It was just the first time I did not feel ashamed of being wrong. I learned to write a warning line under every analysis after that night: "The model is only a probability, not a prophecy." Not to protect myself. But to remind myself that any number can be judged by a shot from 25 meters. The next chapter of this story, I will write about how a basketball model of mine collapsed after only forty-eight hours of operation. But that is another story, and I noted at the end of this piece that I will return to it. Like every topic I abandoned before. Like a match without extra time that I once watched. Every model is a set of unfinished lines. And the most beautiful line is the one you have to delete the next day.