xG as an Unreliable Witness: When the Model Was Right and I Still Could Not Sleep

Summer 2026, matchday 18 of the Chinese Super League. On my screen the expected-goals readout glowed: Shanghai SIPG 2.8 xG, Shandong Luneng 0.4 xG. I printed that line on an A4 sheet and taped it to the wall above my monitor. Before kickoff, almost every traditional pundit at the big sports papers picked a draw. They cited recent form, head-to-head history, the late-July heat, the away-day pressure in Jinan. I did not argue. I published the number, predicted SIPG would win 3-1, and signed my name under the analysis. Final score: 3-1. The piece drew 50,000 views in 24 hours. My editor called, his voice like he had just won the lottery. I sat still in front of the monitor, staring at the A4 sheet, and understood something more frightening than a model being wrong: the model was right. When a spreadsheet is confirmed by reality, people start to trust it. Once they trust it, it stops cross-examining itself. Back then xG was still new in China. Sports platforms imported the concept from Europe but mostly used it as decoration — printed on broadcast graphics while nobody re-priced a match from it. My team did the opposite. We took raw event data from providers and assigned our own weights by shot location, by shooter, by pressure around the box. Every match became a test: did the scoreboard match the shot quality? SIPG's squad then included Hulk and Oscar; for Shandong, Graziano Pellè played as the highest man in the attack. Those names made the numbers harder to ignore. My mental image of xG has barely changed since. xG does not score goals, but it makes people argue more than the actual ball does. It does not predict. It testifies. A shot from the edge of the box under pressure from two defenders is a witness that the attack created a chance, and also a witness that the defensive block was in the right place. The problem is that we keep making witnesses deliver the verdict. A trial cannot close just because the witness told a fluent story. In 2026 I told that story too fluently. The success of matchday 18 took me to a betting company as lead analyst. Summer 2026, the World Cup in Russia. My new model ran on PPDA and the average height of the back line. I do not remember every variable, but I remember the feeling when the group stage closed: the model was right in matches where nobody dared back the underdog. South Korea beat Germany 2-0. I posted it on social media and told everyone to follow the underdog. It felt like standing on a rooftop, watching a sleeping city below. By the round of 16, the model said Brazil would beat Belgium. Not because Brazil attacked better — because their defensive metrics were superior. I went on live television and said so. That night Belgium won 2-1, through a set piece and a late counterattack in the second half. Clients who listened to me lost money. I argued bitterly online with a colleague who had warned me correctly. Afterward I spent three weeks rewriting the code. Those three weeks taught me more than the previous three years combined. I added a 'competition' variable, because intensity and playing style in qualifying are nothing like the knockout rounds. I added a randomness component — not to make the model stronger, but to remind it that it is weak. Yet the biggest lesson sat outside the code. Working as an analyst for the Chinese market showed me how data migrates across borders and degrades: a metric born in England, used in Italy, trusted in China, then turned into a kind of superstition in Vietnam. At every station it picks up a layer of meaning the original author never sent along. Every model is wrong, but a few are wrong in a useful way. In Brazil versus Belgium, the model never predicted a scoreline. It priced the likelihood of scenarios. People wanted a prophecy, and the presenter — me — handed them a prophecy, even though on paper it was only probability. That was my fault, not mathematics'. There is a paradox attached. The more data people have, the more they want a single conclusion. After 2026, when global football went through a historic interruption, I wrote less about tactics and more about chance. Randomness cannot serve as an excuse for me to dodge analytical responsibility. It is a variable that must be declared before a model says anything at all. Football stopped rolling in 2026, but randomness has never taken a lunch break. From another angle, I have also learned to treat metrics as data with a shelf life. One season after birth, xG is still useful. Three seasons in, it begins to age. Five seasons in, it becomes a social convention. Many pundits still cite xG like a ten-year-old truth, while teams have changed how they create chances and defenders have changed how they clear the ball. Data going missing is not a loss of data — it is a kind of data. Sometimes it is the most suspicious kind. On injury and return, I carry a belief that is hard to shake: medical confidentiality keeps fans and media blind. Clubs disclose only what benefits the share price or the dressing-room mood. Every time a star is absent and the public information is boiled down to the word 'injury,' I treat it as a hidden variable, not an established fact. A spreadsheet may not lie, but whoever publishes the spreadsheet can. On youth development, I hold a persistent doubt too. Academies opened by former stars are mostly commercial stunts; what is desperately missing is systematic investment in grassroots coach education. A good data model cannot compensate for a generation of coaches with no foundation. This is the kind of variable xG never looks at, and the kind that decides who lifts the trophy ten years from now. All of this brings me back to a small question I have never fully answered: if I cannot predict accurately, why do I still do this every day? The answer probably lies elsewhere. People say I am good at predictions. Wrong. I am only good at saying 'not enough data' at the right moment. And that skill — saying a negative when everyone is waiting for a positive — is what I learned from my own wrong models, not from the right ones. Looking ahead, I am tracking something traditional metrics have not touched: the relationship between structured randomness and psychological volatility in penalty shootouts at major tournaments. The geometry of the goal is already well mapped; what is missing is data on fear. Perhaps next season I will try to put a variable for it into the spreadsheet. Every spreadsheet is a meditation, except that when you finish meditating, you have lost money. And perhaps, as usual, my model will be wrong again. If it is wrong in a useful way, that meditation was still worth sitting through.