The xG Model That Forgot Corners: A Lesson on the Limits of Data

July 2026, I sat in a library in Liverpool, reopening all 64 matches of the World Cup in Russia across three screens. Croatia had just reached the final. Weeks earlier I had written that Luka Modrić's side had advanced on a run of luck, and I was mocked exactly as expected. What kept me awake was not the criticism but the feeling that my model — the thing I believed was the most honest — was hiding a flaw. Every time I re-ran the data, the metric sat there, silent and confident. It took until the eleventh day for me to see it: my model had forgotten corners. Data whispers, and those who listen will hear the miracle — but only if they sit long enough to hear all of it.

The backdrop to this story begins a year earlier. In 2026, aged 42, I was working as a transfer-market administrator in Liverpool, watching Juergen Klopp take the club into the Premier League top four with 78 points on the back of manic pressing. I calculated Liverpool's average PPDA at 8.2 — the lowest in the league — while a conservative side like Manchester United sat at 15.7. PPDA measures the passes an opponent is allowed before each defensive action; the lower the number, the earlier and fiercer the pressure. I wrote a long piece on gegenpressing and published it on my personal blog, then was fiercely criticised for being "too mechanical". The 4-3 win over Manchester City on 19 January 2026 reinforced my belief: data does not lie, it only needs to be read correctly.

That belief was what I carried into Russia. I built a homemade xG model — Expected Goals, the probability a shot becomes a goal based on position, angle, type of delivery and defensive pressure. I ran it across every group-stage match, totalled it, and concluded France would win because their chance-creation index was the highest, averaging 2.4 xG per game. I also wrote that Croatia had "low xG but good fortune". The piece was widely read, and mocked just as widely when Modrić and his team-mates beat Denmark, Russia and then England to reach the final. I was emotionally drained, locked the library door, and spent two weeks reviewing my own data.

The error lay where I least expected. My model scored every shot from open play, but set pieces — corners in particular — were recorded only as a final shot, ignoring the chain of duels, blocks and second balls before it. Croatia scored a substantial share of their goals from exactly those situations. When I returned corners to the model and counted second balls, their xG rose by nearly a third. A model is not wrong where it measures, but where it does not — and what goes unmeasured is often what decides. I learned at Anfield that belief is itself a variable, and that variable appears in no spreadsheet I have ever built.

By March 2026, that variable came to collect. Aged 45, I watched world football stop because of the pandemic. Liverpool were 25 points clear of Manchester City, all but certain of the Premier League title, and the season was suspended. I lost faith quite deeply: if data cannot predict a pandemic, what does it mean? I wrote three drafts and deleted all three. When football returned in June with empty stadiums, I noticed something no model of mine had ever accounted for: the home-win rate fell from 46 per cent to 39 per cent. Crowds do not score goals, but their absence rewrote almost the entire weighting of home advantage.

I call this data context — the environmental conditions surrounding every metric, something a purely statistical table never declares on its own. The same PPDA of 8.2 means something different with 50,000 people in the ground and with none. The same 2.4 xG means something different when the opponent has run out of legs in extra time. An empty stadium does not distort data, but it makes the truth feel hollow.

In the transfer market, this lesson repeats at a far higher price. A player who has not played even 50 top-flight matches can be valued at 100 million euros, and that price is justified by beautiful metrics from a league two tiers below. I have sat in meetings where a data table was presented as scientific proof while the sample was only 900 minutes of football. The bubble in youth prices bursts not because the players are poor, but because people read the data without reading the sample size. Data makes no mistakes. The people reading it do.

The most counter-intuitive part of this whole story lies elsewhere. We tend to believe correlation is evidence of causation once the sample is large enough, but football rarely gives us a large enough sample. A season is 38 matches; a young player is a few hundred minutes; a knockout tie is 90 minutes that can be decided by a penalty in the 88th. In that space, even the strongest model is only a hypothesis presented neatly. And there is a kind of data we usually ignore: empty data. When I receive a dossier with no title, no information points, no entity to analyse, the most honest conclusion — and the hardest to write — is "not enough information to judge". Refusing to say so is the moment analysis turns into speculation dressed in jargon.

The same holds for VAR. The space for subjective judgement in VAR is far larger than television suggests, and the phrase "clear and obvious error" is itself an ambiguous clause. On the same incident, three video referees can draw three different lines of infringement, and each can justify it with their own frame. Technology does not erase subjectivity; it merely moves it into a more closed room.

So what do I take from thirty-five years of watching this industry? Not that data is useless — quite the opposite. Precisely because data is powerful, the responsibility of whoever holds it is greater. Every number in a transfer table is a life waiting to be written onward, and whoever writes it should remember that behind it stands a twenty-year-old player, a family, a career that can collapse because a model misread its sample size.

Those who are right before their time always pay in loneliness. I once thought that loneliness was the price of seeing ahead. Now I think differently: it is the price of sitting long enough to check whether you truly saw, or were only staring at a model that had forgotten the corners. Next season, when a young player is priced on a sample of nine hundred minutes, I will open my spreadsheet again — and this time, I will check first what it does not measure.