When the Data Label Lies: A Lesson from a Football File with No Football

On August 13, 2026, I opened a stage-one analysis file on my workstation. The domain label read clearly: football. Beneath it were 25 information points, an author-stance statement, and a line recording the article's purpose: to inform. I read all 25 points. Then I read them a second time. Then a third.

No club. No league. No player. No coach. No match.

The original headline concerned the last surviving newborn from a fire at PIMS Hospital in Pakistan, who had died. The twenty-five information points described a blaze, deaths, medical complications, and a death certificate. I sat still for a long while in front of that screen. Spectators watch the play; I watch 22 indices moving — and wait patiently for them to tell a different story. This time, there were no indices to watch at all.

Context: the upstream valve

The analysis system I operate runs in several stages. Stage one reads the source article, extracts information points, and assigns a domain label. That label becomes the input for stage two — where I run nine analytical dimensions: tactics, club finance, the transfer market, results and public-opinion cycles, league landscape, rules and governance, the dressing room, risk profile, and industry transmission chains.

Every one of those dimensions assumes a single premise: that there is football in front of it.

The domain label is the upstream valve of the whole pipeline. When that valve opens in the wrong direction, everything downstream flows the wrong way — but it flows wrongly in silence, because no meter alarms when water runs down the wrong channel.

I have spent many years as a data gatekeeper in this trade. I once sat breaking down every single passage of play across 14 V.League clubs just to build an xG table by hand, before the Vietnamese market knew what xG was. The first xG table I wrote by hand on a bus, back when nobody called it data. That experience taught me something no data school teaches: mislabel one cell at the source, and the entire spreadsheet behind it becomes a desert.

Nine out of nine empty results

What I saw in that file was not a model error. It was a labeller error.

All nine stage-two analytical dimensions returned empty. Not because the model was weak. But because there was nothing in front of it to analyse. The tactical dimension returned N/A. The financial dimension returned N/A. The governance dimension returned N/A. The risk dimension returned N/A. Nine out of nine. A technically perfect result, and a meaninglessly catastrophic one.

The model did not crash. It simply went quiet. And that silence is the most trustworthy signal I have ever received from a machine.

Let me be clear about why that matters to anyone working in football.

Had that file landed on a system without a validation valve, the outcome would have been entirely different. The machine would not have said “I have no data”. The greatest fear in analytical work is not a wrong model, but a confident one. When input data carries the wrong label, the model still runs. It still produces twelve pages of conclusions. It still talks about pressing intensity, wage structures, injury risk. Except that all of it is built out of thin air.

In the xG table I built myself for V.League's 2026 season, I recorded Phan Van Duc — then twenty years old, a winger for SLNA — posting 0.48 expected goals per match, higher than the average of foreign strikers in the league. He scored only five goals. Many called me a man deluded by numbers. But what I grasped was not the five goals. What I grasped was where he appeared, the quality of his finishes, and how often he touched the ball in dangerous areas. In 2026, he scored the decisive goal at the AFF Cup.

That prediction only held because the input data was correct. Every passage was assigned to the right match, the right player, the right minute. Had I misassigned a single SLNA match to another club, the whole xG table would have looked beautiful and been wrong as a dream.

The same principle applied at the 2026 World Cup in Russia. I used PPDA — a measure of pressing intensity adjusted for the passes an opponent is allowed — and found Croatia under Zlatko Dalic posting a PPDA of 7.9 against Argentina, with Luka Modric as the hub of midfield. That was lower even than Spain, the side famed as masters of possession. Croatia pressed head-on with extraordinary efficiency against a woeful Argentine defence, across nearly forty percent of contested time. I wrote a long piece predicting Croatia would reach the final, and placed my view on the scales. The result: they beat Argentina, Russia and England in turn.

But that prediction was not good because I was brave. It was good because the input data was clean. The world saw Croatia as an underdog; I saw them as a string of coefficients nobody had dared exploit. That string of coefficients exists only when someone sits down and classifies each match, each opponent, each context correctly.

By 2026, when the pandemic wiped out the fixture list, I spent six months excavating V.League data from 2026 to 2026. In 2026 the stadiums were empty, but every ball still fell into a cell of the model, and I understood that data never keeps company with a pandemic. I found a pattern: clubs that changed presidents mid-season saw their win rate fall by as much as 23% over the following five matches, owing to governance disruption. After that five-part retrospective series ran, a club executive phoned to thank me for helping them avoid sacking their head coach at a sensitive moment. That pattern only had value because every presidential change was logged on the right date, every win and loss assigned to the right season, every club correctly identified.

Transfer-window noise is also a form of mislabelling

In a transfer window, mislabelling occurs densely. A rumour is labelled “deal done” when behind it lies a single short post. A negotiation is labelled “about to sign” when in fact there has only been one phone call. Transfer-window noise drowns out signal, and the only way to filter it is to rank sources by evidence: deposits paid, release clauses, agent movements, and wage-bill structure. Release-clause structure and the wage bill are the real story — the rest is usually just a misapplied label.

Now return to the file of August 13. The football label sat on an article about a fire. What happens if I just keep running it?

The tactical dimension would be forced to generate a verdict on shape. The transfer-market dimension would be forced to invent a deal. The industry-transmission dimension would be forced to draw a line from academy to national team. And at the end of that line, a reader would be reading a football analysis written out of a hospital fire.

That is the worst-case scenario in my trade. It is bad not because it is hard. But because it is easy. It costs one click.

The paradox lies elsewhere

The intuitive reaction of most people on seeing nine analytical dimensions return N/A is: the system is broken. They want an answer. They want the model to say something. And in that fever for answers, they look for ways to force the system to speak.

I hold that the correct reaction is the opposite. Nine empty results are not a system failure. That is the system doing the hardest thing well: refusing to answer when there is nothing to answer with.

In data analysis we are taught to distinguish correlation from causation. But there is an earlier error, and a less-taught one: confusing the label with the content. A classification tag is not the truth. It is an assumption placed there by people, and people place it wrongly.

The frightening part is that when the label is wrong, everything behind it still proceeds smoothly. My model does not cry, does not celebrate, but after every match it owes me a lesson. This time the lesson is: a good process is not designed to always give an answer. It must be designed to detect when an answer is impossible.

There is another temptation worth naming. On seeing something grave involving human lives, the reflex of a content-maker is to find a way to exploit it. But a medical tragedy does not belong to the raw material of football analysis. The only right action is to record it in its proper place, under its proper label, and not drag it into a framework it does not belong to.

I don't trust managers, I trust models. But I listen to managers in order to fix models. This time, the voice I had to listen to was not a manager. It was the upstream valve itself.

The signal for the next round

The file of August 13, 2026 now carries the correct label. It has been pulled from the football analysis pipeline, and a domain-validation valve has been placed before stage two.

What I carry away from this case is not a lesson about football. It is a question for anyone operating a data system: are you measuring the quality of the answer, or the quality of the question?

Because if the upstream valve opens the wrong way, then every beautiful xG table behind it is just a map drawn accurately of a land that does not exist.

My signal for the next round is this: whenever a file labelled football comes through the gate, the first thing I do is not run the model. The first thing is to look for a ball inside it. If I don't find one, I stop. And stopping, sometimes, is the most precise calculation I can perform.