Wrong Label First, Wrong Model Later: A Data Classification Lesson from Vietnamese Football

2:47 a.m. in Shanghai. I open a CSV with 214 rows. The third column reads "Position." Row 87 reads "Defensive Midfielder." I drag across to the progressive carries column and find a player receiving the ball nine times a match between the lines, carrying past the halfway line more than anyone in the squad. No defensive midfielder plays like that. That label was typed in 2026 onto a match registration sheet, and every model fed by that column since has been confident and wrong in exactly the same place.

I have met that same failure in a completely different field: a file labelled as one thing but filled with another, run through an entire analysis pipeline where nobody bothered to interrogate the header row. Beautiful regression output. Wrong subject. A label does not describe a thing. A label is a governance decision, made by one person, on one day, for a reason that has nothing to do with your analysis.

Vietnamese football data runs on three layers of labels, and each lower layer inherits from the one above without anyone re-checking. Layer one is the administrative label: position on the registration sheet, foreign or naturalised category, year of birth, shirt number. Layer two is the market label: transfer fee, salary, the phrase "marquee signing" or "cast-off." Layer three is the media label: fifteen words in a headline, "crisis," "wonderkid," "dressing-room wrecker."

These three layers never cross-check each other. Layer one carries legal force: it determines foreign-player quotas, registration eligibility, and the squad structure a coach is permitted to build. Layer two carries financial force. Layer three carries emotional force. Your model sits on layer four, where nobody is accountable for the three layers above.

Based on my experience watching matches in the V.League and the Chinese Super League since 2026, most of the error in the prediction models I have built does not come from the algorithm. It comes from the header row of the data file.

The first chain of evidence sits in playing position. I took a sample of 180 V.League players with at least 900 minutes and cross-checked their registered position against their action map: average receiving point, passes into zone 14, frequency of involvement in second balls. About a quarter of the players had a position label that diverged from actual behaviour by enough to change their tactical role. The most badly mislabelled group was young central midfielders who were placed in the "defensive" box in school football because of their build, then judged for an entire career by an attribute they never possessed.

The cost is not in the number. It is that a mislabelled player gets bought at the price of the label, not the price of the behaviour. And the buyer never learns he just paid for a footnote.

The second chain of evidence is the "naturalised" label. This is a label with teeth: it is tied to quota rules, to registration conditions, to how a coaching staff calculates available slots. But it also drags along a package of assumptions never written down anywhere — that a naturalised player is a short-term fix, that he came for money, that he will never commit to the system.

Nguyễn Xuân Son is the cleanest example I have tracked. Before the 2026 ASEAN Cup, most public argument revolved around whether he deserved that slot. On the pitch, he was Vietnam's leading scorer at that tournament, scored in the first leg of the final at Việt Trì on 2 January 2026, then broke his leg in that very match. Vietnam won the title across the two-legged final against Thailand.

What is worth recording is not that he scored. It is how precisely the "naturalised" label predicted public behaviour, and how poorly it predicted behaviour on the pitch. A label has its greatest predictive power over public opinion, and its weakest over football. xG does not score goals, but it makes people argue more than the actual ball does.

The third chain of evidence is the results label. A team goes three rounds without a win and the press calls it a crisis. I do not object to the word — I object to using it without answering a foundational question: how many three-match winless runs does a title-winning side typically have in a V.League season? If the answer is not zero, then "crisis" is a label, not a finding. And if you have not calculated that answer, you are not analysing — you are reading a caption.

The process I use now has four columns instead of one. The original label column, kept intact, for comparison. The label-provenance column, recording who typed it and when. The behaviour column, computed from event data rather than copied from a list. And the divergence column, measuring the gap between the first two. The fourth column is the only one I actually read when building a model, because it is the only one that tells me how much my data can be trusted.

I have paid for skipping that step. At the 2026 World Cup, my model built on PPDA and defensive-line height correctly predicted South Korea beating Germany 2-0. I went on air and told people to bet it. In the round of 16, the model said Brazil would beat Belgium, I stated it again, and Brazil lost 1-2. Many clients lost money. Three weeks later I rewrote the code and added tournament variables. But the real error was not in the variables. It was that I believed I was measuring what I was measuring.

This is where I have to be straight with myself. The first instinct when a model collapses is to blame noise. All models are wrong, but a few are wrong usefully. Before I allow myself to write the word "random," I have to answer: how many intervening variables have I excluded, and is the label among them? Usually the label is the largest intervening variable and the only one nobody touches, because fixing it wins nobody a prize. There is no medal for correcting one position field in a database.

The counter-intuitive part is not that the model is weak. It is that fixing labels is a political act, not a technical one. Administrative labels are tied to regulation, market labels to money, media labels to page views. If even one of the three layers has an interest in a wrong label, that label will outlive the career of the player it was stuck to. Data disappearing is not a loss of data — it is a kind of data.

And there is a discipline I learned in fields with nothing to do with football, where people are obliged to call someone "the accused" until a final ruling exists. Football has no such discipline. We convict a player in round four and publish the sentence in round five. The label becomes data before the event has had time to become data.

What is worth tracking next round is not any team's form. It is the first headline. Who owns it, when it was typed, and when someone last checked it. If the answer is never, then every conclusion built on it — including the prettiest ones — is just an old error restated in new language.