Mislabeled Data and the Blind Spot of Football Analytics

One April morning in Chengdu, I opened the input record of a football analytics pipeline and saw the familiar line: Domain label — football. Below it were 21 information points. I read the whole thing. No clubs. No players. No coaches, no stadiums, no competitions, not a single shot.

What I read about was Klebsiella pneumoniae, an antibiotic-resistant bacterium found in dogs and cats across 25 countries, together with the genetic relationship between strains isolated from companion animals and from humans. The original article carried a Spanish headline: “¿Tu mascota puede portar bacterias resistentes a antibióticos? Esto se encontró”. The study was published in Transboundary and Emerging Diseases, led by Professor Stephen Fordham of Bournemouth University.

A veterinary health file sat neatly inside a football drawer, and nobody noticed through the early stage of the pipeline. That moment made it clear the problem was not the bacterium.

Context: a label applied by habit

Digital sports content runs on speed. A typical data pipeline moves in three steps: a crawler gathers text, a classifier assigns a domain label, and an entity extractor pulls out names of people, organisations and numeric values. That second step decides which drawer the document lands in. In this case, it landed in the wrong one.

The science in the original article was, in fact, careful. The team analysed 712 animal samples and compared them with more than 38,000 human-origin samples. They recorded that 87% of the bacterial strains in companion animals were closely related genetically to human strains, that 43% of samples were antibiotic-resistant, and that multidrug resistance reached 80% in cats and 56.3% in dogs. The ST147 strain was flagged as a lineage worth watching.

Then the authors cooled their own story down. They stated plainly that the study does not prove pets transmit the bacterium to owners, and that there is “no reason for owners to be alarmed”. A piece of research that sets its own limits — something our football analysis trade rarely manages.

So where did the football label come from? The uncomfortable answer is a default field in a form. Someone built the pipeline with a fixed list of domains, and any file that fits nothing falls into the first empty slot. I have seen the same thing in V.League datasets: a pre-season friendly labelled as a competitive fixture, and for three weeks every internal table was skewed.

The original article also lends a concept worth borrowing: One Health, the view that human, animal and environmental health form one linked chain, so antimicrobial surveillance must include companion animals. Football works on the same logic. Academies, clubs, the national team and the scouting network form one chain, and a broken link at academy level surfaces at national-team level five to seven years later. We call that youth development. People only call it a chain when it snaps.

The core: three kinds of label error Vietnamese football keeps making

The most visible label error is string collision. In the original article, “Bournemouth” appears as a university, entirely separate from AFC Bournemouth in the English Premier League. The word “transmission” is used in its epidemiological sense, far from the footballing sense of transition. The phrase “25 countries” is public-health geography, not transfer geography. Three strings, three meanings, and a sloppy entity extractor merges all three into one drawer.

In Vietnam this collision happens daily. A young player shares a name with a company. A province has a professional club and a grassroots tournament of the same name. A foreign coach shares a surname with a former European star. Without an entity-type gate, any statistical table can be contaminated and nobody notices.

Sample asymmetry is easier to miss. The research team set 712 animal samples against more than 38,000 human samples, a ratio of roughly one to fifty, and they stated that ratio to remind readers the animal-side evidence is thinner. Ignore that note and a sensational headline can push readers toward a conclusion stronger than the data supports.

Vietnamese football runs on samples that small. A striker scores four goals in five games and is immediately called irreplaceable. A midfielder starts a few matches abroad then returns home and is declared a failure, as in Nguyễn Quang Hải’s spell at Pau FC. I once analysed Oscar’s GPS data in the 2026 Shanghai derby and found 14 movements into the right half-space creating space for Vương Thâm Siêu to run into. One match, one sample, and I stated in the piece that the conclusion was only a hypothesis.

Data does not replace instinct, but it maps out where instinct is fooling itself.

Scouting networks in developing football nations are the clearest illustration of the sample problem. The same network finds one genuine talent and simultaneously produces trial invitations that lead nowhere and families that fall apart. The data on successes looks beautiful. The data on failures barely exists, because nobody collects it. A dataset that records only the bright side is not a dataset; it is a prospectus.

There is one more device: the question-form headline. “Can your pet carry antibiotic-resistant bacteria?” creates just enough alarm and just enough cover if challenged. The transfer market uses the same construction: “Will this star leave V.League?” The question commits to nothing but harvests the engagement of an assertion.

I learned the value of entity verification through a stumble. At the 2026 World Cup, commentating live on Spain against Portugal, I misread the name Diego Costa three times as “Diego Castro”. Viewers caught it instantly. I did not make excuses; I spent four weeks rewatching all 12 group-stage matches and taking notes in the present tense: the moment possession was lost, the positions of the midfielders. Since then, every player’s name is checked in three languages before I go on air.

That is precisely the entity-type gate a football data pipeline needs, except mine sits in my head rather than in the code.

Pitch sound does not lie; pictures always know how to colour things in.

The point the authors actively defused is the one most likely to shock: correlation is not causation. They found genetic relatedness and stated that relatedness does not mean pets make owners ill. In football we are rarely that honest. A team wins seven matches with its first-choice centre-back starting, and the conclusion is written immediately: he is a lucky charm. Seven matches, no control group, nobody checks who the opponents were.

Space is the culprit, time is the witness.

The contrarian angle: the mislabel is not the most dangerous part

The visible problem is the wrong label. The bigger concern sits elsewhere: football analytics almost never audits its own inputs.

We check expected goals, argue over minutes played, pass completion and transfer values. Very few people go back to ask where the source file came from, who labelled it, and whether that label was right. A veterinary health record landing in a football drawer is the pipeline’s fault. That the error survived an entire stage unnoticed is a culture problem: we reward confident conclusions and do not reward provenance.

I do not believe the opposite extreme either. Some will say: drop all external data and use only internal numbers. That is blinding yourself. It is the external data — 712 against 38,000, the 80% and 56.3% figures, the ST147 strain — that gives the story weight. Its value lies not in whether it belongs to football, but in showing a way of setting research limits that football should learn.

What should be dropped is not outside data, but the act of unconscious labelling.

The moment possession changes is when the match truly begins. And in analysis, the moment data changes drawers is when the error truly begins.

What to verify

Three questions belong at the door of every dataset. What entity types does this file contain. From how many observations was its largest value drawn. Does the strongest conclusion in the source come with any stated limit. If all three go unanswered, better to leave that file outside than to put it on the analysis table.

A pitch is not a map, it is the coordinate of split-second decisions. And football data, before it becomes a tactical map, has to be a photograph taken in the right place.