Mislabeled in the Analysis Room: When Football Data Gets Contaminated

In June 2026, I discovered a file in my personal archive that had been mislabeled. It sat inside the folder “Serie A 2026-2026”, between two reports on Gasperini’s Atalanta. I opened it and found an article about the battery of a Vietnamese electric motorbike. It had been sitting there for nearly a year. I had even cited it as a data source on wide attacking phases.

That was the moment I understood that even the most careful analyst can plant an error in their own system without realizing it. And when a mislabel slips through the door, it does not stay still. It spreads.

I am telling this story not to talk about motorbikes. I am telling it because during a major tournament season, when every match is recorded and every phase is tagged with numbers, our databases are growing faster than any human can verify. Mislabels like these are becoming a real problem.

When a file wanders onto the pitch

This was not the first time. Over 29 years of observing the industry, I have seen analysis rooms fill up with similar cases. An advertisement tagged as “tactical analysis”. A market report stored as “transfer market”. A press release about a sports drink landing inside an injury database. These are not acts of intent. They are the natural consequence of an auto-tagging pipeline running too fast.

Technically, these systems work like this. An article is ingested from a source. A classification model reads the title, the keywords and sometimes an excerpt, then assigns it a topic label. The process takes milliseconds. When it is right, it lets an analyst like me filter through thousands of documents in an afternoon. When it is wrong, it drops a pebble into the gears.

In this case, the article had a headline about a “3 million VND promotion” for two electric motorbike models. Its content covered a 2.4 kWh battery, a 2,250 W motor, a 70 km/h top speed, a 262 km range, 35 litres of storage, and an IP67 rating. All of it was vehicle specification. There was not a single full-back in it. But because the headline contained certain keywords that overlapped with sports context, the article slipped into the football document set.

A heat map shows position; an intent map shows thought. But to draw an intent map, you need a correct label. A wrong label turns a heat map into a meaningless image.

How a small error spreads into a wrong conclusion

Suppose the article had never been caught. Suppose it stayed in the database, and three months later a young analyst needed a number on the conversion rate of right-wing counterattacks. She opens the file, skims it, sees a figure that looks credible — a number, a unit, a source. She cites it. A colleague reads her piece, believes it, and puts it into a report sent to a coach.

It took me three months to realise I had been reading this position wrong. In my case, it was Gosens’ position inside Gasperini’s system. In the case of the database, it was the position of a file inside a folder. The same kind of mistake, the same mechanism of transmission.

This is why I never say a number is the final truth. Numbers do not lie, but they do not tell the whole story either. A number is only as credible as the process that produced it. And in most modern pipelines, there is a tagging step almost nobody re-checks. That is the structural weakness.

If you read a tactical report based on 4,500 situations, ask one question first: who labelled those 4,500 situations, and how. That is not a question of suspicion. That is the question of someone doing serious work. Ask what the system has hidden before you judge a defender.

The problem is not the model

There is a natural reflex when an error is found: blame the classification model. But in most of the cases I have observed, the model only does what it was taught. The problem lies in the assumption that tag speed can outrun tag quality without a cost.

I once witnessed the opposite. In the summer of 2026, a colleague in Turin spent two weeks manually checking 800 documents before feeding them into a prediction model. He was told he was slow. When the model ran, it produced a prediction against the consensus — and it was right. He was not faster, but he was more trustworthy. The difference was that he treated tagging as part of the analysis, not a preparatory step.

This is the counterintuitive point. In an industry that praises speed, slowing down at the first step often produces more value at the last. But slowing down does not mean doubting everything. It means building check-gates at the high-risk points — where a commercial article can slip into a technical category, where advertising data can become a performance metric.

Emotion is not data noise; it is data that has not yet been decoded. Ironically, it is emotion itself — a young analyst’s suspicion when reading a number that looks too good — that is sometimes the only line of defence before an error spreads through the system. Tools cannot replace that unease. They only amplify it.

What I changed

After finding the electric motorbike file inside the Serie A folder, I changed my routine. I added a step: before citing any document, I must read all of it, not just skim the headline. I added a column to my tracking sheet recording the date I checked a label by hand. And I set a rule for myself: if a document cannot be verified against a second independent source, I do not let it into a final conclusion.

This sounds basic. But most modern analysis rooms do not do it at scale. They rely on an auto-tagging model and trust that it is right. When I write reports for Italian clubs, I always add a short appendix on data sources and tagging. Most readers skip that appendix. The ones who do not are the ones I want to work with.

A question for this season

When a major tournament season arrives, every number gets accelerated. Analyses spring up within hours of the final whistle. Models update continuously. That agility has value — it keeps the conversation close to what actually happens on the pitch.

But I want to ask a question of myself and my readers. In your database, how many labels have you never checked by hand? How many articles have you cited only because the headline looked like the right topic? If you cannot answer, that may be a weakness larger than you think.

There is no conclusion here. Only a habit I suggest you try in the next match: before trusting a number, ask where it came from and who tagged it. The answer will sometimes surprise you, and sometimes make you drop a citation. That is the value of asking.