Mislabeled Input: The Most Expensive Data Error in Football Scouting

A data file was labeled “football”. Inside it: thirty incident reports per day, eleven thousand households served, a list of Mexico City boroughs, and a fire-department director named Juan Manuel Pérez Cova. No team. No player. Not one minute of play. Not one pass, one shot, one press.

I read that file three times. An automated classification system had seen a document about gas leaks, about deflagration risk from electrical sparks, about a fire-service prevention program, and stuck the label football on it. That label travels downstream. It sits in a training table, becomes a weight, then becomes a number inside a scouting report that nobody bothers to trace back to its source.

On the pitch, being wrong means losing. In data, being wrong keeps winning, until the bill arrives as a contract.

In more than twenty years in this trade, I have watched football move from the scout's notebook to the API. A single V.League 1 match now generates thousands of data points: ball position per fraction of a second, touches, running direction, pressing distance. Clubs do not produce that data. They buy it. And what they buy is not football, but a description of football written by a person or an algorithm.

Which means there is a layer sitting between the grass and the spreadsheet. I call it the labeling layer. A passage of play exists in the data only if someone decided it was a cut-back rather than a blocked cross. A player exists as a winger only if someone, on an afternoon four seasons ago, clicked that box.

In 2026 I stayed behind after Sichuan Longfor lost 0-6 to Beijing Renhe in China League One. I rewatched the tape, annotated every phase, then cross-checked it against the data. Sichuan's entire midfield only passed sideways and backwards; not one key pass into the box. I wrote a 3,000-word piece titled “Sichuan does not need a new coach, it needs an algorithm”, using twelve recent matches to show a disjointed pressing system. 0-6 in Sichuan was not a defeat, it was a doorway into the world of data. Before 2026 I watched football with my eyes. After 2026, I watched it with numbers that know how to cry.

In the summer of 2026, while most of the media praised Germany after the win over Sweden, I published a contrarian piece: Germany would go out in the group stage, and Mesut Özil was not the real problem. The number I staked was 41 percent, Germany's duel-success rate in central midfield. The piece was mocked. Germany lost 0-2 to South Korea and went out; the article was shared more than 50,000 times in 24 hours. tôi đã nói rồi.

Positional labels are where the errors are easiest and most expensive. Trent Alexander-Arnold is registered as a right-back, yet for several seasons he has operated as a right-sided playmaker. João Cancelo was deployed at left-back and then inverted inside. Joshua Kimmich moved from right-back to central midfield. These are not outliers; they are the normal state of modern football, where function no longer matches the position printed on the team sheet.

When a valuation model reads the label right-back, it compares Alexander-Arnold with purely orthodox full-backs. His key-pass numbers become an anomaly. The model does not say this player is good. It says this data is broken. And clubs pay for the model, not for the truth.

The event label layer is worse. Definitions of a key pass differ between data providers. PPDA, the passes allowed per defensive action, shifts with each provider's counting method. Two providers watching the same match can produce two different pressing portraits, and both claim to be correct. Put those two numbers side by side on one table and you are comparing two things that do not share a unit.

The Sichuan case runs the other way. I labeled it myself, because I sat through the tape and counted. Twelve matches. Every phase. That is why I dared publish such a provocative headline: I was not arguing from feeling, I was arguing from a count sheet I could open for anyone.

In 2026 I stood in an empty stadium where nobody sang, and for the first time I heard this sport breathe. When European leagues returned behind closed doors, I spent hours on old tape and found a gap: in the 2026-20 Bundesliga, home win rates dropped by roughly 12 percent compared with matches played in front of crowds. That produced the piece “Football without crowds is a different sport”, and the idea of a virtual home advantage built on loudspeaker noise.

The correct label for that season was not an interrupted season. The correct label was a different sport. Scouting models do not collapse at the algorithm layer; they collapse at the labeling layer, where a human or an automated system decides what a thing is before any calculation begins. A wrong label at the input passes through every layer above without detection, because the layers above are not tasked with checking it. They are tasked with computing.

A civil-safety file labeled football does not cost anyone a match. But the same mechanism, applied to a 19-year-old in V.League 1, makes a club pay for someone it has never watched for a full ninety minutes. Today's valuation systems measure what is measurable very well: speed, touches, distance covered at 19. They cannot label the thing that decides a season: a dressing room. Nobody calls that data, so it does not exist in the file.

By the same logic, a goalkeeper labeled good with his feet carries a higher transfer fee than a goalkeeper with better reflexes. The label good with his feet is easy to apply. The label saves low shots is not.

I have to argue against myself, because this line of reasoning slips easily into a moral lecture about data. Football survived 150 years on dirty data and still produced great teams; the human eye handles a bad label faster than any model. Data providers are not standing still either: they revise definitions, standardise events, hire cross-checkers. And that mislabeled file may have caused no harm at all, because nobody acted on it.

I am not someone who is always right. I called Germany correctly in 2026 after staking a number, but I have misread plenty of other matches, and I have never hidden it. In the end, enlightenment through data may just be another way of trusting a spreadsheet you labeled yourself.

One point I still hold: manual labeling does not scale. No team has enough people to rewatch every phase of every match in every league on earth. So the labeling layer will keep being automated, and the errors will keep multiplying. The issue is not whether errors exist, but at which layer they get caught.

A testable prediction: within 24 months, the biggest transfer mistake by a Southeast Asian club will be explained by a labeling error, not by a modelling error. If I am wrong, I will write it up again, with data, as always.

What I want to leave behind is not a model. It is an odd habit: open the file, read the first lines, and ask who labeled it.