When a File Tagged Football Contains No Football: A Classification Failure and What It Reveals About Sports Data Infrastructure

Hook

The number 18 appeared twice in the same file, and neither occurrence had anything to do with a football match. Eighteen was the count of information points in an analysis tagged "football" — and also the age of the victim in a fatal shooting outside the CETIS 33 technical education campus, in the Azcapotzalco borough of Mexico City. I opened the file at 9:47 a.m. Madrid time, after pouring coffee and turning my notebook to a fresh page. Three minutes later, I still had not written a single line.

There was no club name in the file. No player. No scoreline, no PPDA, no xG, no starting XI, not a single line of GPS data. There was only a criminal incident, an open investigation by the Mexico City Attorney General's Office, police sealing off the scene, and two women treated for nervous shock on site.

I have spent nearly thirty years in stadium stands and press rooms doing exactly the opposite: turning dry data files into an understanding of playing rhythm and team state. But this time, there was nothing to convert. No team to analyse. There was only a domain classification error that occurred at the input-processing layer, and it passed straight into a specialised football analysis workflow without being stopped.

Context

Since 2026, when I began my career at a small newsroom in Madrid as a beat reporter following the team, the information intake structure has changed in a way that is almost irreversible. Back then, every story passed through an editor's hands before classification. A crime reporter read crime stories. A football reporter read match results. There was a professional boundary, and that boundary was protected by humans — slow, expensive, but effective.

By 2026, when Real Madrid granted me special access to the Valdebebas training complex throughout pre-season preparation, I realised something else was changing at a similar speed: the way data was collected, tagged, and distributed. Zinedine Zidane was then trialling a new-generation GPS positioning system on 18 players, including Luka Modrić at 32. I spent nine days cross-referencing load metrics against the results of 11 friendly matches. The team's average pressing intensity fell 14%, but finishing efficiency rose 28%. I wrote a conservative analysis, warning about the risk of over-reliance on midfield counter-attacking speed.

That piece had two career consequences for me. First, I began requiring any data provider to specify the source and software version — something I had previously treated as an administrative detail rather than an analytical condition. Second, I understood that the more a system automates the classification stage, the more invisible that stage becomes — and invisible errors are the hardest to detect. When Valdebebas stopped trusting intuition, I began trusting data. But I also learned the reverse: data that has not been cross-checked is just intuition packaged more carefully.

From Valdebebas to Kazan, I learned that football's rhythm is not in the goals — it is in what gets recorded before the goals arrive. But there was one dimension the 2026 training camp did not teach me, and Kazan 2026 taught it to me in its most expensive form: bad input data can destroy the entire downstream output, no matter how carefully that downstream is processed.

Core

It took me about forty minutes to read this file according to the exact procedure I apply to every document entering my notebook. I split it into three columns. Column one: the core incident — a shooting outside the CETIS 33 campus, Azcapotzalco borough. Column two: the involved units — district police, the Mexico City Attorney General's Office, witnesses, two women treated for nervous shock. Column three: investigation status — motive and perpetrator identity still unresolved, scene still sealed.

Those three columns, laid side by side, connect to the "football" tag at no point. I checked three times. No football club has its headquarters in Azcapotzalco in that file. No league is mentioned. No player, coach, referee, or football official appears across the 18 information points.

The nature of this incident, seen from a data-infrastructure angle, is not a single editorial mistake. It is a systematic classification error, and to understand why it matters, one must look at the three layers of any modern sports data analysis pipeline.

Layer one — entity extraction. An automated system scans the document and attempts to identify named entities: people, organisations, locations, events. For a crime report, entity extraction can easily produce noisy results. Abbreviations such as CETIS, SSC (Secretaría de Seguridad Ciudadana), or the place name Azcapotzalco may be misread by a pattern-matching rule set as sporting entities — especially if that rule set was trained on a large but contextually insufficiently diverse dataset.

Layer two — domain tagging. This is the deciding layer. After entity extraction, the system matches those entities against a predefined domain taxonomy: football, basketball, economics, crime, politics, and many others. If a geographic entity such as "Mexico City" is flagged by an auxiliary keyword like "World Cup 2026" appearing somewhere in the same data space — even if not in this file — the system can push the document into the football domain. This is a typical error of a classifier running on probabilistic mechanisms.

Layer three — task routing. This is the most dangerous layer in this specific case. Once the document is tagged "football", it automatically enters the deep-analysis queue reserved for football. No cross-check step blocks it. The nine-dimension football analysis framework — from tactics, club finance, results and public-opinion cycles, to governance and industry transmission — is applied to a document containing not a single football entity. The result is that all nine dimensions return non-applicable values, and the only genuine value this process produces is the discovery of the classification error itself — a data-quality value, not a sporting one.

Cross-referenced against my own working process in Kazan in 2026, the similarity is structural. In June that year, I was assigned to cover the German national team at their training camp in Kazan. During the match against Mexico at Luzhniki, an editor asked me to commentate live on radio to replace a colleague who had fallen ill. I mispronounced the name of forward Timo Werner three times in the first half, calling him "Wermer". The content director reprimanded me publicly. I did not offer excuses.

Instead, I hired a local assistant to record the correct pronunciation of nine German players, then filmed myself practising thirty minutes every evening for two weeks in the hotel. I also compiled a list of easily confused proper names across the Spanish — English — Russian language systems. From that incident, I built a "personal transliteration table" before every tournament, with at least 50 core player names from major teams. At Kazan, a wrong name can change the flow of an entire match. And I understood that an error at the identification layer is not an administrative error — it is a delayed tactical error.

Today's incident in Azcapotzalco is the industrialised version of the same lesson. If my mispronouncing a name three times is one reporter's problem, then a wrong-domain file entering a nine-dimension analysis pipeline is the problem of an entire infrastructure chain. And the most worrying thing here is that nobody misread a name. No human made a mistake at the human layer. The error lies precisely where no human was close enough to be able to make a mistake.

Contrarian

There is a counter-intuitive way to see this incident: the problem is not the classification algorithm. The algorithm does exactly the job it was given — it looks for patterns, and patterns exist: a major Mexican toponym, a public-security context, a sequence of abbreviations that could be mismatched. If anyone is responsible, it is the people who designed the process without a domain gate.

The real blind spot is not at the tagging stage, but at the trust stage. A system can only be automated to the extent that its operators are willing to take responsibility for the output. When an organisation places speed above traceability, it does not increase productivity — it merely shifts risk from an easily detectable layer to a hard-to-detect one.

I have spent my whole career saying this in another form: data is the visible part. I have spent my career looking for the submerged part. But in a fully automated environment, the submerged part is no longer what is hidden — it is what nobody thinks to check, because it appears already checked.

There is a reasonable counter-argument: without large-scale automation, we cannot process the volume of modern-season sports data. That is true. But automating the processing stage does not mean automating the judgement stage. And a process with no domain gate — where the minimum requirement is at least one entity from the target domain before routing a document — is not an automated process. It is a process that skips the most important step.

Takeaway

In my Valdebebas notebook, I keep a rarely used column that must always exist: "source mismatch". It is not for mistakes I have found. It is for mistakes I have not yet found. Today's file incident deserves an entry there — not because a crime report was mislabelled, but because that error passed through four processing layers without being stopped. The question I carry into my next working session: if a file with eighteen lines and not one line of football can enter the football analysis queue, how many other files have entered it that nobody has noticed?