Labeling Error in the Football Data Pipeline: 18 Information Points With No Player

This month, inside a batch of documents tagged "football" for deep-dive analysis, I hit a file that made my hands stop mid-keystroke. Eighteen information points. No club. No player. No coach, no match, no league table, not a single line about transfers or financial fair play. What appeared before me was a 20% window-tint limit, exhaust regulations, taillight film, and fines of "several thousand pesos" for drivers in Mexico City. The whole file was about the Traffic Regulations of Mexico's capital, and nothing more.

The first xG table I ever drew by hand, on a bus, back when nobody called it data. That day I learned something still true nearly thirty years later: dirty data does not live in the number, it lives in the label stuck onto the number. A wrong cell can be fixed in minutes. A wrong label quietly drags an entire chain off course, and nobody notices until someone sits down and reads every line again.

I tell this story not to catch a system out. I tell it because it lands exactly on my job — the job of a data gatekeeper.

When the label runs ahead of the content

For readers to grasp what I am talking about, I have to describe the pipeline that every modern football analysis passes through. An article, a wire report, a raw data file — they all pour into one first stage. This stage does not analyze. It does one thing: it labels. This is football. This is economy. This is life. That label decides who receives the file, which analytical frame it uses, and where it eventually flows.

To me, the labeling stage is like the assistant referee judging offside. It scores nothing, assists nothing, but let the flag go up one beat too early and the finest move is erased. And like VAR, its mistakes are rarely fixed on the spot — they just move the argument off the pitch and into another room, where the crowd sees nothing but a redrawn line.

During a transfer window, this stage comes under brutal pressure. Every day, thousands of rumor fragments, hundreds of contracts, countless injury updates and agent moves pour in. People want speed. And speed is the first enemy of accuracy. A machine that runs fast tends to label by probability, by keyword, by habit — not by what the content actually is.

The file I met was born of exactly that pressure. It carried the "football" label, but inside was an explainer about traffic law — the kind of utility content written to answer "what percentage of window tint gets you fined." That is service journalism, not sports journalism. Service journalism has its own worth; a piece that helps a driver avoid a ticket is useful content, except it belongs in a different library. The problem is that the label dragged it into a library that is not its own, and from there every later stage treated it as a football document. That label was wrong. And that wrongness does not fix itself.

The evidence chain: eighteen points, not one player

I opened the file and read it the way I read a match report. The only difference was that no one was playing.

Information points one and two discussed sanctions that could apply during 2026 — an administrative date, not a season. Points three and five named the "Reglamento de Tránsito de la Ciudad de México" outright — the Mexico City Traffic Regulations — alongside "the capital's authorities": municipal regulators with no role in football. Point four mentioned "economic sanctions reaching several thousand pesos." I read that line three times, because in my trade every monetary figure must sit beside a financial context. Here there was no contract, no wage, no transfer fee, no amortization. It was a traffic fine.

Point six said that "ignorance of the rules does not spare a driver a ticket" — a principle of legal liability, not a claim about form. Points seven, eight, nine, ten and eleven described lights, films, exhausts and vehicle accessories. Point twelve was the author's advice: check the technical specs before modifying a car. Points thirteen and fourteen covered the 20% limit — the only quantitative figure in the whole file, yet it measures how dark a pane of glass is, not anything on a pitch. Points fifteen and sixteen described the medical-exemption process and restated the role of traffic authorities. Points seventeen and eighteen noted that multiple infractions can stack into a larger fine.

Having read all eighteen points, I took out pen and paper and did what I always do with any match: I counted. Clubs named: zero. Players: zero. Coaches: zero. Leagues: zero. Matches: zero. xG, PPDA, possession figures: zero. Moves described: zero.

A string of zeros like that, inside a file tagged as football, is not a bad match. It is a match that does not exist.

Drawing on my experience watching matches, I hold one rule: no sample, no conclusion. I once spent an entire season counting every move by fourteen V.League clubs, only to find Phan Văn Đức — a twenty-year-old winger at SLNA — posting an xG per match of 0.48, above the average for foreign strikers in the league. He scored just five goals, and I still wrote that he would become a national-team pillar within three years. Many laughed. But I dared to write it because I had a sample, a denominator, a control. Here, I have an empty sample. And an empty sample feeds no conclusion.

I wondered what would happen if I fed that 20% figure into my model. It would sit beside other indices, get normalized, get compared — and manufacture a false correlation. The model cannot tell a car parameter from a tactical metric. It only knows there is a quantity, and a quantity always wants to be used. That is where dirty data begins to breed: not because someone typed wrong, but because someone trusted a label.

The difference between an xG table I verified by hand and a machine-assigned label is this: the xG table admits its limits. It states the sample size, the pitch conditions, the variables it cannot measure. A label never says "I might be wrong." It just sits there being right until someone catches it being wrong.

The counterintuitive angle: a labeling error is not small

People often treat a labeling error as a trifle — a stray file, delete it and move on. I disagree. Correlation is not causation, and one stray file does not by itself wreck a model. But it is a symptom, and symptoms always tell the truth.

Imagine I forced this file into the football framework just to fill the format. I would have to invent a club, a tactic, a player. I would have to write about the "loose defense" of a team that does not exist, and turn a 20% glass limit into a tactical parameter. The tables would be full. And all of it would be zeros inflated into something real.

That is the trap anyone in data has nearly fallen into: forcing a model to answer when the data is not big enough. Turning three matches into long-term form. Turning a wrong label into a right conclusion. My model does not cry, does not celebrate, but after every match it owes me a lesson. This time the lesson came from a file with no match in it.

The crowd watches the move; I watch 22 numbers in motion — and wait patiently for them to tell a different story. But when no one is on the pitch, those 22 numbers amount to nothing. A gatekeeper's job is to say so, instead of staging a match to look good.

The worry is not one file. It is this: if the labeling stage erred once on a Spanish-language text about traffic law, how many more times might it err on less obvious texts? An injury report on a player could be tagged "health" and drift out of the football stream. A club-finance analysis could be tagged "business" and vanish from fans' sight. A labeling error does not kill one article. It kills an entire information stream, quietly and cleanly.

There is a paradox worth pondering. Precisely because this file was mislabeled so glaringly, it became a perfect test case. An obvious error is an easy error to fix. What worries me are the invisible ones: a football file correctly labeled but holding wrong data, an xG table correctly formatted but built on a wrong sample, a conclusion that sounds reasonable with nothing behind it. No one rereads every line to find those. And those are what poison a model.

The transfer market is a game for those who see far, not those who see much — and seeing far, first of all, means knowing what belongs to you and what does not. When noise drowns the signal, readers need a reliability filter, not a confident machine.

Takeaway

I will send this file back where it belongs — an "out of domain" stack, with a note requesting a re-audit of the labeling stage. Not to catch an individual out. But because in a transfer window, when noise drowns the signal, readers need a filter more trustworthy than a machine's confidence.

The question I leave for the next analysis cycle is not which file was mislabeled. It is: across how many correctly labeled files is an empty match still waiting to be inflated into a beautiful conclusion?