Wrong Labels and Leaky Pipes: The Cost of a Single Misclassification in Football Data

Wrong Labels and Leaky Pipes: The Cost of a Single Misclassification in Football Data

At three in the morning in Chengdu, my dashboard lit up with a strange data record. Sitting among hundreds of entries about the weekend's matches — the Manchester derby, the relegation race in Serie A, the transfer deals heating up in the Premier League — was an article about Pakistan's fiscal budget, neatly tagged by the system: "football". I read it three times. No players. No coaches. No stadiums. Only the International Monetary Fund, a fuel-subsidy scheme, and figures about poultry farming. At the end of the record, a single field read two characters: N/A.

That was the moment I understood something thirteen years in the trade had not fully taught me: our analytical pipelines do not collapse from a lack of data. They collapse because too much slips inside without anyone checking its passport.

In modern football, we are used to stories about data abundance. Every Premier League match generates more than 3,000 recorded events — passes, duels, shots, the position of every player down to the hundredth of a second. A mid-table European club can track thousands of players worldwide through platforms such as Wyscout or StatsBomb. Scouting departments no longer send people to watch ten matches a week; they filter hundreds of thousands of profiles by algorithm before picking a few names to watch with their own eyes.

But behind every xG model, every PPDA metric, every transfer-value ranking, lies something few ever see: the data pipeline. And every pipe has joints. Every joint can leak.

A fiscal article inside a football system

Let me recount exactly what happened to that record at three in the morning. The cause was not a typo. It was a chain of decisions, each reasonable in isolation, that together produced something meaningless.

Step one: a news-collection system scans tens of thousands of articles a day from around the world. Step two: an automated classifier tags each article by topic — sport, economics, politics, entertainment. Step three: articles tagged "football" are pushed into a dedicated pipeline, where entities are extracted: which player, which club, which league, which event. Step four: those entities feed deeper analytical models.

At step two, the Pakistan article was tagged "football". Perhaps because the headline contained an overlapping keyword. Perhaps because the classifier drifted after an update. Perhaps because of a source-mapping error — the article was really about finance, but it sat in a feed the system had defaulted to sport. Whatever the cause, the result was identical: a document with nothing to do with football had walked through the door and sat down at the analysis table.

The frightening part is not that it got in. The frightening part is that, if the analytical layer is not alert enough, it will try to find football inside it. And when a model is forced to find something that does not exist, it will invent it.

I have seen that happen. In 2026, when the Bundesliga restarted in empty stadiums, I analysed 88 matches and found the home-win rate fell from 42% to 30%. I built a dedicated xG model for teams defending deep, and from that data I predicted Leipzig could not overturn PSG. The prediction was right. But the lesson was not "data always wins". The lesson was that data can collapse, and when it collapses, so do you. That year's numbers collapsed, and so did I — then I learned to rebuild from the fragments of doubt.

The record at three in the morning was a smaller version of the same lesson.

N/A is not a failure — it is an immune system

When the deep-analysis layer read the Pakistan article, it returned a remarkable result. Nine analytical dimensions — tactics, club finance, match results, league context, rules and governance, dressing room, risk, media, industry transmission — all returned the same sentence: N/A, insufficient information.

That sounds like a failure. But look closer. The system did exactly what very few football analytics systems dare to do: it refused to invent.

In the world of scouting and analysis, the pressure to always have an answer is brutal. When a coach asks about a player, no one wants to say "I don't know". When the board needs a report before the transfer window, no one wants to hand in a blank page. So analysts learn to fill the gaps. They infer. They extrapolate. They turn a small sample into a trend, and a trend into a conclusion.

And that is precisely where the pipeline gets poisoned. A mislabelled fiscal article may do no harm if it is removed at once. But if it enters a machine-learning model alongside thousands of other records, it becomes a data point. And a wrong data point, repeated often enough, becomes a wrong rule.

The three leak layers of a football data pipeline

Over years of working with sports data, I have found that incidents like this always leak at three layers. The first is source mapping. A feed is configured to tag everything arriving from it as sport, because 99% of the time it truly is a sport source. But the remaining 1% is an article about a national budget. No one checks that percentage until it appears on my dashboard.

The second is the classifier. Labelling models learn from past data, and they learn the biases of the past too. If, in the training set, the word "release" appears often in articles about players' release clauses, then an article about the "release" of a fuel-subsidy scheme will easily be pulled toward football. Language is a space, and within that space words collide in ways no analyst anticipates.

The third — and most dangerous — is the interpretation layer. When a young analyst receives a strange record, the first reflex is not "this is wrong" but "what do I do with it". They will try to find a football angle. Maybe a metaphor. Maybe a link about financial governance. Maybe a lesson about how an organisation comes under external pressure. It all sounds very clever. And it is all fabrication dressed up in professional language.

I have been at that third layer. In 2026, as a third-year student in Chengdu, I wrote a 3,000-word analysis of France's 4-3 win over Argentina at the World Cup. I did not focus on the goals. I decoded how Didier Deschamps set up a shifted midfield diamond to exploit the space behind Argentina's midfield line. I counted exactly 11 line-breaking passes by Kylian Mbappé in the second half. The piece drew 15,000 reads on a forum. But what I am proudest of is not that number. It is that I began everything with a pitch diagram, not with a narrative line.

A pass is just a pass, until you read the intention of the whole block of space. And a record is just a record, until you read what it is truly telling you.

A cost that can be measured

If this were only the story of one lost article, it would not be worth writing. But its cost can be measured, and that is the part that keeps me up at night.

Picture a club preparing for the transfer window. Its data department runs a player-valuation model built on hundreds of thousands of text and numeric records. If 0.5% of them are contaminated, then on a dataset of one million records we are talking about 5,000 wrong data points. It does not take much to skew a model. In machine learning this is called label noise, and just 5-10% of label noise is enough to destroy a classifier's performance.

But the real loss is not in the number. It is in time. Every hour an analyst spends checking a junk record is an hour not spent watching footage. Every week a scouting department spends tracing a pipeline error is a week not spent evaluating a rising 20-year-old centre-back like Joško Gvardiol, or a young midfielder like Jude Bellingham reshaping the middle of the pitch. And in football, time cannot be bought back.

I know this because I paid with my own time. In 2026, at the World Cup in Qatar, I noticed a weakness in Croatia: their defensive transition when the ball was lost in midfield. I wanted to build a perfect model, with a pressure index on Gvardiol — then just 20 years old. Chasing perfection, I delayed publication by three days. Another analyst published a similar piece the next day and drew big attention. I arrived late because I wanted a perfect map; it turned out the match had already redrawn itself.

That lesson applies straight to the pipeline story. A wrong record ignored for three days multiplies into ten, then a hundred, because data is copied, shared and reused. The pipeline does not just leak. It spreads.

Quantifying a systemic crisis

This is where I must be most careful with myself. Space is my preferred measure, and scepticism toward numbers is my instinct. But if I dismiss all data just because it can lie, I fall into the very trap I always warn others about. Space does not lie — only people fool themselves with numbers. The problem was never the number. The problem is how people use it.

So let us quantify this crisis properly. A football data pipeline at a professional club can process anywhere from 50,000 to 500,000 records a month, depending on scale. The classification error rate of a good automated labeller typically runs between 1% and 5%. That means each month a club may face thousands of contaminated records. Most are filtered out by checking layers. But not all. And what is not filtered out is the dangerous part, because it carries the appearance of plausibility.

I once wrote that Team A no longer "lost focus" but "lost its ability to organise when its pressing pressure dropped 12%". That phrasing is not to show off numbers. It is to force a judgement to be accountable. If I say a system leaks, I must say how much, where, and at what speed. Otherwise I am just adding another layer of noise to the very pipeline I am criticising.

The blind spot of automation

There is a widespread belief in the industry that automation will solve the data problem. More, faster, fewer humans. But looking at that record at three in the morning, I see the opposite.

It was a human — an alert analyst at three in the morning — who was the final checking layer that caught the error. Had I slept well that night and let the system run, that record would have gone straight into a model, and from there into a report, and from there into a transfer decision. The biggest blind spot of automation is not that it is wrong. Its blind spot is that it does not know it is wrong, and has no one to ask.

In football we are far too used to worshipping models. We speak of xG as a truth, of transfer value as an objective measure, of a pressure metric as a verdict. But behind every number is a pipeline, and behind every pipeline are people deciding what gets in and what gets blocked. Transfer value is the story, but I prefer reading the footnote.

The transfer window: a stress test for every pipeline

There is no moment when the dirty-data problem becomes more serious than the transfer window. In a few short weeks, every club must make decisions that can shape years. A good signing can lift a team to the top; a bad one can sink a budget.

Amid that flood, the data pipeline comes under several times its normal pressure. Thousands of rumours, hundreds of scouting reports, dozens of valuation models — all running in parallel. And it is precisely in that flood that contaminated records have their best chance of drifting through unnoticed.

I once saw a club nearly put a player on its watchlist simply because a model misread an injury article. The article was about a different player with the same name. Same name, different person, different position, different league. The entity-extraction tool merged them into one. The result: a profile with wrong injury data, a wrong valuation, and a decision that almost went wrong.

This is why I always tell young analysts: never trust a profile just because it is complete. A complete but wrong profile is more dangerous than an empty one. An empty profile forces you to go and look. A wrong profile gives you the feeling that you already know.

And in the transfer window, the feeling of "already knowing" is the most expensive thing in the world.

When the media helps the pipeline leak

There is one more layer I have not mentioned: the media. When a wrong record slips through the pipeline and becomes a number, it does not sit quietly in a database. It gets cited. It gets shared. It becomes part of the story fans read every morning.

I have seen player rankings built on contaminated data. I have seen commentary citing a metric whose origin no one verified. And I have seen players misjudged simply because of a labelling error from months earlier.

This is where the role of the sports journalist changes. We do not just report. We are the final checking layer between a model and the public. If we cite without verifying, we are not journalists. We are loudspeakers for a leaking pipe.

A contrarian view: do not just fix the label

At this point the natural reflex of anyone is to go and fix the label. Change "football" to "economics". Done. Problem solved. But I think that is exactly the biggest mistake.

If we only fix each wrong label as it appears, we are treating the symptom, not the disease. The disease is that our systems are designed to always answer, rather than to know when to stay silent. A fixed wrong label is a small victory. But a system with no capacity to refuse is a large disaster waiting to erupt.

The paradox sits here: we fear dirty data, yet we build systems that cannot tolerate dirty data. We fear wrong models, yet we reward models that are always confident. What we truly need are models that can say "I am not sure" — and analysts who value that answer.

Moreover, there is an uncomfortable truth few want to admit: dirty data is not the exception. It is the rule. Every pipeline in every industry leaks. The question is not how to stop the leak, but how to detect it before it spreads. And detecting a leak is a human skill, not an algorithm's. An algorithm can scan a million records in a second, but it cannot tell a football article from a fuel-subsidy article — unless someone teaches it, or someone checks it.

And here is the most contrarian part: sometimes the wrong record itself is the most valuable asset. It is the stress test of the system. A pipeline that has never met a wrong record is a pipeline that has never been tested. The record at three in the morning did not break my system. It proved that my system is alive, still able to refuse, still able to say N/A.

Takeaway

So what should we do with a leaking pipeline? The answer is not to shut the valve and stop using data. The answer is to build checking layers that can say "I don't know". A mature analytical system is not one that always has an answer. It is one that knows when the right answer is N/A.

The record at three in the morning taught me that. It is not a disaster. It is an immune system at work. And the question I carry into next season is simple: is your pipeline brave enough to say "I don't know" before it manages to invent an answer it should not have?