Sports Data Failure: When a Football Analytics Pipeline Misreads a Non-Football Story

6:12 a.m., Marseille time. I open the internal dashboard where news records are pushed before an analysis day begins. The top row carries a clear domain label: football. I click in. No team. No player. No coach, no scoreline, no transfer, no league table. The tactical data block is empty; the fields I know by heart, such as pass counts, running distance and touches inside the box, are all marked 'not applicable.' The subject column carries a name that belongs to an entirely different field. A celebrity human-interest item, centred on a family in mourning, has landed in exactly the slot I reserve for football.

I sit still. Eleven years covering this industry have taught me that data lies in subtle ways: small samples inflated, denominators cherry-picked, chart axes trimmed to please the eye. This time the data did not lie. It simply stood in the wrong room. And that ordinary moment worries me more than any tactical error, because it does not sit on the pitch. It sits inside the pipeline that carries data about the pitch.

Football is a game of chess played with pawns that can run. But before we talk about the pawns, I have to talk about the board they are placed on.

Most fans see only the two ends of a long process. At one end is the match, with the ball, the whistle, the stands. At the other end is the article they read the next morning. In between lies a chain of intermediate layers nobody puts on a poster: collection, classification, labelling, routing, and only then analysis. At the first layer, automated machines scan thousands of items a day from every kind of source, from major papers to specialist sites, social media and wire services. At the second layer, a classifier reads headlines, keywords and entities, then assigns each record a domain label: football, basketball, tennis, entertainment, politics, business. That label decides which analytical frame the record is pushed into.

The trouble starts with a very comfortable belief: that labelling is a harmless step. It is just data administration, a stamp applied before shipping. Nobody builds a tactical model on a domain label. But the domain label decides what the tactical model looks at. A record that walks through the wrong door will be reshaped by a framework built for an entirely different subject. And that framework has no shame. It will fill every empty field, pose every tactical question, even for a news item that contains not a single pass.

I once met a milder version of this problem as a second-year economics student in Marseille. On the night France beat Argentina 4-3 in the 2026 World Cup round of sixteen, I sat noting every phase. I recorded that France had only 38 percent possession but produced 14 shots to Argentina's 12; Mbappe alone made six counter-attacking accelerations covering 312 metres in total. I wrote a 4,000-word piece on how Deschamps built a low 4-1-4-1 block to invite the press and then break at speed down the flanks. It drew 12,000 reads in 48 hours, twenty times my previous average. Had a single data record from that match been mislabelled, I could have written a completely different conclusion, highly persuasive and completely wrong.

The lesson about dirty data arrived a year later. In the summer of 2026, with leagues suspended by the pandemic, I was stuck in Marseille and bought the tracking dataset of ten Atalanta matches from the 2026-20 season to decode Gasperini's pressing. Drawing on my experience watching matches, I counted an average of 56 high-intensity pressing actions per game, 23 of them inside the final 40 metres of the opponent's half. I also noticed that when both full-backs pushed high along the vertical axis, the team's total misplaced passes fell 18 percent if one midfielder dropped deep to form a V shape. But to reach those numbers I spent nearly two weeks cleaning the data: removing dead-ball phases logged incorrectly, synchronising timestamps across two camera systems, merging duplicate player identifiers. Skip that step and I could have built a very persuasive analysis of a team that did not exist in the way I described.

By Euro 2026 the habit had become reflex. I spent the week before the final analysing Mancini's Italy, counting 612 passes in the semi-final against Spain, 23 of them line-breaking passes into the final third. I noticed their 4-3-3 was never fixed: in possession, one full-back tucked inside to form a 3-2-4-1; out of possession, it snapped back to a 4-1-4-1. To tell that story of state transitions, I had to trust that every data label I read was correct. That trust is the weak point.

That is why a mislabelled record catches my attention more than a shock defeat. It is not a failure at the analysis layer. It is a failure at the ingestion layer, where people rarely look, because they believe the machines have handled it.

The record that stopped me was structured into 24 information points. I read them one by one. The first concerned a famous artist reported to have spent hours talking to encourage a young person in the family of a close friend who had just died. The second covered handwritten letters and phone calls. The third addressed efforts to lift the spirits of someone struggling with mental health. And so on to the twenty-fourth. Not one point mentioned a club, a competition, a player, a coach, a transfer, a contract or a cash flow.

The core is here: a pipeline built for football received an item with no football element at all, and instead of refusing it, it still tried to analyse it. The full framework was assembled anyway, covering tactics, finance, rules, the dressing room and risk. All of them returned a single value: not applicable, insufficient information. A complete, elegant and utterly empty framework.

I read the source section, and that is where I put down my pen. Most of the most emotionally vivid details, the letters, the calls, the hours of conversation, rest on anonymous sources relayed through a British tabloid and then retold by a secondary outlet in another country. A three-link relay chain, with the first link unnamed. That is the sourcing structure with the lowest verifiability in journalism classification.

The backbone is different. It contains facts recorded in the public record: the friendship between the famous artist and the husband of a well-known supermodel stretching back to the 1990s; the tequila brand co-founded in 2026 by those two and a third investor; the friend's marriage announced in 2026. These markers can be looked up, cross-checked and verified. But they do not confirm the specific emotional details told in the voice of an insider. Put another way: the frame is real, while the emotional filling is largely the account of unnamed people.

In the language of data work, this is a record with two layers of reliability stacked on top of each other. The lower layer is solid. The upper layer shimmers. The most damaging thing is when the upper layer is told in such a confident voice that readers merge the two into one.

But the most discussion-worthy part remains the pipeline. Picture such a record landing in a library that feeds predictive models. An automated assistant answering football questions reads it. An aggregator picks it up. A sentiment model scans it and assigns a score. None of them was designed to ask the first and most important question: does this record talk about football. They were only designed to process what already sits in their drawer.

This record exposes three levels of risk. The first is domain misclassification: an item that is not football is labelled football and enters the football pipeline. The chance of this happening in a single case is very low, but the chance of it happening at a scale large enough to pump noise into a system is not low at all. A small rate across millions of daily records is a quiet river of noise.

The second is source quality. Most of the specific details rest on anonymous sourcing with no independent verification mechanism. This is the kind of information that, if reused in a data product, replicates itself with no one able to trace the origin.

The third is downstream contamination. When a bad record slips through the gate, it does not stop there. It flows down into the lower layers: models, aggregators, automated answering tools, and writers like me, who sometimes trust a data table simply because it looks organised.

I want to be clear about how I read analyses of this kind, because it is in my professional blood. When a framework returns only empty values, the first reflex of a newcomer is to think the tool is broken. The reflex of someone experienced is to think the question itself was set wrongly. The framework is not broken. It works exactly as designed: it answers a question it was programmed to answer, even when that question has nothing to do with the subject in front of it. The breakdown is not in the algorithm. It is in the step that decides what is placed in front of the algorithm.

Here I see a parallel with my own craft of tactical analysis. We are used to praising sophisticated models, from models predicting goal probability from shot location to models quantifying the value of a possession phase. But I have learned that the value of a model does not lie in how complex it is, but in whether it knows how to refuse to answer. A good model is not one that always has an answer. It is one that can say: I do not have enough data to answer this, and I will not invent one. That mislabelled record should have been returned at the gate with a simple line: this subject does not belong here.

The counter-intuitive angle, and the place where I want people to argue with me: the frightening mistake is not the stray record itself. The frightening mistake is the belief that the pipeline is neutral.

We are used to imagining the pitch as the place where football truly happens, and the data on the screen as a faithful mirror of that pitch. But that mirror is polished by thousands of human decisions: which sources to choose, how to label, what to ignore, what to keep. Each of those decisions is an unwritten assumption. And a system made entirely of hidden assumptions is confident in exact proportion to its ignorance of what it is assuming.

Tracking data does not say who is right, it says who showed up on time. I still use that line whenever I am asked why I do not worship numbers. But it holds even more at the upper layer: a pipeline does not say what the truth is, it says what managed to get through its gate.

The first reaction of many people to an error like this is to demand more barriers. More filters, more checks, another labelling layer. I understand that reflex, and part of me agrees. A semantic check before ingestion could stop records like this. But adding barriers without changing how we think breeds a new kind of confidence: the belief that because there are more gates, whatever gets through is far more trustworthy. Each new gate brings a new assumption about what looks like football. A human-interest item that mentions a club name in its comments will slip past a keyword-based gate. Barriers cannot replace humility.

There is one more point I want to push further, one that may make my readers nod while system builders squirm. The worrying thing is not the bad record. The worrying thing is the possibility that the bad record is a symptom, not an outlier. If the labelling layer can fail on a case so obvious that any reader spots it in three seconds, how many fuzzier cases is it getting wrong, cases nobody has the patience to click and check. The most damaging errors are the invisible ones. This record is harmless only because it is wrong in a glaring way. A subtly wrong record, right topic but wrong fact, right statistic but wrong context, right source but wrong interpretation, will drift through the gate unnoticed.

In football we have a name for that kind of damage: the feint the whole stadium fails to see, until the ball is already in the net. France 4-3 Argentina, the day organised chaos beat gifted disorganisation. I still use that line to remind myself that victory does not always belong to the prettier side, but to the side that controls the chaos. Controlling chaos, at the data layer, means knowing clearly what you do not allow through. Mancini's Italy did not own the ball, they owned the moment. They won by choosing the right moment to intervene, not by being everywhere at once. A good pipeline must be built the same way: choosing the right moment to refuse, rather than trying to swallow everything and hoping for luck.

If you have read this far and still wonder whether one stray record in a news system deserves this much space, my answer is: the smaller it is, the more it matters. The thing that slips through loudly gets caught by someone; the thing that slips through in silence simply stays, becoming part of the foundation on which thousands of later decisions rest.

The work now is not to convict a single record. It is to ask the system a question it has never been asked: before analysing something, are we sure it belongs here. And if the answer is no, do we have the courage to send it back where it belongs, instead of cramming it into the framework.

For me, a tactical analyst who earns a living from clean data, this error is not a stain. It is a road sign. It says that every number I put in an article must stand up to the first and simplest question: does this data belong to the question I am asking. Because many bad analyses do not fail at the conclusion layer. They fail right at the input layer, where nobody stops to ask.