A Marvel Film Landed in the Football Database: When the Labelling Engine Gets It Wrong, Who Pays?

On 2 June 2026, at 6.47 a.m. Hamburg time, I opened my machine and found a new item sitting in a folder called football. The piece ran less than a page. It concerned Cinépolis and Cinemex, Mexico's two largest cinema chains, confirming an early screening schedule for Avengers: Doomsday: midnight showings, presales opening ahead of the release date, Mexico opening one day earlier than the United States, and chain websites crashing for hours under demand. I read it once. Then I read the label at the top of the record again: Domain — football.

There is not a single club in it. Not a single player. No scoreline, no line-up, no contract, no league table, no regulation. There is only a classification engine that named the wrong thing, and a second-stage analysis process — the very process I sit here to run — forced to answer the hardest question in my trade: when there is no data, what do you write?

I chose to write N/A. Not out of laziness, and not out of cowardice. Because it was the only honest thing left.

One folder, twelve lines of data, one wrong label

I have worked with football data since I was sixteen. In 2026 I sat in my bedroom in Hamburg, rewatched twenty-three HSV U19 matches, mapped the movement on one hundred and eighteen attacking sequences, and found that left-back Josha Vagnoman pushed an average of fourteen metres higher per possession. I wrote a two-thousand-one-hundred-word piece recommending the staff move him to wide midfield. It got three hundred and seventy-six views. One youth coach at the academy read it, wrote back, and invited me to a coaching meeting.

I sat at the back of the room and listened to four men argue about a 4-3-3 for forty minutes. What I remember is not their conclusion. It is that they were willing to sit and listen to a sixteen-year-old present one number.

Since then, every analysis I write starts with a raw fact. A gap. A percentage. A minute. And I always ask myself before writing: does this fact actually say what I am about to claim it says?

That question is why I sat for a long time in front of a football folder containing a Marvel press item.

In the sports data industry, every item passes through a chain. Collection first. Extraction second. Then topic labelling — what the trade calls the domain label. That label decides where the item is routed: tactical analysis, club finance, rules and governance, or media and public opinion.

Labelling is the cheapest step in the whole chain. It is also the most dangerous. A wrong label does not damage the source item — the item stays there, intact. It damages everything built on top of it.

I have seen this at a smaller scale. In 2026, when the pandemic emptied Bundesliga 2 stadiums, I analysed eighty-seven behind-closed-doors matches for the Hamburg data firm ProData. Home win rate fell from 43 percent to 34 percent; average goals fell from 2.6 to 2.1. I dug into St. Pauli, the club I love, and found their defensive block pressed toward the touchline 18 percent more often without crowd noise.

But to get those numbers, I had to drop eleven matches from the sample because their home and away labels had been misassigned during a fixture-list update. Eleven out of eighty-seven is not enormous. Had I not caught it, that 34 percent could have been 36, or 32, and the whole conclusion about crowd noise would have tilted with it.

87 matches, 43% to 34%, 2.6 to 2.1 — I thought I was reading numbers. I was reading the loneliness of the game.

The lesson sits elsewhere: a small labelling error at a low layer can flow upward into the final conclusion, and by the time it gets there it is wearing the clothes of data. It has a chart. It has citations.

That football folder was the same error class, one scale up.

Nine analytical dimensions, nine returns of N/A

The second-stage process I run has nine dimensions. Tactical and technical. Club finance and the transfer market. Results and the public-opinion cycle. League landscape and team positioning. Rules and governance. Management and the dressing room. Risk profile. Media narrative and expectation. Football industry transmission.

I opened each one against the Marvel item.

The tactical dimension asks how a team builds, how it presses, who receives in the third line. The item has no team. No PPDA, no xG, no possession share. It returns N/A.

The finance dimension asks about revenue structure, wage share, financial-fair-play exposure. The item has cinema ticket revenue. Box-office revenue and football broadcast revenue do not share an accounting basis. It returns N/A.

The results dimension asks where a team sits against expectations and how its recent form looks. There is no table and no fixture. The expectation in the item is a film audience's anticipation, which shares no unit of measurement with expectation in a league table. It returns N/A.

The league dimension asks what tier a club occupies. The item contains a release-window race between Mexico and the United States — distribution strategy, not competitive hierarchy. It returns N/A.

The governance dimension asks about FFP, registration, disciplinary exposure. No football entity here is regulated by FIFA, UEFA, a national association or a competition organiser. It returns N/A.

The management dimension asks who decides, how coaching authority is structured, whether the dressing room is stable. The decision-makers in the item are a studio and two cinema chains. It returns N/A.

The risk dimension asks for sporting, financial, personnel, rules and reputational risk levels. There is no football asset to score. It returns N/A.

The narrative dimension asks what story is being told and where the heat cycle sits. The story is nostalgia for midnight premiere culture. It returns N/A.

The transmission dimension asks how effects flow through academies, agent ecosystems and derivative markets. Not one link of the football value chain appears. It returns N/A.

Nine times N/A.

In my trade, N/A is treated as failure. People pay for analysis, conclusions, forecasts. But there is a distinction few sports data people will admit: 'no data' and 'data that says nothing' are entirely different things.

No data means go collect more. Data that says nothing means you are asking the wrong question. For the Marvel item, the second is true. No additional volume of information about Cinépolis and Cinemex could help me analyse a team's tactics, because the object of analysis does not exist.

This is where the trade must choose between two roads. The first is to force data into a mould. You could write that a midnight screening resembles a kick-off whistle; that crashing cinema websites resemble a sell-out stadium; that a Mexico-versus-US release race resembles a title race. It reads smoothly. It reads professionally. It is metaphor, not analysis.

The second road is to keep the N/A, log the misclassification, and move the item where it belongs.

Where the wrong label actually does damage

The item itself is not at fault. Cinépolis and Cinemex published a schedule, Mexican audiences responded, websites buckled — all true and all valuable within its own field. The fault sits a layer above: a routing system stamped football on a cinema event.

If nobody catches it, no club is harmed. The data store is.

Picture a training set for a match-result model, assembled from hundreds of thousands of items labelled football. If a small share of them are actually about films, music or elections, the model is learning from noise. Small noise may not break a model immediately, but it makes the model harder to explain, harder to audit, and worst of all it creates a grey zone nobody wants to own.

At the dashboard level, consequences arrive faster. An editor on deadline opens the football feed, sees a Cinépolis and Cinemex entry, and has two choices: delete it, or write something. Many choose the second. That is the moment a data error becomes a published falsehood.

I once sat in a ProData meeting where someone proposed skipping the label check to hit a quarterly deadline. The argument was reasonable: low error rate, high review cost, invisible to the client. I objected and was called a perfectionist. Six months later a client found four matches with the wrong result in their reports. Four matches. Nobody lost a contract, but part of the trust was gone.

The real blind spot of the analysis trade

Here I want to be blunt, even if it is uncomfortable for my own colleagues.

The classification failure in a Marvel item is a symptom. The disease is elsewhere: sports analysis is paid to always have an answer.

Look at how we handle a match. Team A wins 1-0 with three shots on target and 39 percent possession. Team B loses with nine shots but faces eleven tackles inside the box. Within thirty minutes of the final whistle, hundreds of pieces are published, and most call Team A's win a tactical lesson, the coldness of a perfect machine.

I wrote that piece. In 2026, at the World Cup in Russia, a football fan site asked me to analyse the semi-final where France beat Belgium 1-0. I sat for a long time with the numbers, torn between Belgium's attacking beauty and France's coldness, and finally titled it 'The Heart Behind the Tactics'.

Belgium had 9 shots, France only 3 — but the ticket belonged to the colder side, not the side that dreamed more.

I still think I was right to write it. I also know what I was not allowed to do: I was not allowed to say that three shots on target proved a football philosophy. Three shots are three shots. One match is one sample. A sample is always smaller than the story people want to tell from it.

The heart behind the tactics — I do not ask which team deserved to win. I ask which team dared to lose as itself.

The blind spot is here: we teach each other how to read data, but not how to refuse data. We have rituals for reaching conclusions but no ritual for saying there is not enough to stand on. An analyst who returns N/A nine times in a row is judged as weak, even while being the most correct person in the room.

That pressure does not come from data. It comes from incentive structure. Feeds need filling. Page views need raising. Algorithms need new content every day. When a system rewards volume, people produce volume, even when the raw material only covers half of it.

Deeper down, I see the same disease in youth development. Academies below under-eighteen level are pushed to run more, heavier, earlier, to get results in youth competitions. Youth results are an easy metric to measure and an easy one to report. Foundation technique is hard to measure and takes a decade to show. So people pick the easy metric. Just as people pick writing about a midnight screening over logging that the label was wrong.

There is a test I apply to myself before publishing. I ask: if my only data source vanished, would this conclusion still stand? If the answer is no, the conclusion is borrowing the authority of data rather than being produced by it.

Readers never see the label, but they see the consequence

There is a reason misclassification is hard to catch: the end reader never sees the label. They see the product. If the Marvel item had been rewritten as a football piece, readers would read it, feel something was slightly off, and mostly keep going, because the prose flows and they trust the outlet's brand.

Fluency is the best camouflage for false information.

This holds for the most serious analysis too. A piece about the space behind a left-back can be entirely correct about geometry and still wrong about meaning, if the author never checks whether the team left that space deliberately. That is why I always return to the raw footage after building a hypothesis from numbers.

The gap behind him was exactly fourteen metres wide — but the real dead zone sat where nobody bothered to look.

The place nobody bothered to look, in this case, is the labelling layer.

What is needed is a gate before analysis

Fixing this is not about adding reviewers. Humans cannot read the volume modern systems process daily. The fix is a simple gate: before an item moves to the analysis layer, the system must confirm it contains at least one football entity. A club. A player. A competition. A governing body. A match.

If none of those is present, the football label must lose confidence, not keep it.

This is something anyone in data work knows and often forgets: a confidence threshold is not a technical detail. It is professional ethics. A wrong label caught at the entry layer costs seconds. The same label passing through five layers, into a dashboard, into a client report, into a piece cited three thousand times — that cost is no longer measured in seconds.

I remember defending my data on a 2026 World Cup contract. I presented to four people. One of them asked me three words: 'Are you sure?' It took me fifteen minutes to prove I was. And when I admitted there were three items I was not sure about, the contract was still signed. They signed not because I was confident, but because I knew the boundary of my confidence.

I sat at the back of the room, watching them argue about a 4-3-3 — the biggest lesson was that they were willing to listen to a sixteen-year-old.

Honesty about the limits of your own understanding is the greatest asset an analyst has. It is not taught in schools. It only comes from having been wrong and having owned it.

What I keep from an item that was never mine

The Avengers: Doomsday item in Mexico does not belong in my folder. It belongs to another industry, with other professionals, and I have nothing to say about the quality of their work — they are doing their job. Mine was to move it back and log the error.

But had I stopped there and closed the machine, I would have missed the biggest lesson. Labelling failure is not a private matter for one routing system. It is a miniature of how we handle information generally: label fast, process fast, and rarely revisit the first label.

I have seen the same in how people discuss a new manager after three defeats. The label 'failure' gets attached before the data is thick enough. I have seen it in how an eighteen-year-old is judged after one season. The label 'prospect' or 'not good enough' is attached while the player's development curve is long and depends on roughly ten other variables.

376 views do not make a tactical mind — but one youth coach who reads to the last word might.

By the same logic, one off-topic item does not make a data crisis. But one gate added in the right place can prevent thousands of similar errors.

If my readers — young coaches, sports students, junior data people — take only one line from this piece, I do not want it to be a number. I want it to be the thing nobody taught me ten years ago: saying 'I do not have enough to go on' is not surrendering to a hard question. It is how you ask the question properly.

I still check my own data labels every Monday morning. Nobody requires it. But in 2026, when I removed eleven matches from a sample of eighty-seven, I understood something: the frightening thing is not the error. The frightening thing is an error nobody sees.

The empty stadium cut home win rates from 43% to 34% — the human being is the most hidden tactical variable of all.

And across every data layer, that human layer is not in the number. It is in whether someone sits down to check the number.

From next season, when young coaches read my work, I want them to ask one question before every conclusion: if this analysis had no numbers in it, would my conclusion still stand? Whoever dares to ask that every week will move a little slower than their peers. They will also be the only ones still holding clean data five seasons from now.

As for me, that morning, I moved the Avengers: Doomsday item to its correct folder, logged the misclassification, and started the day with a question pinned on my screen: how many more wrong labels are sitting in this data store, waiting for someone to read to the last word?