Mislabelled in the Transfer Window: When Football Data Deceives Itself

On the night of 12 July, in an office overlooking the Vieux-Port in Marseille, I opened a file that the system had tagged "football" and was about to feed into my transfer-evaluation workflow. I read the first line and stopped. No clubs. No players. No league table, no transfer fees, no wage bill, no release clause. The content was about the first trailer for Netflix's Narnia film, directed by Greta Gerwig, set to Led Zeppelin's "Stairway to Heaven."

I sat still for a few seconds. This was not the writer's fault. It was a label mismatch: a document belonging to cinema, tagged as sport, and pushed into the very pipeline I use to value deals. In twenty-nine years of observing this industry, I have learned that the most dangerous mistakes are not the large errors that stare you in the face, but the small errors buried in the base layer, the classification layer. A single mislabel at the input can become a wrong conclusion at the output, and nobody in the meeting room notices until the contract has already been signed.

I tell this story not to catch out a particular system. I tell it because it exposes what professional football faces every single day of the transfer window: we have built an industry worth tens of billions of euros on data, yet we spend very little resource checking whether that data actually belongs to football.

Start with the number. A club in a top European league now tracks roughly three thousand players in its scouting database. Each player may carry hundreds of fields: date of birth, height, preferred foot, minutes played, goals scored, expected goals, key passes, pressing index, distance run per match, top speed, injury history, salary, contract length. Multiply it out and you have millions of data points updated continuously. No analyst reads them all with their eyes. We rely on labels to filter.

The label is what I want to talk about. The label is the classification layer that decides which drawer of the organisation's brain a document, a video, a scouting report will go into. When the label is right, the system runs smoothly. When the label is wrong, the entire downstream chain is poisoned without a sound. A mislabel does not raise an error message. It sits there quietly, waiting to be read, to be believed, and to be used to make a decision.

What troubles me is the frequency. Over the past three months, when I reviewed my own intake log, I found the rate of topic-misclassified documents to be under one percent. That sounds small. But if you ingest five thousand documents a week, one percent means fifty wrong documents slipping into the system every week, two hundred a month, nearly two thousand five hundred a year. For a club making data-driven decisions, that is a stream of noise flowing silently through every decision.

I remember a scouting meeting at a Ligue 1 club I once advised. An assistant presented a shortlist of ten full-backs. One name stood out with exceptional attacking metrics. Three weeks later, when we re-checked, it turned out that player's data had been merged with the data of a midfielder with the same surname but a different person, a different league, a different season. An entire profile built on a wrong label. The club had nearly opened negotiations for a player they had assessed completely wrongly.

That is why I always tell younger colleagues: our job is not to read numbers. Our job is to check whether the number belongs to the right person.

Modern football runs like a transmission system. A signal travels from the pitch to the satellite, from the satellite to the server, from the server to the spreadsheet, from the spreadsheet to the decision-maker's head. At each stage the signal can be distorted. But the most dangerous stage is the first: labelling. Because if the label is wrong from the start, every calculation afterwards becomes perfectly meaningless. You can build a probability model as sophisticated as you like, but if you hand it a document about the Narnia trailer and call it football, it will return an answer about football, and the answer will be wrong, and it will be confident.

I have seen the same thing at a larger scale. In 2026, when I published my analysis of Marseille against Paris Saint-Germain, my expected-goals data showed Marseille created the more dangerous chances despite losing three-nil. I received hundreds of hostile comments. But the notable thing was not the disagreement. The notable thing was that at least seven critics quoted my figures with completely different numbers, taken from unclear sources, and declared my data wrong. They did not check the label. They did not ask where the data came from. They simply reacted to the number.

An error in the label layer never reports itself. It wears the clothes of truth and waits to be believed.

Three months later, PSG's metrics dipped and they lost to Lyon. My read was vindicated. But I did not celebrate. Because I realised that what I had won was not a debate about football, but a battle over the integrity of data. And that battle never ends.

Data is the only thing I trust after watching too many promises break. But I have also learned that trust in data must come with a discipline of checking data. A correct number from a wrong source is a more dangerous number than a wrong number from a correct source, because it is more believable.

Look at the transfer window. This is the moment when our industry consumes the most data and checks the least. Every day, thousands of transfer rumours are born, spread, and die. Each rumour is a data point. Each data point needs a label: tier-one source, tier-two, tier-three, verified, unverified, agent's rumour, club's rumour, fan's rumour. When labels are ignored, a rumour from an anonymous account can be processed on equal footing with an official club announcement.

The transfer market does not buy players, it buys stories. And stories, when unlabelled, become false truths.

I have spent years tracking how transfer rumours are made. The pattern repeats. An account posts a line. A page aggregates it. A journalist cites the page. A club reads the article and assumes something is happening. An analyst puts it in a report. And so a wrong label is born, not from a liar, but from a chain of honest people who did not check the source.

In the transfer window, noise is always larger than signal. The task of an analyst is not to collect more noise, but to filter noise and keep the signal. And to filter, you need labels. The label is the first and most important filter.

I am often asked how to know whether a transfer story is credible. My answer always disappoints the person asking. I say: look at the structure, not the emotion. Look at the release clause, the wage bill, the remaining contract length, the player's age, the club's positional need, the fixture schedule, the injury history. Those are signal. The rest is an unverified label.

But there is one thing I have never said publicly until now: those signals themselves can be mislabelled. A release clause can be misreported. A wage bill can be estimated from stale sources. A player's age can be wrongly listed on data sites. I once saw a player profile with a date of birth wrong by two years, which rendered every judgement about his development curve meaningless.

This is where the Narnia trailer story becomes meaningful for football. It is not a joke of the system. It is a warning about what happens when we build decision machines on foundations we do not check.

A club can spend tens of millions of euros on a player based on a data model. But if that model is fed mislabelled data, the club is buying a shadow. And that shadow will carry a real price.

I once worked with an analytics group using a machine-learning model to predict the potential of young players. The model performed well on test data. But in the real world, it consistently overvalued a group of players from a particular league. The cause was not the algorithm. The cause was that league's label: it had been merged with another league of far higher competitive level. The model was not wrong. The input data was mislabelled. The result was a batch of scouting reports that overvalued an entire generation of players.

This is exactly where I think football analytics is at its most naive: we believe in the algorithm, but not in the labelling process. We invest in models, but not in data hygiene. We hire the best data scientists, but not enough people to check whether the data actually matches the topic.

Numbers have no bias. The bias lies with those who lack numbers. But I would add a clause: the bias also lies with those who have numbers but do not check their source.

So what creates a mislabel? I have observed three main causes.

The first is keyword matching. An automated system scans text and looks for words like "transfer", "contract", "value", "record". A film article discussing Netflix's "acquisition of rights" and a "music licensing deal" can easily be caught by those keywords. The system does not understand context. It only matches patterns. And in football, where financial language and entertainment language overlap heavily, pattern-matching errors are everyday events.

The second is the default label. When a document's topic is unclear, a system sometimes assigns it a default label rather than leaving it blank. That default label may be football, if football is the most common topic in the system. This is a subtle design flaw: the system would rather be wrong than blank, because blank requires a human to step in.

The third is a stale schema mapping. A data field that once meant one thing, after the system changes, comes to mean another. But the old label remains, and it continues to be used as though nothing had changed.

All three causes share one feature: they do not produce a visible error. They do not crash the system. They quietly poison the result.

In football, the consequences of a mislabel can be measured in money. A misjudged player can cost a club tens of millions. A model fed contaminated data can lead a team to build its tactics around an unsuitable player. A scouting report built on a wrong label can cause a talent to be overlooked or an ordinary player to be bought at a high price.

I once saw a club sign a striker based on a very high expected-goals figure. But that figure was calculated from a season in which the player competed in a league with far weaker defending than the destination league. The league label was not adjusted. The player arrived at his new club and scored a third of what the model had predicted.

This story is not rare. It is the rule. And the sad part is that it is always retold as a story about the player's failure, not about the data's failure.

This is where I want to argue against myself, because I know I have a tendency to believe everything can be reduced to a perfect spreadsheet. If every mislabel could be detected by a better algorithm, the story would be simple: just build a better algorithm. But reality is more complex.

Some mislabels are not in the machine, they are in the person. A scout looks at a player and labels him "high potential" because he resembles a player he once admired. A coach labels a player "unsuitable" because of a personal memory. A sporting director labels a deal "priority" because of fan pressure. These labels are in no database. They are in people's heads, and they are more dangerous than any algorithmic error.

A risk model saves no one, but it gives them a chance. And that chance only has value if the foundational labels are checked before the model runs.

I have come to realise that the most important skill of a football data analyst is not building models. The most important skill is knowing when to say: I do not have enough information to conclude. In an industry where everyone wants a decisive answer, the ability to say "I don't know" is a form of professional courage.

When my system received a document about the Narnia trailer labelled football, the correct response was not to try to find football meaning in it. The correct response was to stop, flag it, and say: this document does not belong to the topic. Timely silence is a valid output.

This is what I want young colleagues to remember. In the transfer window, when everyone around you is shouting about a deal, when news sites are reporting on a player, when fans are demanding the club act, the pressure to produce a conclusion is enormous. But the worst conclusion is the one produced from a wrong label.

I once watched an analytics room get swept up in a transfer rumour and spend two weeks building a detailed report on a deal that never existed. When the truth was confirmed, no one among them asked: why did we believe it. They simply moved to the next rumour. The process learned nothing. The same error will repeat.

That is why I began applying a rule in my team: every incoming document must pass an entity test. If a document contains no club, no player, no competition, or no verifiable match event, it is returned. The rule is so simple it borders on naive. But it has blocked countless mislabels.

The entity test is not a perfect solution. It will not catch subtle errors, for instance a football document that discusses an irrelevant aspect. But it is a first fence, and in a complex system, the first fence is often the most important one.

Look at the big picture. Football is going through a data revolution. Clubs hire data scientists. Leagues sell data to analytics companies. Players are judged by metrics that did not exist twenty years ago. This revolution has brought much good. It helps find talent in places few notice. It helps small clubs compete more fairly. It helps expose long-standing bias.

But this revolution also creates a new weakness. When you build a house on ground, you test the ground. When you build a decision on data, you must test the data. And data does not test itself.

I believe that in the next ten years, a club's greatest competitive advantage will not be the best model, but the best data-hygiene process. Because in a world where every club can access the same set of metrics, the winner is the one who knows which metric is real.

Go back to Marseille and PSG in 2026. What I learned from that debate was not that I was right and others were wrong. What I learned was that in an industry where information moves faster than truth, the person who holds on to one correct number will beat the person who has many numbers but checks no source.

When I was criticised for using expected goals to argue against PSG's win, the criticism did not come from data. It came from emotion. And emotion, though part of football, is not a credible label.

I chose to trust the misses. PSG won that year, but I chose to trust the misses. Because misses carry more information than goals. A goal can be luck. A miss exposes a person's true decision logic.

And in data football, a mislabel is a miss at the foundation layer. It is not pretty, it is not remembered, but it tells you how the entire system is operating.

I wonder: if every club spent one percent of its analytics budget on data auditing, how much would it save? I do not have the exact answer. But I know one thing: the cost of a wrong label is never recorded on any balance sheet. It lives in bad deals, disappointing seasons, missed talents.

Croatia in 2026 taught me that heroes also have biological limits. And a data system also has its biological limit: it is only as good as its foundational label layer. A team can run three hundred and eighteen kilometres in the group stage, but if the fitness chart is misread, people are still surprised when they collapse in extra time.

I write these lines in the middle of the transfer window, when the noise is at its peak. Every day there are hundreds of rumours. Each rumour is an unverified label. My task, and that of anyone doing this job seriously, is not to let that noise become truth merely because it has been repeated enough times.

People see a comeback, I see a chart breaking. And in this case, people see a football document, I see a wrong label waiting to be corrected.

The question I leave you with is not how to build a better model. The question is: when did you last check whether the data you trust actually belongs to the topic you think it does? If you cannot remember, then perhaps it is time to look again at the foundation layer of every decision you are making. Because in football, as in any complex system, the most lethal mistake always begins with a mislabelled tag — and ends with a belief nobody checked.