Wrong Labels, Wrong Lineups: The Classification Gap Inside Professional Football Data

In the news feed I maintain for a K League 1 club, there is an entry on row one thousand one hundred and forty-seven. The industry tag reads: football. The headline looks unremarkable. The body below it, however, describes an Apple TV series starring Jessica Chastain, adapted from a 2026 Cosmopolitan article, dealing with domestic terrorism on American soil, with a spring 2027 release window after several schedule slips.

I read the whole thing. No club. No player. No coach, contract, transfer fee or single minute of play. Twenty-five information points in the source document, and all twenty-five belong to the entertainment industry.

What stopped me was not the error itself. It was its position: sitting in the labelling layer — the first layer, the one every downstream model assumes to be correct. Thirteen years in this trade taught me something simple: mistakes at the last layer are loud, mistakes at the first layer are silent. And silent things are what end a season.

FOUR LAYERS OF A DECISION

Every decision at a modern professional club passes through four layers. Collection: event data from vendors, video, scout reports, media feeds. Labelling: sorting each piece of information into a drawer — football, medical, market, finance, media. Modelling: computing with those drawers. Decision: the coaching staff and sporting director act.

The last three layers have someone checking them. The first layer has almost nobody.

I used to think this was dry technical talk, until I realised it mirrors a situation on the pitch. When a team loses the ball in midfield, the viewer's eye snaps to the ball. But the goal is decided elsewhere — in the position of the left-back who just pushed up, in the thirty-metre gap behind him. The label "counter-attack" says nothing about that gap.

A gap does not vanish on its own; it just changes its name to defeat.

Over the past decade, football analytics has shifted from counting to classifying. Event data vendors no longer merely record that a pass happened in a given minute; they tag it — progressive pass, line-breaking pass, pass under pressure. Every tag is a claim about intent. And every claim about intent can be wrong.

What makes it dangerous is that tags do not live alone. They stack. A mis-tagged pass skews a player's ball-progression metric. That metric skews his heat map. The heat map skews his role description. The role description goes straight into the scouting report. The scouting report goes straight into the spending decision.

A wrong label in the first layer can become a wrong contract in the last. The distance between those two layers is worth several months of wages, depending on the club.

WHEN A LABEL IS WRONG, THE DATA'S SHAPE IS WRONG

What matters is the consequence. An entertainment item slipping into a football dataset does not simply add one junk row at the bottom of a table. A wrong label does not create noise — it creates a sample with a different shape, and every downstream model will learn exactly that wrong shape.

Think about it the way an analyst would. If your model is trained to recognise "football articles" from the frequency of words like match, player, tactics, then a piece about domestic terrorism using words like attack, target, plan will sail through the filter easily. The labelling layer guesses by semantics. Football uses a great deal of military vocabulary. That intersection is what makes this error far from rare.

I have met that exact mechanism at a much smaller scale, and in a far more sensitive place: the pitch itself.

In 2026, when the pandemic left K League 1 stadiums empty from May to August, I sat at SportsData Korea and collected data from 142 matches without spectators, comparing them with 142 pre-pandemic matches. Home win rate fell from 47 percent to 41.5 percent. Average goals per match rose by 0.7.

The usual explanation is "loss of home advantage". But home advantage was never a single variable. It is one label stuck onto at least five different things: crowd noise acting on referees, psychological pressure on away players, familiarity with the surface, travel schedules, and the match tempo a crowd generates. An empty stand does not lose you the match; it strips away the decorative layer of emotion.

When the crowd disappears, all five variables disappear at once. But if your model bundles them under the single label "home", you can never separate them — and you will draw the wrong conclusion about cause. Worse, you will draw the wrong conclusion about the future: you will predict that home advantage has permanently declined, when all that declined was one of its five components.

That is the lesson I carried through my career: a label is a tactical assumption written in a single word.

I also learned a harsher lesson from that same project. I built a predictive model based on pressing intensity and attacking start positions, then kept rewriting it because I wanted perfection. The report was only finished in December. My boss still rated it highly, but a colleague said something I have never forgotten: good data, but published too late it is no different from predicting after the match.

Since then I write short, predictive pieces, accept uncertainty, and state the data limitations at the end of every article. A correct label delivered late is still a useless label.

THREE WRONG LABELS AND WHAT THEY COST

The first, at the 2026 World Cup.

That night I sat in front of a screen in Incheon, 23 years old, watching South Korea play Germany. The score was 2-0. Most viewers talked about luck, about a shock, about the defending champions going out in the group stage. I spent three days rewatching the tape and counting. Germany played the ball into the box 87 times. Shots on target: two.

Germany's failure came not from a shortage of talent, but from a surplus of certainty.

The label the media stuck on that match was "dominance". It was right about possession, right about chance volume, and completely wrong about structure. Germany pushed its line high for most of the match, exposing the space behind the back line — and with every Korean counter, that space grew wider. The label "dominance" hid the fact that the team was dominating an irrelevant area.

I wrote a five-thousand-word blog about it. It was called convoluted. But an editor at a tactical analysis site got in touch, and that was the start of my writing career. I learned that deep analysis needs structure, that I had to select data rather than dump it, and that one gap diagram argues better than ten paragraphs of description.

The second, at the 2026 World Cup.

Morocco reached the semi-finals — the first African team ever to do so, and the first Arab team as well. The world talked about spirit, about will, about a desert fairytale. I spent five days rewatching their six matches and logging every ball loss.

The answer lay in a number: 2.3 seconds. That was the average time for Morocco to shift from an attacking state into a 5-4-1 shape after losing the ball. Within those 2.3 seconds, Achraf Hakimi — who advanced an average of 58 metres per match — had already begun dropping, while Azzedine Ounahi covered the right channel until the full-back was back in position.

Morocco did not need to control the ball; they controlled what the opponent was allowed to dream.

The label stuck on them was "counter-attacking defence". Correct, and useless. Because the label says nothing about 2.3 seconds. And those 2.3 seconds are precisely what made every through-ball from opponents meaningless. I wrote a 3,500-word piece with twelve heat maps. It reached 1.2 million views, and on the back of it a K League club invited me to work as a part-time tactical consultant.

Between two phases of play, time exposes the decisions the naked eye misses.

The third, at a much smaller stadium, with no camera recording it.

I once received a scouting report on a player whose first line read: attacking full-back, good pace. That label came from his starting position in the formation. But when I redrew his heat map across ten matches, a different story emerged: he pushed high for only about 40 percent of the time, and for the other 60 percent he played as a pure full-back, sitting lower than his left-sided centre-back.

If the club bought him based on the label, they would be buying a player profile that does not exist. If they bought him based on the heat map, they would be buying a defensive full-back — far cheaper, and far better suited to their system.

LABELS DO NOT ONLY COME FROM MACHINES

This is the part rarely discussed. Mis-labelling is not a data speciality. Humans label far more, except we do not call it labelling — we call it experience.

Every time a scout says a player has leadership qualities, he is sticking a label on an undefined set of behaviours. Every time a coach says this lad fits my system, he is sticking a label on an untested hypothesis. Neither of them calls it bad data, because both are interpreting direct observation. But direct observation has an error rate too; it is just never recorded anywhere.

And in the transfer market, one group labels more professionally than anyone: agents. They are the market's largest hidden cost, and their primary product is the label itself. A player branded "top young talent in Asia" can sell for three times the fee of an identical player by every metric who lacks that label. The noise they generate distorts prices, and distorted prices distort tactical decisions too: the club feels obliged to start the expensive signing, even when he does not fit how the team presses.

Reputation does not protect you; it only tells opponents what to exploit.

I have seen this most brutally at goalkeeper, where the market labels crudely. The label "ball-playing goalkeeper" is now priced above the label "good shot-stopper", while the thing that decides points remains the saves. Distribution has been sanctified to the point where a keeper with a declining save rate still commands a high transfer fee, simply because he can pass short under pressure. That is a label being paid to exist independently of results.

There is a very close example showing this problem is not abstract: VAR. The intervention threshold is written as "clear and obvious error". It sounds rigorous. But clear and obvious are words that need labelling before use, and nobody labels them. The result is that for the same collision, three referees can reach three different decisions, and all three are being honest to the standard in their heads. The space for subjective judgement inside VAR is far larger than the written law suggests. That is precisely a labelling problem, placed where an entire stadium can see it.

THE CONTRARIAN VIEW: THIS ERROR IS NOT AN ERROR

The natural reaction to an entertainment item tagged as football is to treat it as a fault to be fixed. I do not think so. I think it is the inevitable consequence of how modern data systems are designed.

Every classification system must choose between two kinds of error: missing something it should have kept, and keeping something it should have discarded. In football, the first error costs far more. Missing a good player means losing an asset. Keeping a junk item costs an analyst a few seconds. So every data pipeline is tuned to catch wide, and every wide-catching system has a false-positive rate. One hundred percent is an extreme figure in this case, but the mechanism is entirely ordinary.

That leads to a harder conclusion. The biggest risk is not that a junk row sits in the dataset. The risk is that we habitually fix errors at the model layer — adjusting weights, adding variables, changing algorithms — while the error sits in a layer nobody bothers to open.

And there is one more blind spot, inside people like me. An analyst likes to think of himself as the one checking other people's labels, but the analyst's eye is also a labelling machine with its own biases. When I rewatch a match I have already reached a conclusion about, I tend to count the phases that confirm it. I once did this with a team I support, and it took me two weeks to realise I was analysing an image, not a match.

Every tactic is a hypothesis until the opponent forces you to answer. So is every dataset.

WHAT TO DO NOW, WITHOUT WAITING

Data only means something when we ask at the right moment; ask wrongly, and every number is noise.

For people in my trade, the task is not to remove all labels — without labels there is no analysis. The task is to treat every label as a hypothesis with an expiry date, conditions, and a moment when it stops being valid. A label with no expiry date is a label that will soon become a prejudice.

For a club, the task is far cheaper than people assume: take thirty randomly selected labelled rows from the past week, open each one, and read. No algorithm needed. Just a person willing to read. If those thirty rows are clean, your pipeline is fine. If not, you have just saved yourself a contract.

As for the mis-tagged entry on row one thousand one hundred and forty-seven, I have left it where it is. Not out of laziness. Because it is the cheapest reminder an analyst can have: that the first layer of every football story never checks itself. And this weekend, when you see a team described as high-pressing, try asking a different question: in the first three seconds after losing the ball, who moves first, and in which direction. The answer lies outside every statistics table, and it is often the only correct answer there is.