Mislabeled Signals: What Football's Data Pipeline Learns From a Classification Error

Last month, a colleague in the data analysis department sent me a 38-page scouting report. Page one said: left winger, 23 years old, left-footed, tendency to drift inside. Page eleven said: right back, 27 years old, right-footed, tendency to push high. Same name. Two datasets. Two different players.

It took me nearly four hours to trace it back. The data provider had merged two same-named players from different leagues into a single identifier. Nobody cross-checked. The report went straight into a meeting room, and nearly went straight into a transfer proposal.

Data errors exist everywhere. What is notable lies elsewhere: there is an entire industry that handles this exact kind of error with a rigorous process, and football has almost never adopted it.

One observation is not enough to call it a detection

In August 2026, a group of astronomers announced they had recorded auroral-type radio emission coming from Beta Pictoris b, a gas giant roughly 63 light-years from Earth. They used the MeerKAT radio telescope array in South Africa. The signal was so faint that, to separate it from background noise, they had to compare the recorded position against the planet's known orbit, then statistically exclude both the host star and a second companion planet in the same system.

The most important line in that announcement sat at the very end: the result was still awaiting independent peer review. No other group had confirmed it. Nobody had reproduced the measurement.

That standard is so strict that many outsiders find it absurd. A single observation, however clean, is not yet called a detection. To become a detection, someone else, with different equipment, elsewhere, must repeat it and reach a similar result. Until then, it is merely an open candidate.

Football runs the opposite way. A single observation — one match, one half, one passage of play — is instantly elevated into a conclusion, then into a headline, then into policy. A forward who scores twice in one evening is labelled a complete number nine. A team that loses after holding 70 percent possession is labelled finished. Nobody waits for peer review, because in football no peer review mechanism exists.

The gap between those two ways of handling information is what I want to examine. Not to praise one industry or criticise the other. Rather to point out that most errors in football analysis do not occur at the measurement stage. We measure very well, far better than fifteen years ago. The error occurs at the stage where we label what we just measured.

Three layers of labels and three layers of error

Over more than thirty years observing the football industry, with the last eleven spent working with professional data in Valencia, I have noticed that classification errors appear at three distinct layers, and the higher the layer, the harder the error is to spot.

The lowest layer is the event label. Who touched the ball, where, and how. It sounds simple, but this is where data providers diverge most. A pass that a defender grazes before it reaches a teammate: is it a completed pass, an incomplete pass, or a clearance? Three different providers can give three different answers about the same incident, and all three are confident they are right. When you merge data from two sources for one match, disagreement at this layer can account for a significant share of the recorded events.

The second layer is the role label. This is where money is lost. Winger and inside forward sound nearly identical, but in a player valuation model, those two labels sit in different career-age brackets, different reference wage levels, different transfer fee levels. A player mislabelled at this layer is not mispriced once. He is mispriced for the rest of his career, because every model downstream learns from the original label.

The third layer is the conclusion label. This is the layer the media lives on: character, weak mentality, winning DNA, small club. These labels cannot be measured, cannot be verified, and therefore can never be refuted. They are simply passed from one article to the next, one season to the next, until nobody remembers they ever had an origin.

Data does not lie, but it also does not tell stories on its own. Every number has passed through the hands of someone who labelled it. The problem is that the labeller rarely signs their name.

There is one very concrete example of the event-label layer that I still use when training young analysts in my department. The same shot from about 12 metres, slightly angled, might be given an expected-goal value of about 0.08 by one model and about 0.14 by another. Both are mathematically defensible. But when you add 0.08 for every shot a team takes across a season, and add 0.14 for every shot its opponents take, then place the two results side by side on a chart, you create a conclusion that is nearly double the real difference. Nobody lied. Two people simply used two definitions for the same word.

The role-label layer works the same way. When a centre-back playing in a back three is compared against the label of centre-back in a back four, his numbers look worse than reality, because his job is different in nature. He faces fewer one-on-one situations, but must cover more space, and those covering runs barely appear in standard statistical tables. A central midfielder at a weak club, where he touches the ball perhaps sixty percent as often as a counterpart at a strong club, will be undervalued by the model, because his sample is smaller rather than because he is worse.

This leads to a consequence the scouting world rarely states outright. The best players at mid-table clubs are usually priced below their true value, and most of that error comes from labels, not from skill. The biggest clubs exploit exactly that gap. They do not buy the best players according to indicator rankings. They buy the players those rankings have mislabelled.

Lessons from 31 hours of tape

In 2026, while tracking Levante UD in Valencia, I sat down with 31 hours of video from a single season. The initial goal was very narrow: find out where this team conceded from.

The standard statistical table gave me an answer that looked entirely sensible: the defensive line was misaligned, the goals conceded were evenly spread, there was no clear pattern. But when I coded each situation myself instead of trusting the pre-set labels, the result changed completely. Most goals conceded came from the left channel, and a subset of those came from corners exploited through one repeating running pattern. I drew 214 attacking diagrams, and the points dropped from that cluster of situations were enough to change the team's final league position.

What is striking is that the raw data was not wrong. Every single event was correct. The error lay in the label the system assigned by default. The system called it a goal conceded from a set piece, a label far too broad to generate action. To generate action, it had to be relabelled more narrowly: a goal conceded from a left-side corner, when the second line failed to drop in time, and when the full-back on that side was drawn inside.

Tactics are not a diagram; they are how a team responds to chaos. And to see the response, you must first be able to classify the chaos.

After I presented the findings, an editor told me he was not convinced, because the official statistics said otherwise. I suggested he do one simple thing: rewatch the next four matches and record every goal conceded using my new classification. Three of those four matches unfolded exactly as the model predicted. That was the first time I understood that the value of analysis is not in stating what is true, but in offering a different classification that can be tested by the next match.

1,029 passes and eight shots

In 2026, at the World Cup in Russia, I was invited to provide tactical commentary on the Spain versus Russia round-of-16 match. Spain completed 1,029 passes and held roughly 74 percent possession. They were eliminated on penalties, having registered only eight shots on target across 120 minutes. Russia's goalkeeper Igor Akinfeev was the man who stopped that shootout.

The 1,029-pass figure instantly became a headline. It was used to prove several contradictory things: that Spain were still playing their true identity, that Spain had run out of ideas, that possession football is useless. All three conclusions were drawn from the same number, and all three skipped the only question worth asking: where those 1,029 passes actually went.

I mapped out 47 of their build-up sequences. Most of the passes were lateral circulation in the area in front of the penalty box, where the Russian defensive block had dropped deep enough to make every sideways pass a safe pass. Around eighty percent of the passes fell into the category of circulation that created no breakthrough angle. The share of progressive passes into dangerous areas was so low that it explained the eight shots on target on its own.

The ball is only a variable; the way it moves is the message. The figure of 1,029 was not wrong. The possession label attached to it was the wrong part, because it merged two entirely different things — holding the ball to find space and holding the ball to avoid losing it — into one box.

When I presented this argument live on air to roughly two million viewers, the reaction was not gentle. Someone said outright that a woman does not understand tactics. I did not argue. I simply presented the data on the average position of each pass, the number of entries into the penalty area, and the number of transition situations after losing the ball. Days later, when other analysts redid the count, my numbers held. But what I remember most is not being right. What I remember most is how many people were willing to draw a major tactical conclusion from a number without ever checking how that number had been labelled.

When the stands emptied and the home-advantage label vanished

In 2026, when football returned after the lockdown, I reviewed 63 La Liga matches played without crowds and set them against 63 pre-pandemic matches. The goal was to test a label the whole industry treats as a law: home advantage.

The results diverged clearly. Successful pressing dropped by around 12 percent. Goals from fast counter-attacks rose by around 18 percent. The average height of the home team's defensive line fell by about 4 metres. These three indicators do not explain the whole story, but they point to something the old label had concealed: most of what we thought was tactical home advantage was in fact a psychological effect acting on people — the opposing players, and the referees too.

An empty stadium does not erase the match; it strips away the excuses. When forty thousand spectators disappeared, home teams no longer dared to push their defensive line as high. It turned out they pushed high not because it was the optimal tactical choice, but because the crowd behind them made sitting deep emotionally expensive.

Three weeks after I published the 12-page report, an assistant coach in La Liga cited it in an official press conference. That was the first time I saw an independent data analysis travel directly into the language of a coaching staff, rather than staying in the newspapers.

Four steps to self peer-review

From those three stories, I distilled a small procedure I still use before any conclusion leaves my computer.

The first step is to trace the label backwards. For every indicator I intend to use, I ask myself: who labelled this event, under which definition, and when did that definition change. Just knowing that two major providers define a clear chance differently is enough to stop you comparing their indicators as if they shared a unit of measurement.

The second step is to separate the sample. One match is not a sample. Five matches are not either. For systemic conclusions, I need at least a season, and I need a comparable control group. When testing home advantage, 63 matches against 63 matches was the minimum for me to say it out loud.

The third step is to hunt for counter-examples. Before asserting a model, I go looking for the case where it fails. If I cannot find a single counter-example, the likely explanation is that I have not looked hard enough, not that the model is perfect.

The fourth step is to let someone else redo it. This is the step football skips most often. In the analytics world, we rarely share our code, our indicator definitions and our raw datasets for outsiders to check. The result is that everyone builds their own model, believes in it, and nobody is in a position to refute anybody.

Good data does not answer questions; it teaches us to ask better ones.

Mislabelling in the transfer market

There is one consequence of this problem I have tracked for years and seen become clearer each season. During the annual league campaign, when a mid-table club outperforms expectations, the reward is rarely keeping the squad intact. It is being taken apart. Two or three key players are signed away by bigger clubs within a single transfer window. Success becomes the opening act of a dismantling.

What is notable is that this dismantling is usually driven by labels rather than by peer-reviewed data. A player who performs well inside the specific system of his parent club is labelled a complete forward, bought, then placed into a different system where his real qualities are no longer used. The following season he is given a new label: poor adaptor. Nobody goes back to check whether the first label was correct.

Another example sits in load management. For years I have heard a great deal about scientific rotation programmes and planned rest. But when you look at the actual schedule of a player at a big club, you see long trips wedged between important matches, usually for commercial friendlies. Load management exists, but it usually has to give way to that calendar, and when the player is injured, the label applied is fragile physique or lack of professionalism.

Both examples are errors at the third label layer. We explain outcomes through character, while the cause sits in structure. And because character labels cannot be verified, they outlive all the data.

A contrarian view: when classification becomes bias

There is a paradox I have not seen anyone in the industry state plainly. Over the past fifteen years, football has moved from too little data to too much. We solved the measurement problem, but created a new one: over-classification.

When every passage of play has a label, every player has a label, every team has a label, the label becomes a mould the data is forced to fit. A player who does not fit the mould gets pushed to the margins, not because he plays badly, but because the model has no box for him.

The real blind spot of modern football is not a shortage of data. It is the confidence that our classification system is neutral. In astronomy, people spend months proving that a signal does not come from another source. In football, we spend about thirty seconds deciding what category a player belongs to, and then spend years justifying that decision with data.

I am not proposing that football wait for peer review the way astronomy does. Nobody can wait several months to decide whether to sign a striker. But there is a shortened version of peer review that any analysis department can apply immediately: before using an indicator, write down its definition, its source and its labeller; and before concluding anything about a player, check whether he was mislabelled from the start.

This does not make analysis slower. It makes analysis refutable, and that is precisely the difference between an opinion and a conclusion.

What to check next matchday

This week, when I rewatch the La Liga fixtures, I will do one very small thing: for each team, I will write down the label the media is currently using for them — in crisis, flying high, playing counter-attacking football — and then compare that label against the three driest indicators: entries into the opponent's penalty area, entries allowed into their own penalty area, and the quality of their transition situations.

If the label and the data match, the label is trustworthy. If they diverge, the divergence is usually the label.

A match without spectators is still loud enough, if we know how to listen to every touch of the ball. And perhaps the thing most worth listening to is not the sound of the ball, but the sound of whoever just labelled it. If you are holding a dataset of your own, try asking yourself this week: how many of those labels did you write, and how many were simply inherited from someone else?