Mislabeled Data: The First Crack in Every Tactical Analysis

Over the past four rounds of K-League 1, I ran three different event-data providers side by side on the same fixture. Three xG values, differing by nearly half a goal. Three different counts of line-breaking passes. Three different definitions applied to a single aerial duel. All three providers followed their own rulebooks to the letter. Nobody broke the rules. Nobody made a mistake. But when I reopened the footage at quarter speed and stepped through it frame by frame, most of the divergence did not sit in the passage of play. It sat in the labeling layer, decided long before the passage existed in any system.

Every collapse begins with a crack on the tactical map that nobody bothers to look at. In football data, that crack sits at the lowest layer: a person in front of a screen, with less than a second to choose between "pass" and "clearance," between "shot on target" and "blocked shot." The entire analytical building above is constructed on that choice.

A single match in a top European league generates roughly three to four thousand logged events. Each event carries coordinates, a timestamp, the acting player, the receiving player, an outcome, and at least one classification label. In major leagues, one to three taggers usually work each match, in real time or post-production, under pressure to reconcile with a verification partner's log. That volume multiplies across hundreds of matches per season and dozens of competitions, producing one of the largest datasets the sports industry has ever operated.

The three largest providers today do not share a dictionary. Opta belongs to Stats Perform; StatsBomb was acquired by Hudl in 2026. Both systems coexist in the market while defining nearly every defensive concept differently. One provider's progressive pass is another's line-breaking pass. A pressure event may count only when the defender is within two metres; another system stretches it to five. A successful duel may include instances where a player merely stood nearby and the ball bounced loose. No definition is wrong. They are simply different, and that difference is erased the moment the number lands in a comparison table.

None of this is new. It becomes urgent when clubs, scouts and even league bodies begin signing, extending and sacking on the strength of indicators born from non-identical definitions. A centre-back rated highly in dataset A can sit below average in dataset B for the same season. Nobody is wrong. It is simply that nobody checked whether the label matched reality.

Working in Busan, I keep the habit of cross-checking every important metric through three layers: event data, slow-motion footage, and my own handwritten notes. That rule took shape after I once called a player by the wrong name on a live broadcast, and I have never again let a single source shape my judgement.

Take the aerial duel. For the same defender, the aerial success rate can differ by as much as fifteen percentage points between two providers. The cause lies in scope: provider A counts only duels where both players actively jump for the ball, while provider B counts every ball delivered into that defender's zone of operation. Centre-back Kim Min-jae is the case I have tracked closely since he moved from Napoli to Bayern Munich. The differing scouting reports on him across his first Bundesliga season did not disagree about ability; they disagreed about definitions. One report called him an aerial sweeper, another called him passive in the air. Both were produced from real data, and neither could support a decision unless the reader knew which dictionary stood behind it.

The same problem appears with attackers. Lee Kang-in was grouped as a winger at Mallorca, then grouped as an attacking midfielder at Paris Saint-Germain, even though his action set changed far less than his positional label. When a player changes label without changing behaviour, every position-adjusted comparison loses its value. Son Heung-min is the mirror case: he held the same position across several seasons while his actual role shifted with each manager, and no positional label recorded that shift.

The key point is this: most debates about players in modern football are not debates about ability, but debates about dictionaries. People believe they are comparing two footballers. In practice they are comparing two sets of definitions.

Data only recounts the past. The good tactician is the one who hears the echo of the future inside the numbers. But to hear that echo, the first step is verifying that the number speaks the same language you do.

The same problem surfaces one layer higher: the heat map. For years, the heat map has been used as the final proof of a player's role. It shows where a player was. It does not show what the player did, and still less what the manager asked him to do. A winger with a heat map glowing red along the touchline has not proven that he hugged the flank. It may only prove that he kept abandoning his assigned position to hunt the ball, while his real position was the gap between centre-back and full-back that no heat map ever draws.

Modern football is not won with the feet, but by reading space before an opponent can set foot in it. And space is misread when the map cannot distinguish presence from intent.

Football analytics spends enormous energy hunting for faults in the data layer, yet almost nobody audits their own labeling layer. We examine frame after frame to catch a provider's error, then accept, whole and unexamined, a transfer story built on a single source, usually an agent with a direct interest. Agents do not lie in the ordinary sense. They relabel the truth. An enquiry becomes a negotiation, a compliment becomes a desire to move, a release clause becomes a blockbuster about to explode. The largest hidden cost of the transfer market is not the fee. It is the noise generated before a single euro changes hands.

The same mechanism explains why a back three gets labelled modern. In most cases I have tracked, the back three does not appear because of tactical progress, but because the back four has already been punctured and the manager needs a new label to protect his reputation.

The transfer window is a chess game in which the majority look at the pieces, while the quietest person looks at the whole board. Misclassification in the data layer and misclassification in the media layer are the same disease. They always appear together, and they are always ignored together, because both are disguised by the same thing: the appearance of precision.

I do not believe in miracles, but I believe in a squad the whole world has rushed to cross out. The problem is that to believe in it, I have to relabel it myself from scratch, rather than borrow someone else's label.

At the quantitative layer, I always build three conditional scenarios for any assessment. Scenario A: providers move toward a shared dictionary, at the cost of losing role-specific actions. Scenario B: the status quo holds, and every cross-provider comparison carries an unrecorded systemic error. Scenario C: video verification comes first, with data used to widen the sample rather than to conclude, at three to four times the time cost per report.

Among those three, I choose C for major decisions and B for daily observation, because not every match is worth paying three times the time.

The worry is not that data is wrong. The worry is that the more formally precise data looks, the less anyone checks it. A roughly rounded number makes a reader suspicious. A number with two decimal places makes a reader believe instantly. That illusion of precision spreads into journalism too, where a headline declaring that matters are now clear is often just repeating a source's claim rather than verifying anything.

After every round of fixtures, I still do the same old thing: pick one metric, reopen three providers, reopen the footage, and relabel it myself from scratch. Not because I distrust everything. Only because I once stumbled on live television, and I know that stumble did not come from the passage of play in front of me — it came from a labeling step completed long before, when I trusted my own memory too much. The season is long, and every round opens one more dictionary that still needs checking.