Data Mislabeling: The Invisible Enemy of Professional Football Analytics

Last Tuesday morning, at my desk on Nguyen Thi Minh Khai Street, I opened a file marked "football intelligence — high priority." Inside were 25 data points. No team appeared. No player, coach, competition, or transfer was mentioned. The entire content described allergic rhinitis, the nasal mucosa's reaction to air conditioning, and sinusitis complications in young children.

I read it three times. The first time to check whether I had opened the wrong file. The second time to search for any tactical metaphor I might have missed. The third time out of pure curiosity: how could a professional classification system err this badly?

After three minutes of silence, I wrote the familiar line in my notebook: "Every number is a confession, if we are patient enough to listen." This time, the confession came not from a player or a coach. It came from the very data system I had trusted for 46 years.

Twenty-five data points. Not a single name. Not a single sporting event. Only nasal mucosa, sinuses, and air conditioning.

To properly assess the severity of this error, we must start with the numbers.

A mid-tier V.League club now collects between 800 and 1,200 data points per match. A top European side like Manchester City or Liverpool can generate over 10,000 points per match, including GPS coordinates for every player every 0.1 seconds, xG for each shot, pressing counts within 5 seconds of losing the ball, and pass completion into the final third.

No analyst reads every data point manually. Companies such as StatsBomb, Opta, and Wyscout, along with clubs' internal systems, use automated pipelines to assign labels. When a file enters the system, a classification algorithm decides which domain it belongs to: match data, transfer data, medical data, financial data, or media data.

The label determines the processing queue. A wrong label sends data to the wrong department, the wrong analytical model, the wrong decision-maker.

The labeling mechanism has three typical layers. The first relies on keywords: if a text contains phrases like "player," "coach," "injury," or "transfer," the algorithm may classify it as football. The second relies on entity frequency: if team or player names appear above a certain threshold, the label is reinforced. The third relies on machine-learning language patterns: a model trained on millions of sports articles predicts the probability that a text belongs to the football domain.

These three layers complement each other, but they also create three blind spots. The keyword layer is fooled by contextual overlap. The entity layer is neutralized when there are no entities to count. The language layer is corrupted when a medical text describes symptoms in language close to describing sporting injuries.

I witnessed a miniature version of this error in 2026, at Ho Chi Minh City FC. Our system tracking 12 movement metrics mistakenly merged recovery session data into the tactical session because both took place in the same time slot. The subsequent report showed a midfielder's pressing index up 40%, when in reality he had only jogged lightly around the pitch all session. The coaching staff nearly adjusted tactics for the Hanoi FC match based on that wrong number.

The difference between the 2026 error and last week's error lies in scale. A club-level error can be detected and fixed within 24 hours. An error in a global analytics pipeline can spread through hundreds of reports before anyone notices.

When a medical file leaks into a football analytics queue, the damage does not stop at wasted reading time. It spreads across three levels.

The first level is local contamination. The wrong file sits in the queue, takes the place of the right file, and slows processing. For a small club, this is a performance issue. For a data analytics corporation, it is a computational cost issue.

The second level is statistical contamination. If the wrong file is not discarded but fed into an aggregate model, it alters the data distribution. An xG prediction model or an injury-risk model trained on contaminated data will produce skewed results that no one detects, because the output looks perfectly reasonable.

The third level is decision contamination. This is the most dangerous. When wrong data leads to transfer, tactical, or medical decisions, the consequence is no longer a number on a screen. It becomes a wrongly signed contract, a wrongly purchased player, or a missed injury.

At the third level, cost is no longer measured in time or bandwidth. It is measured in money and careers.

Back to last week's file. The 25 data points included: the sneezing mechanism, the role of the nasal mucosa, the effects of air conditioning and sudden weather changes, allergic rhinitis symptoms, sinusitis complications, the risk of otitis media in young children, impacts on sleep quality and academic performance, and nasal hygiene measures.

Nothing related to football. So why was the label "football intelligence"?

The answer lies in three keywords: "injury," "treatment," and "performance." In the medical text, "injury" appears when describing mucosal damage. "Treatment" appears when recommending hygiene measures. "Performance" appears when discussing the impact of rhinitis on work and study.

The algorithm reads these three words. It does not read context. It does not know that "mucosal injury" is entirely different from "knee injury." It labels the entire document as football.

This is the classic error called "domain mismatch." The document belongs to the medical domain, but the label belongs to the football domain.

Notably, there was not a single football entity in the document. No team, no player, no coach, no competition, no transfer. Logically, the system should have refused to label a document without any football entity as football.

But the system did not. It only counted keywords.

This is precisely the point I want to emphasize: entity verification is a mandatory gate before accepting any label. If a file is labeled football but contains not a single team, player, or competition, that label must be automatically rejected.

Without this gate, every analytics pipeline is at risk of contamination.

The financial consequences of mislabeling are far larger than outsiders imagine.

The global transfer market handles roughly 15 to 20 billion USD per year, and most decisions are based on analytical reports. Each report is a chain of selected, classified, and aggregated data files. If one link in that chain is contaminated, player evaluation can skew in ways no one expects.

I know of a 2026 case, when a V.League club considered signing a foreign striker based on a high-index report. That report was built from data across two different leagues, but the system mistakenly merged one league's data into the other's cross-check. As a result, the player's goals per 90 minutes was pushed to 0.78 when the true figure was only 0.41.

The club signed him at a salary 30% above his true value. After 12 matches, he scored three goals. He was terminated mid-season.

The cost of that error, including wages, agent fees, and opportunity loss, was estimated at around 800 million VND. Not a huge figure by European standards, but for a V.League club, it was a quarter of a season's budget.

Every transfer decision begins with data. When the data is contaminated, everything downstream is contaminated too.

Domain mismatch is not exclusive to football. The financial industry has seen automated trading algorithms buy shares in one company because of news about another with a matching ticker. The medical field has had diagnostic systems mislabel patients because a disease code matched a drug code. The media industry has broadcast sports news within financial bulletins because the keyword "market" appears in both domains.

The common denominator is faith in automation without cross-verification.

In football, that faith is especially dangerous due to time pressure. A match lasts only 90 minutes. The report must be ready before the referee blows the whistle. If the system mislabels, no one has enough time to detect and correct it.

World Cup 2026 taught us: emotion is the hardest noise to filter from data. But there is another kind of noise harder still: noise from a wrong label. Emotional noise can be recognized by cross-checking metrics. Label noise sinks into the structure, seeps through every analytical layer, and only surfaces after the damage is done.

Back to the Ho Chi Minh City FC story of the 2026 season. The match against Hanoi FC on matchday 18 is a lesson I still tell in data workshops.

Young midfielder Nguyen Trong Huy ran only 8.2 km in 90 minutes, 15% below the team average. I proposed substituting him in the 60th minute. The coaching staff ignored it. The team lost 1-3.

Notably, the 8.2 km figure itself might have been wrong, because the system at the time had no cross-check gate between GPS and camera data. If the error had gone the other way, I would have proposed substituting a player who was performing well. After the match, I spent two weeks re-verifying all GPS data against match footage, and discovered three other discrepancies that had never been reported.

Since then, I have established a two-layer process: every metric, before entering a report, must pass cross-verification between at least two independent data sources. If only one source exists, the figure must be flagged "unverified."

This is a simple but costly lesson: an unverified number is more dangerous than no number at all.

I want to dedicate a section to the real costs when an organization accepts a wrong label without verification.

The first direct cost is wasted analysis time. An analyst reads the wrong file, tries to find meaning, writes notes, and sends an internal report. That time is unrecoverable.

The second cost is cross-contamination. If that analyst writes notes into the shared system, the contaminated notes will be read by colleagues. A single error becomes a chain.

The third cost is loss of internal trust. When one analyst discovers a colleague is working with wrong data, suspicion spreads. If one label can be wrong, every other label is suspect. This is the hardest cost to measure, but the longest-lasting.

The fourth cost is external reputational risk. If a report is published based on wrong data, the organization's credibility suffers. In the social-media age, a small error can become a topic of debate within hours.

Summed up, these four costs are typically 20 to 30 times larger than the cost of building an automated verification gate.

From last week's incident, I draw four methodological principles that any football data analyst should apply.

Principle one: always verify entities before trusting a label. A file labeled football must contain at least one team, player, or competition. Otherwise, the label is void.

Principle two: cross-verify against at least two independent sources. No backup source, no conclusion. This is a principle I have applied since 2026 and have never regretted.

Principle three: record data provenance. Every number in a report must be traceable to the original file, the collection time, and the collector. If it cannot be traced, that number does not exist.

Principle four: explicitly mark what is unknown. An honest report must clearly state information gaps, never fill them with speculation.

These four principles do not require advanced technology. They require only discipline.

There is a contrarian view I want to place on the table, knowing it will irritate many colleagues.

In 46 years of work, I have realized that most efforts to improve football data quality are heading in the wrong direction. Organizations invest millions of dollars in more complex algorithms, deeper machine-learning models, more precise sensors. But they invest almost nothing in the most basic step: checking whether a file belongs to the right domain at all.

The paradox lies here: the more complex the algorithm, the harder the error is to detect. A simple linear model with three variables yields results that are easy to understand, easy to check, easy to spot anomalies in. A neural network with millions of parameters can absorb contaminated data and still output results that look reasonable.

I do not oppose advanced technology. I oppose using advanced technology as camouflage for laziness in basic verification.

If the majority is right this time, would I dare abandon my conclusion? I have asked myself that question many times. But the evidence from World Cup 2026, Euro 2026, and the hundreds of data files I have checked all leans in one direction: the more complex the system, the more it needs verification at the lowest layer.

This is what I call the "paradox of sophistication": more sophisticated does not mean more accurate.

Last week's incident is not an isolated event. It is the manifestation of a systemic problem spreading across the global football analytics industry.

The signal for the next cycle is clear. Professional clubs will soon face the question: how do you prove a number is correct before placing it into a decision report? The organizations that answer this first will gain a competitive edge in the transfer market, where one wrong decision can cost millions of dollars.

Those that cannot answer it will keep signing contracts based on contaminated data, and keep being surprised when the players they buy fail to perform as the report predicted.

Turning 62 has not slowed me down; it has told me which data is worth waiting for. And in this case, the data worth waiting for is a verified number — not a number that was labeled in haste.

The question I leave readers with: does your club's system have an entity verification gate, or is it silently accepting every file tagged as football without asking a single question?