Mislabeling: The Silent Trap Distorting Football Data

Mislabeling: The Silent Trap Distorting Football Data

In late September, a file landed in my inbox with exactly one classification tag: football. I opened it, read it from the first line to the last, then read it again because I thought I had opened the wrong file. Inside were twenty-seven information points. Not one mentioned a club, a player, a coach, a competition, a transfer, or a passage of play.

The file discussed an international delegation being slimmed down under a national austerity policy, the speaking schedule of heads of state at a multilateral forum, regional tensions, and a list of speakers with time slots. The tag said "football." The content contained not a single word about football.

My first reaction was not irritation. It was a sense of familiarity that was uncomfortable in its intensity. I have seen this kind of error many times, differing only in scale. In an internal data file, it is a domain-classification error. In a transfer report, it is a name attached to the wrong club. In a post-match statistics table, it is a metric measured one way but cited as though measured another. All three share a single root: mislabeling.

In modern football, mislabeling is the most dangerous form of data contamination. It does not corrupt the number. It corrupts trust in the number. And trust is the only thing that makes a metric meaningful.

"Data does not lie, but the people who collect it do." I wrote that sentence years ago and have not yet found a reason to revise it.

When Vietnamese Football Entered the Data Era

Until the mid-2010s, football data in Vietnam was essentially synonymous with the scoreline. Goals, cards, possession rounded to the nearest whole percentage, and a few lines of shot statistics. Everything else lived in the eye of the viewer.

The turning point came in two waves. The first was achievement. When Vietnam's national team reached the quarter-finals of the 2026 Asian Cup and lost 0-1 to Japan in Dubai on 24 January 2026, the domestic sports media market realised it needed a new language for talking about football, one more detailed, one comparable with the major football nations. The second wave was infrastructure. International data platforms such as Opta, Wyscout, and StatsBomb expanded their coverage into Southeast Asia. V.League began appearing in structured data packages. Tracking cameras, though not yet synchronised across the whole league, gradually entered a number of stadiums.

The result was a paradox. Vietnam had football data faster than it could produce people able to read football data.

I call that gap the middle layer. It sits between the raw data source and the end consumer, the audience. The middle layer consists of extractors, metric translators, editors, labelers, headline writers. In many major football markets this layer is thick enough to constitute a profession: data journalists, tactical analysts, research leads. In Vietnam the layer is thin, often worn as a second hat, often untrained in statistics, and almost never publicly accountable when it errs.

When the middle layer is thin, three categories of error appear with alarming frequency.

The first is semantic error around metrics. PPDA, the number of opponent passes allowed per defensive action, is a measure of pressing intensity. But a low PPDA does not automatically mean good pressing. A team defending in a low block, ceding the ball entirely and waiting to counter, can post a very low PPDA in certain matches, simply because the opponent passes a lot without being challenged. Without reading it alongside ball-recovery locations and block height, PPDA becomes decoration.

The second is error of unit and sample. xG is calculated differently by each provider. A shot from the same location can yield 0.08 in one model and 0.12 in another, depending on whether the model accounts for goalkeeper position, the number of players between ball and goal, or the type of pass leading to the shot. Citing xG without naming the provider is an act of mislabeling. You are assigning to the metric a precision it does not own.

The third, and most serious, is domain error. An article, a data file, a summary tagged entirely wrongly. This is precisely the September file in my inbox.

These three categories are not equal. Semantic error weakens analysis. Unit error falsifies analysis. Domain error collapses the entire system downstream, because everything built on it is meaningless. A prediction model trained on mislabeled data will learn exactly what we taught it: confusion.

Based on my experience covering matches over nearly three decades, I can say that Vietnamese fans today are far more sophisticated than the level of service they receive. They are used to xG, to touch maps, to heat charts. But they have never been given the tools to check where those things come from.

Anatomy of a Mislabel: Three Layers

A data file travels from raw text to the reader's hands through three layers. The first is entity extraction. The second is domain classification. The third is source attribution. An error at any one layer is enough to break the whole chain, but only when all three fail together do we get a "football" file containing no football.

Layer One: Entity Extraction

If a system reads text and identifies the entities mentioned, it must return clubs, players, competitions, coaches, stadiums, and governing bodies. When the entity list contains only heads of state, ministers, spokespeople, and the names of international organisations, that is a red flag that cannot be ignored. A sound system must refuse the football tag, or at minimum drop its confidence to near zero and route the file to a manual review queue.

In the September file, the entity list was government officials and heads of state. Not one club. Not one competition. Not one player. Yet the tag remained football. That means the entity check was skipped, disabled, or never installed in the first place.

This is the kind of error I fear most, because it is invisible. Nobody sees it on the dashboard. Nobody receives an error notification. The file still travels through the pipeline, still files into the correct drawer, and only when someone actually opens it and reads does the truth emerge.

Layer Two: Domain Classification

Domain classification assigns a subject to the whole document. It relies on a language model, a keyword set, or a combination of both. It is also the layer most easily fooled.

The phrase "austerity measures" can lead a weak model to associate it with a club's transfer-budget cuts. "Delegation" can be wrongly mapped to a squad. "Head of state" can be read as head coach. "Speaker" can be read as commentator. These faulty mappings do not come from the data source. They come from the model.

What is worrying is that a misclassifying model still produces confident-looking output. The probability may be 0.94. That number does not reflect truth. It reflects the model's confidence in a wrong conclusion. This is why I never accept a label merely because it comes with a high score.

Layer Three: Source Attribution

In the September file, most information points carried the line "Source: None." Only a few were attributed to "sources at the delegation." In data analysis, an unattributed information point is worth half, or nothing, depending on how strict the process is.

An unverified number is more dangerous than a wrong opinion. A wrong opinion can be argued with and refuted. A source-less metric cannot be refuted, because nobody knows what to refute. It exists in a grey zone, gets cited again, and gradually becomes default truth.

These three layers do not operate independently. They form a chain. When extraction fails, classification loses its foundation. When classification fails, attribution becomes meaningless. The end result is a fully mislabeled file, with a fully legitimate appearance.

Four Cases, One Lesson

I do not tell the stories below to praise myself. I tell them because each is a mislabel that was corrected, and each correction left a specific lesson.

2026: The Shanghai Derby and Fifty-Four Pressing Actions

In July 2026, aged thirty-five, I wrote an analysis of the Shanghai derby between Shanghai Shenhua and Shanghai SIPG, which finished 1-3. I argued that SIPG won not through luck but through fifty-four pressing actions executed in the final third of the opponent's half. I counted each one, cross-checked against the video record, and stated my counting method at the end of the piece.

A former male star on national television mocked: "What does a woman know about football?" My article drew negative comments for a week. I stayed silent. When Opta published tracking data confirming the figure of fifty-four, a number of colleagues apologised to me privately.

The lesson was not that I was right. The lesson was that people had placed a label on the article before reading it. The label "woman" determined how they read the analysis. They did not check the counting method. They did not open the video. They only read the label.

The Shanghai derby forged in me a healthy instinct to distrust data.

2026: Croatia and the Geometry of Pressing

In 2026, aged thirty-six, on the strength of the previous year's credibility, I was invited onto a live analysis panel for the World Cup. Before the semi-final between Croatia and England at Luzhniki on 11 July 2026, I wrote a prediction that Croatia would win. My reasoning did not rest on inspiration. It rested on structure.

Three days earlier, in Sochi, Croatia had beaten Russia in the quarter-final. In that match, Luka Modrić touched the ball one hundred and twenty-eight times. That figure does not merely say Modrić was heavily involved. It says Croatia controlled the central corridor through rotating triangles among Modrić, Ivan Rakitić, and Ivan Perišić. Each triangle opened a passing lane, each pass erased a gap in the opposition shape, and each erased gap created a new one.

The media leaned firmly toward England. They labeled a young team "title contenders" and an older team "past it." Croatia won 2-1. Several major newspapers cited my name.

Croatia 2026 taught me: pressing is geometry, not a sprint. The geometry of pressing lives not on the screen, but between the running lines.

2026: Empty Stadiums and the Limits of Tactics

When the Bundesliga returned after lockdown, I was thirty-eight and barred from the stadium. I analysed Borussia Dortmund's match at an empty Signal Iduna Park. The data showed the home side winning only 58 percent of duels, a significant drop from 76 percent with crowds the previous season.

I wrote a piece titled "The Silent City: Is Atmosphere a Player?" The central argument: crowd pressure had been masking part of Dortmund's pressing weakness. When the crowd vanished, the weakness surfaced. The label "home advantage" we had used for decades turned out to be a composite label, blending tactics, psychology, referee habit, and atmosphere.

The empty stadium of 2026 showed me the limits of tactics. 2026 made me realise: football is emotion before it is data.

The piece was widely shared in sports-science circles. But more importantly, it forced me to add one more variable to every analytical framework thereafter: crowd context.

Vietnam: VAR, xG, and the "Source: None" Habit

When VAR was introduced to V.League 1 from the 2026-2026 season, arguments over referee decisions multiplied exponentially. What was notable was that most arguments did not revolve around the decision, but around the data used to reach it.

How many frames per second define an offside situation? Which camera angle? Who selects the freeze frame? What is the margin of error? These questions are almost never answered in news reports. Fans receive only the conclusion, accompanied by a straight line that looks deeply convincing.

That is mislabeling in its most subtle form. Nobody lies. But a metric whose method has not been published is presented as absolute truth.

The same happens with xG. Some domestic data tables for V.League publish xG values without naming the model, the sample size, or the provider. When those tables are cited again in commentary on Nguyễn Quang Hải, Nguyễn Tiến Linh, or Đỗ Hùng Dũng, readers do not know they are comparing numbers produced by two different models.

For attacking players operating high up the pitch, model deviation may be small. For midfielders and defenders, deviation can be large enough to reverse rankings. This is why I always insist on seeing a metric's definition before trusting any ranking table.

The Blind Spot: More Data Cannot Fix a Wrong Label

The industry reflex upon discovering a data error is to buy more data. More sources. More models. More dashboards. This reflex is wrong, and it is expensive.

If the labeling layer is broken, more data only spreads contamination faster. An error at the classification layer multiplies when it passes through aggregation steps. Each time data is merged, the error is diluted but simultaneously replicated.

Composite indices are the most dangerous place. When you merge many sources into a single index, one mislabeled source skews the entire index without leaving a trace. Player ranking tables by action metrics are the classic example. If one match is tagged with the wrong competition, the player's minutes in that match are still added to the total, but placed against the wrong opponent context. The result is a ranking that is both wrong and appears objective.

The betting market is the end of this chain. Bookmakers read the same data source, apply the same label, and price on the same logic. When a mislabeled file enters that pipeline, it does not merely produce a wrong article. It produces a wrong price.

The deeper blind spot sits with people, not technology. We tend to trust the label over the content. When a metric is tagged "football," our brain stops checking. This is a widely documented cognitive bias, and it is stronger than any algorithm.

The fix does not lie in buying more data, but in installing a domain-confidence gate. If the entity list contains no club, player, or competition, the system must auto-reject. And at the human layer, the editor must be empowered to say one simple sentence: "This file does not belong to football, I will not process it."

In Vietnam, granting that authority matters more than buying a new data package. A middle layer that knows how to refuse is worth more than one that knows how to publish fast.

Three Rounds of Verification

I do not predict from data alone; I predict from data that has passed three rounds of verification.

The first round is source verification. Who measured? With what device? Over how many matches? Does the measurer have an interest in the metric leaning one way?

The second round is definition verification. How is this metric calculated? What is the margin of error? What question was it designed to answer, and does that question match the one I am asking?

The third round is contradiction verification. Does this metric conflict with any independent data source? If so, which side has stronger evidence?

Only when a metric passes all three rounds do I let it into an article. This is why I write slowly. Every piece carries a data-source note at the end, as a way of protecting myself and building long-term trust.

For Vietnamese football, these three rounds do not require expensive infrastructure. They require time and habit. One simple question in the newsroom, "where does this data come from?", can prevent more bad articles than any analytics software.

What I want to see next season is not more xG, more heat maps, or more ranking tables. I want to see one note at the end of each article, stating where the metric came from, how it was measured, across how many matches. Such a line does not make an article less compelling. It makes it credible.

And if some file ever again arrives labeled "football" while containing not a word about football, I hope the person who receives it will do what I did: open it, read it through, read it again, and refuse it.