The Crack Starts at Data Entry: When Vietnamese Football Data Gets Mislabeled

In November 2026, after the coaching-staff meeting, I opened a file containing fifteen information points that the system had tagged "football tactics" and routed into my professional queue. I read the way I always read: first point, middle point, last point. No club. No player. No competition. Not a single passing metric, not one corner kick, not one match minute.

The second point mentioned two transnational religious organisations. The sixth listed a series of conflict geographies. The twelfth spoke of negotiation, justice and international law. The fifteenth promised that positive results would emerge soon.

The system label still read: football.

The match rate between label and content was zero. Not nearly zero. Zero, across all fifteen points: no football entity, no tactical content, no financial transaction, no rule of the game.

In the dressing room, I do not listen to voices; I read the position of the boots. On the pitch, I do not read the scoreline first; I read the data-entry line first. A wrong entry line does not concede a goal immediately. It sits there, waits three months, and walks out as a scouting report.

Four layers, one of them dead

Vietnamese football ingests data through three channels. The first is the international provider, where every event is coded twice by two independent people and then cross-checked. The second is a club's internal unit, usually one or two people, doing it part-time, cutting the camera angle and typing event codes at the same time. The third is independent analysis groups, sometimes one person alone with two monitors and a coffee gone cold.

The three channels differ in money. They are identical in one respect: the labelling layer.

A data pipeline has four layers. Capture — cameras, sensors, the match sheet. Coding — turning images into discrete events. Labelling — assigning those events to a drawer: defensive, transition, set piece, wide attack. Modelling — using the drawer to compute and recommend.

Vietnamese football invests in capture. Occasionally it invests in modelling, usually by buying a software package, installing it, and leaving it there. It almost never invests in labelling.

And labelling is the only layer where a small error travels through all three remaining layers without being stopped anywhere. A capture error is obvious because the picture is blurry. A modelling error shows up slowly, because the predictions drift. A labelling error shows up for nobody, because it does not produce an error. It produces a new fact that sounds entirely reasonable.

1,247 corners and a question nobody asked

In 2026, when the stadiums were shut, I sat down and coded all 1,247 corner situations of the 2026 V.League season. The season stood still, but the corners kept rolling through the spreadsheet.

The first result: one goal for every thirty-seven corners. The comparable Southeast Asian average I could cross-reference for the same period was one goal for every twenty-five corners. A gap of roughly thirty-two percent.

I cross-referenced the near-post ball position against central defenders' body placement and found a systemic hole: V.League defensive lines habitually pushed up before the ball was struck, leaving space behind the first line for a runner arriving from the second line. That is a real fracture. I wrote the report and sent it to a club in Nha Trang without waiting for anyone to assign me the work.

But before sending it, I did something I consider more important than the conclusion itself: I audited my own labelling layer.

From those 1,247 situations I pulled a random hundred and re-watched the footage. Seven had been coded into the wrong category. Four I had logged as "attacking corner" were in fact clearances off the flank that had been recycled. Three I had logged as "short corner" were in fact long balls into the box punched out by the goalkeeper. An error rate of seven percent.

If that seven percent were evenly distributed, the conversion rate would shift from one in thirty-seven to roughly one in thirty-four. The distance to one in twenty-five still stands, meaning the systemic set-piece defensive hole is real and not a product of entry error. But had I not pulled a hundred situations for review, I would have sent out a number whose margin of error I did not know.

Corner numbers do not lie, but they stay silent until you ask them the right way. And the right question is not "what is the conversion rate" but "who labelled each of those situations, and did that person watch it a second time".

The match against Russia and the sixtieth minute

In 2026 I coded all sixty-four matches of the World Cup in Russia using a spreadsheet I built myself. In the France – Croatia final I found that after the sixtieth minute, Luka Modric's high-intensity distance metric dropped twelve percent, while the French attack kept switching its point of attack into exactly the zone Modric had to cover. Croatia shifted into a deep-lying midfield block, but the dropping midfielders arrived late, leaving vast space through the middle.

My piece was later widely shared within domestic coaching circles.

The part I rarely tell: before publishing the twelve percent figure, I re-watched the second half three times. The first pass gave me fifteen percent. The second gave ten. The third gave twelve. The difference came from whether I labelled short bursts and sharp changes of direction as "high-intensity running" too. Only once I fixed the definition and wrote it down did the number stabilise.

A halo does not go out overnight; it begins to crack in the sixtieth minute of the match against Russia. But a crack only means something if the person measuring it uses the same ruler across all three measurements. If you see nothing at the sixtieth minute, rewind to the fifty-ninth. If you rewind while your labelling standard shifts between passes, you will produce three different numbers and tend to pick the prettiest. That is the moment analysis becomes decoration.

And the match in Nha Trang

In 2026, while I was on the coaching staff of Sanna Khanh Hoa BVN, we faced SHB Da Nang at home. Half-time: 0-1 down. I watched the first-half footage twice and counted fourteen opposition build-ups, every one directed into the gap between the right-back and the right-sided centre-back.

I redrew the shape and proposed to the head coach that we switch from 4-4-2 to 3-5-2 during the interval. In the second half, dangerous entries into that same gap fell to two. We won 3-1.

I did not celebrate. I added one more defensive variant for the next match.

Looking back, the remarkable part was not the shape change. It was that I counted fourteen build-ups. Had a different assistant done the counting, the number could have been eleven, or seventeen. We had never sat down to agree what "a dangerous build-up" was, or what "the gap between the right-back and the right-sided centre-back" meant in metres or in body-lengths. The whole staff read the same data, but each read it with a private dictionary.

A pass two metres off target is not a technical error; it is a fracture in the whole cognitive system. But before you conclude anything about the pass, you must be certain the person logging the landing coordinates did not mistype an axis.

Three kinds of labelling error

From those three cases I extract three error types that every V.League analysis unit commits, differing only in whether it notices.

Type one: the label is right but the definition is wrong. A cross may be tagged "chance" while the internal standard for calling something a chance was never written down. If the standard is "ball into the box within five metres", the conversion rate differs sharply from a standard of "ball reaching an attacking player's position". Same match, same coder, two different results.

Type two: the label is wrong but the definition is right. A counter-attack logged as a set-piece attack because the coder was fast-forwarding. Easy to catch if a second person reviews; easy to miss if only one person works.

Type three: label wrong and definition wrong at once. This is the dangerous one, because there is nothing to cross-check against. The coder believes he is right, the reader believes the data is sound, and nobody in the chain has a reason to reopen the footage.

The fifteen-point file I opened last month is an extreme variant of type three: machine-generated labels, machine-inferred definitions, no human confirmation anywhere in the chain. One hundred percent wrong. That kind of error is loud, so it is caught immediately.

The paradox is this: the most dangerous error in Vietnamese football is not the one hundred percent error. It is the five-to-ten percent error — precisely the margin we call "fine".

The labelling layer in scouting and valuation

Now pull this into the transfer market.

In the summer of 2026, Joao Felix moved from Benfica to Atletico Madrid for one hundred and twenty-six million euros at nineteen years old, with fewer than fifty Primeira Liga appearances. That is a naked gamble, and I do not use the phrase in a moral sense. I use it in a statistical one.

A valuation of that size is not built on a scout's gut feeling. It is built on a data chain: progressive actions per ninety, successful line-breaking passes, escapes under pressure, receptions between the lines. Every one of those metrics passes through a human labelling layer, and that layer has never been publicly audited in any market.

In other words: a club paid one hundred and twenty-six million euros for a player based on numbers that nobody in the meeting room could trace to a specific labeller, a specific definition, and a specific sample-review rate.

In the V.League it is cruder still. A young player with eight goals in half a second-division season will be valued clearly above a player with seven goals in the same position. The difference between eight and seven sits comfortably inside the margin of error of a coder who was tired in the eighty-eighth minute of his third match of the week. Nobody writes "eighty-eighth minute, third match, coder tired" into a scouting report.

The young-player price bubble will not burst because the market runs out of money. It will burst because the labelling layer cannot carry the load the market has placed on it. You can build a valuation model with as many tiers as you like, but if the input line logs the wrong event type, the model is merely optimising an error.

Transmission down the infrastructure

A labelling error does not stay in one analyst's file. It transmits.

Down to academies: a sixteen-year-old centre-back is trained against an "aerial duels won" metric while that metric at youth level is recorded by eye with no second reviewer. The boy is taught to jump more, when his real problem is the angle of his screening run before the striker.

Down to the intermediary network: an agent uses a data sheet to persuade a smaller club. The sheet is not wrong about the number. It is wrong about the definition behind the number, and the smaller club has no capacity to audit the definition.

Down to broadcasters and sponsors: viewership, engagement and audience-demographic figures price the sponsorship package. At this layer, a few percentage points of demographic error can mean billions of dong of contract variance. Nobody in the signing room reopens how that audience segment was labelled.

The result is a system in which every layer trusts the layer below to have done it right. The modelling layer trusts the labelling layer. The labelling layer trusts the coding layer. The coding layer trusts the capture layer. The capture layer trusts nothing; it merely records.

The vulnerability map must start at the first line

My job is to defend by drawing the vulnerability map in advance. Without it, I get surprised in matches. But a vulnerability map does not only mean the gap between the right-back and the right-sided centre-back. It also means the list of places where my own data can crack.

That list has four lines. First: which events were coded by one person with no second reviewer. Second: which definitions were never written down and live only in the coder's head. Third: which samples were never pulled at random for review. Fourth: which labels were machine-generated with no human confirmation.

The fifteen-point file labelled "football" that I opened last month falls on the fourth line. It was an automated error, not an editorial one. But in Vietnamese football, most data errors fall on the first three lines, because we have no cross-check layer.

Here is a nearby reference point. On 1 February 2026, the Vietnam national team beat the China national team 3-1 at My Dinh Stadium in the third round of Asian qualifying for the 2026 World Cup. After that match, many analytical sheets circulated in domestic coaching groups. I read about twenty. How many specified their definition of a "clear chance": two. How many specified who coded the data and whether it was cross-checked: none.

That is the problem. We do not lack data. We lack a data-confirmation layer.

The blind spot is not the wrong label

Here I have to state a paradox of the domestic football-analysis trade plainly.

When a dataset is completely mislabelled, everyone's first reaction is to fix the label. Fixing a label is easy. Building an input gate is harder. Neither is the blind spot.

The blind spot is the habit of filling every empty cell in the analytical table. A table has nine dimensions, each with cells, and every cell must contain words. When there is no data, people write inference. When there is no inference, people write feeling. When there is no feeling, people write a sentence that sounds highly reasonable.

With that fifteen-point file, had I been a cell-filler, I could easily have produced eight paragraphs of tactical analysis that read perfectly plausibly from an article about interfaith dialogue. I could have written about "coalition structures", "consensus-building strategy", "pressure from below". Every sentence correct rhetorically. Not one correct in football terms.

Real analytical discipline is not producing a conclusion. Real discipline is writing exactly four words: insufficient information. And accepting that out of nine cells, six will be empty, and the reader will see those six empty cells.

People shine a light on the winner; I shine a light on where he stumbled. But my light only works where there is ground to illuminate. Where there is no ground, I leave the darkness intact and do not build extra lampposts.

Closing

Next season, ask one question before reading any data report: who labelled this line, did that person review it a second time, and what percentage of the sample was pulled for verification?

If nobody can answer, the number in the report is not data. It is a promise.

And a promise cannot be recorded in a spreadsheet.