22 Data Points, Zero Football Entities: How a Labelling Error Distorts Football Metrics

On a Saturday evening I opened a file of 22 information points that had just dropped into the analysis queue. The domain label said one word: football. I read line by line, waiting for a club name, a match, a transfer fee, a manager. Nothing came. The first point described a 35-year-old actress who had just given birth to her second child at home. The third mentioned a blockbuster film franchise. The nineteenth referred to a community venture called Earthling. Across all 22 lines, the number of football entities I could count was zero.

A single mislabelled file, standing alone, harms nobody. The problem is that it does not stand alone. It sits inside a batch of several thousand items, flows through the same pipeline, is added to the same column, and that column answers the question every sports editor asks each morning: what is Vietnamese football talking about today, where is the heat, where has it gone cold.

I spent an entire evening on a birth announcement because it is evidence, not a story to tell.

The first xG table I ever wrote was by hand on a coach bus, back when nobody called it data. I sat in the third row, logging every shot by all 14 V.League clubs into a notebook, building my own coefficients for each shooting zone, then adding them up. Nobody checked me. Nobody made me prove a source. I had to set the rules myself, because a table without rules is just paper with ink on it.

Ten years later the industry runs on different rules. News is collected automatically, labelled automatically, classified automatically, and most of the time it works. Precisely because it works almost all the time, the failures slip through unnoticed. The Vietnamese sports data pipeline today has three layers. Collection reads the headline, matches keywords, assigns a domain label. Analysis checks the body, verifies entities. Editorial decides whether to publish. The first layer runs on machines. The second runs on people, and people have quotas.

My check on the file returned a clean result: no competition, no club, no player, no manager, no governing body appeared anywhere. What the 22 lines did contain was an actress, her husband, a newborn, a three-year-old first child, two former co-stars, a film franchise, a community venture and a social platform. All real, all verifiable, all irrelevant to football.

A domain label is only worth trusting when at least one verified entity stands behind it. That is the rule I apply to every file I receive, including files I collect myself. Without an entity, the label is nothing more than a filter's guess, and a filter's guess does not deserve a place in an aggregate table.

Source quality here is worth recording too. All 22 information points carry either a blank source field or a self-published source from the subject herself. No named journalist, no independent confirmation. On the four-tier scale I use, this is tier three: adequate for a lifestyle item, useless for any football claim. A transfer fee sourced at that tier does not enter my valuation model, even when it happens to match a rumour I am already tracking.

Then comes the most interesting part, and the easiest to skip: date consistency. The post is dated Sunday, 18 October. The second child was born on 28 September. The first child is described as three years old, and another information point places his birth in September 2026. Cross-check: if the first child is three, the announcement cannot predate September 2026. 18 October 2026 does fall on a Sunday. 18 October 2026 fell on a Saturday, and the first child would then have been two. The data is self-consistent under exactly one reading, and that reading is not the one the crawler defaults to.

Now the mechanism. The headline uses the word star attached to a capitalised proper noun. For an athlete-extraction filter, that is the classic false positive: star is polysemous, meaning both screen star and sporting star, and the filter only sees the second half. In this article the word describes a film role and does not describe a player at all. The filter cannot read that distinction because it was never built to read.

The identification failure is not in the filter. It is in the ordering. The system assigns the domain label first, then filters entities. Reverse the order and this file is stopped at the door: no football entity, no football label. One small rule, several hours of manual review saved every week.

I have met this exact failure in my own data, at larger scale. In 2026, with competitions suspended, I rebuilt ten years of V.League from 2026 to 2026. One finding forced me to rewrite the whole opening section: clubs that changed chairman mid-season saw win rates fall 23 percent across the next five matches. I published a five-part retrospective, pulling apart every handover of power. A club executive called me afterwards and said the table helped him delay sacking his head coach at the most sensitive possible moment.

But for that table to stand, I had to remove 41 records from the sample. Forty-one mislabelled rows, mostly because the headline carried a player's name while the body discussed something else. Had I kept them, the 23 percent would have become 19 or 27, depending on how I chose to rationalise it. An error rate of 2 to 5 percent is enough to distort every trend computed on a narrow slice, and enough that nobody notices. Large errors get seen. Small errors get cited.

The transfer window is peak season for this failure mode. Headlines carry player names, bodies discuss family news, a commercial deal, a film premiere. The filter sees the name, assigns the label, and the label travels straight into the rumour ranking. The credibility filter I hand readers starts with one simple step: separate the name from the context, then ask what the name is doing in that sentence.

The hardest part of this case is that it does not look like a failure at all. A birth announcement published on a football channel will generate reads. Those reads flow into the dashboard, and the dashboard goes green. Nobody audits a green box.

Viewers watch the move; I watch 22 numbers in motion — and wait patiently for them to tell a different story. This time the 22 numbers told a story about the ruler itself. For contrast, a real football file looks like this: in 2026, Phan Van Duc recorded 0.48 xG per match for SLNA, above the V.League average for foreign strikers, despite scoring only five goals. I wrote that he would become a national-team pillar within three years. In 2026 he scored the decisive goal at the AFF Cup. A year earlier, in Russia, Croatia pressed with a PPDA of 7.9 against Argentina, lower than the side famed for control. The world saw Croatia as underdogs; I saw them as a coefficient chain nobody had dared to mine.

Those two examples share one property: every conclusion was anchored to a verified entity and a sufficient sample. The birth announcement has neither.

The easiest temptation is to blame the machine. That temptation is too easy, and it is wrong. The filter did exactly what it was programmed to do: catch keywords. What was removed from the process was the human gate that once stood between collection and analysis, the gate where an editor needed ten seconds to catch it. That gate was removed for speed, and speed is the one thing in this industry nobody wants to reduce.

Another temptation is to read this failure as an analogy. High reads on a football channel do not prove a football audience on the other end. They prove one person watched that franchise fifteen years ago and is curious about the actress who played Ginny Weasley. Correlation is not causation, and in this trade the two get read as each other every day.

There is one point where I have to argue against myself. Tier-three sourcing is not bad in this case. The subject published on her own account, so the claim authenticates itself: no third party can distort information whose origin is the person who wrote it. Real risk is low. The issue is not a weak source; it is the right source for one field being used in another.

My model does not cry and does not celebrate, but after every match it owes me a lesson. This lesson belongs to the collection layer, not to tactics.

My proposal fits in one sentence of rule: a file may carry a football label only when at least one verified football entity exists — club, player, manager, competition or governing body. Without an entity, the file goes to the general archive and waits for a human. The cost is close to zero. The benefit compounds batch by batch.

The next cycle of Vietnamese football data will not be won by whoever has more news. Everyone has news. It will be won by whoever has the cleaner gate, because when every pipeline empties into the same place, the only thing left that creates a difference is who dares to discard.