A Monterrey News Report Inside a Football Data Pipeline: The Labelling Error and What It Means for Vietnamese Football Data

On Morones Prieto Avenue in Monterrey, Nuevo León, a pickup truck driven by a 19-year-old lost control and ploughed into the scene of a landslide. Two people working there, reporter Alma De la Rosa and cameraman Cristo Morales, were struck directly. Morales suffered a broken leg and multiple contusions; both were taken to hospital. Initial reports said the driver lost control after being cut off by another vehicle, and Monterrey authorities had not determined definitive legal responsibility at the time of publication. Earlier, a motorcyclist had also been injured at the same landslide point.

That is the entire content of the event. No club. No player. No match, no contract, no table.

And yet the item sat in my data feed under a “football” label.

I sat with that label for a while. My daily work is reading youth-player data, building indices, and trying to imagine how far each of them can go. A traffic-accident report from northern Mexico should have been stopped at the door. It got through. And the way it got through taught me more about today’s football data systems than any scouting report I read this month.

Two events in one data stream

Field reporting in Mexico has long been a life-threatening trade. Annual press-freedom trackers place Mexico among the most dangerous countries in the world for journalists, and most victims are local reporters covering crime, corruption and natural disasters. The Monterrey case belongs to a different, quieter but no less lethal category of risk: occupational hazard at an active incident scene, where the danger is uncontrolled and nobody can guarantee that the next vehicle will stop.

The telling detail is that a motorcyclist had already been injured at the same landslide point before the pickup arrived. That suggests the area had been an open hazard zone for hours, and the two journalists were hurt not through carelessness but because they were doing the hardest part of the job: being present where the event is happening.

The report itself is tightly written. Causation is hedged as “initial reports,” legal responsibility is explicitly stated as undetermined, medical information is attributed to the receiving facility. For an event whose causal chain is still open, that is the correct way to write: describe what is known, fence off what is not.

For me, the problem lies in a different layer, one nobody sees when reading the story.

I run a small data system supporting youth-player tracking and transfer-market work. Every day it pulls thousands of items from feeds, aggregators and multilingual keyword alerts. Before any item enters analysis, it must pass a domain-labelling step: football, economy, politics, society.

In most systems today that step works simply. First, the system looks for known entities: club names, player names, competition names. If it finds one, the label is assigned immediately. If it finds no entity at all, it falls back to a backup layer: keywords and place names.

And “Monterrey” is one of the strongest football tokens in any Spanish-language corpus. The city has CF Monterrey; Tigres UANL sit in neighbouring San Nicolás de los Garza; the derby between them is one of Liga MX’s biggest fixtures. Nuevo León is one of the most football-dense states in Mexico. The place name fired the backup layer. The semantic check — the step that should have asked whether this article is about football at all — never ran.

A landslide, two injured journalists, a 19-year-old driver who lost control, and one data row labelled football.

Anatomy of a labelling error

This kind of failure is predictable, not freak. It is the output of a design choice made for good reasons: recall over precision. In a data system, missing an important item is a loud error — measurable, and spotted immediately by users. Misaccepting an irrelevant item is a silent error. It sits in the table, occupies one row, and nobody re-checks it.

That asymmetry governs the system’s entire behaviour. When the cost of a miss is judged higher than the cost of a false accept, the system is tuned to accept more, and every such tuning pushes more rubbish into the clean store. The rubbish makes no noise. It simply distorts the counts.

At scale, one bad row in fifty thousand is close to harmless. At the scale I work at, it is a disaster. A dataset of fourteen young midfielders with one bad row carries a seven per cent error rate. That is why I hand-check the final list, and why I know exactly what it feels like to find a row that does not belong where it sits.

Uruguayans do not build walls. They build manifestos about space. A classification system is the same. The category “football” is not a box that holds articles; it is a statement about how information space is divided. The moment that category can contain a traffic accident, it has stopped describing football and started describing the geography of football. The shift is silent, and its consequences surface only when somebody sits down and counts.

I knew that feeling long before any automated system. Based on my experience watching matches on the ground, in 2026, aged seventeen, I hand-coded twenty-three matches of U19 Hà Nội and PVF at the national U19 finals. My spreadsheet held more than 1,400 data points: distance covered, pass-completion rate, receiving positions. The headline finding was that U19 Hà Nội generated only fourteen per cent of their shots from the central corridor, leaning almost entirely on crosses. The twelve-page write-up was shared among youth coaches in Hà Nội.

But I remember something else. Every cell in that spreadsheet was a human judgement. When I cross-checked, I found cells assigned wrongly: a duel credited to one player when another won the ball, a pass logged as incomplete when it reached a teammate before being cleared. The errors were small. But they sat in exactly the layer that models would later use to compute.

Under the raw data, I found the first brick of a generation. That brick is only trustworthy if I admit my hand can place it crooked.

The dead-data layer

In Vietnamese football, labelling errors happen daily, in silence, to our own players.

A U19 player who operates as a left-sided midfielder is filed as a full-back. A striker in the second tier has his minutes under-recorded because the match was not fully tracked. A goalkeeper who has sat on the bench for three seasons is sorted into the “insufficient data” bucket and disappears from every scouting list. Nobody decides to discard those players. They are discarded by a bad data row, and the bad row does not incriminate itself.

I call it the dead-data layer: the cohort buried by conventional metrics, not because they are poor, but because the way they were recorded was wrong from the start. That layer is thick in smaller leagues, where there is no analytics staff, no multi-angle camera, nobody typing up every phase of play.

In 2026 and 2026, when the pandemic kept me away from stadiums, I analysed 186 matches played without spectators in the Bundesliga and the V-League. The Bundesliga home-win rate fell from 44.8 per cent to 33.2 per cent. In the V-League, away teams’ expected goals per match rose 26 per cent. I spent an extra two weeks finishing a five-variable home-advantage erosion index, then published it as a series of five pieces.

Home used to be a fortress. The pandemic taught us that a fortress is only a variable.

And “Monterrey means football” is only a variable too — one somebody set long ago, which held for years and was never re-tested until a landslide proved it no longer holds in every case.

When clean labels produce signal

The opposite case is one where the labelling layer was kept clean.

At the 2026 World Cup in Qatar, over 45 days, I built a scoring system for fourteen young midfielders across twelve criteria, from pressing capacity to line-breaking pass rate. Every row was hand-verified, because the sample was too small to tolerate error. Enzo Fernández stood out with a 91.3 per cent pass-completion rate across five matches. Before any newspaper mentioned him, I reported that Chelsea had sent scouts to Qatar. Seventy-two hours later the media confirmed it, and the deal closed at 121 million euros. The piece drew more than 40,000 reads.

The conclusion did not come from a complex model. It came from every row sitting in the right place: right player, right position, right competition, right time window. Had one of those fourteen rows been mislabelled, the final ranking would have changed, and I would have issued a wrong recommendation with entirely undeserved confidence.

A single bad row does not destroy a system. It just makes the system lie very convincingly.

The contrarian angle

Football analytics spends most of its attention on models. Debates circle around which algorithm is better, which metric is more advanced, which model predicts more accurately. Meanwhile most real failures sit in the labelling layer, the lowest and least glamorous step in the whole chain.

The first reflex is to blame the algorithm. But the taxonomy was written by people. The decision to use place names as a fallback layer was a human decision, taken to save time and labour. An accident in Monterrey entered a football database not because machines rebelled. It entered because somebody decided geography was a good enough signal, and nobody went back to audit that decision.

The second reflex is to treat bad rows as harmless noise that large models will swallow. That argument holds only where data is abundant. Vietnamese youth football does not have abundant data. There, a bad row is not noise; it is the entire story of a person.

This is the point I want to hold longest. The mechanism that placed a landslide in Nuevo León inside a football feed is the same mechanism keeping some seventeen-year-old off every scouting list: labelling by surface signal, no re-check, and letting error accumulate in silence. Both are classification failures. One produces a story in the wrong place. The other produces a career nobody looked at.

Some will say the answer is manual checking. I do that every day, and I know its limits: human hands do not scale, and those same hands produced the bad cells in my 2026 spreadsheet. Data discipline is not a choice between machine and human. It is the design of a process in which every label can be challenged.

What remains

The story of Alma De la Rosa and Cristo Morales will close within weeks. Monterrey authorities will publish their findings on responsibility, and one more line will be added to the record on press safety in Mexico. Most readers will read, feel for them, and move on.

For me the value lies elsewhere. I am keeping that record as a test case for my own system: a row that once got through, and from now on must be stopped before it touches any analytical table. One wrong label has been found. That will not help anyone in Monterrey, but it may save a young player from being read wrongly over the next few years.

If a landslide on Morones Prieto Avenue can be filed under football, how many other rows in your dataset are still waiting to be re-checked?