Wrong Labels in Football Data: When a Horse Story Walks Into the V-League Analytics Room

A male horse, roughly 18 months old, was struck by a vehicle on the Santa Catarina highway in the Tláhuac borough, in the south of Mexico City. The Animal Surveillance Brigade (BVA) of the Mexico City Secretariat of Citizen Security (SSC) attended the scene, safeguarded the animal, recorded multiple injuries, and transferred it to a facility in Xochimilco for veterinary assessment. Nowhere in that report is there a team, a player, a coach, a competition, or a single sporting metric. Yet that item sat inside a content pipeline carrying the label football. I read it one evening, right after closing the goal dataset from a V-League matchday. The feeling was strange. On one screen I was counting the origin points of shots; on the other I was reading about a horse on a highway in Mexico and a label that had been attached to the wrong thing. On the night the World Cup feed went down in 2026, I learned to see a match in the dark. Tonight I learned something else: the most dangerous thing in an analytics room is not missing data, but mislabelled data that looks perfectly tidy. That report also carried no publication timestamp. A small detail I will come back to at the end. How a content pipeline works is fairly simple, and precisely because it is simple, it breaks easily. An article enters the system and receives a vertical label, then the system extracts entities such as people, organisations and competitions, then it scores sentiment, and finally it feeds aggregate indices. If the vertical label is wrong at the gate, every layer behind it is contaminated. Entities pulled from an animal-welfare report become nodes in a football graph. The sentiment of a rescue story is scored as fan mood. What mechanism attached the football label to the Tláhuac story? The report uses a vocabulary that easily fools a keyword-driven classifier: brigade, elements, specialised unit, specialists, veterinary zootechnicians, security personnel. Under an over-broad rule set, those words read like a squad list, a coaching staff, a matchday delegation. The Animal Surveillance Brigade is parsed as a sports organisation. The deployed personnel are parsed as a registration list. And so a road-safety incident walks into a football database. In Vietnam, I meet exactly this kind of error every transfer window. Based on my experience watching V-League matches and watching how domestic sports feeds operate, I see three layers of error that usually stack: wrong label, wrong entity, wrong source. The third layer is the one readers see least, and the one that decides everything. In the horse report, only four of fifteen information points carry explicit attribution, and all four trace back to a single source, the SSC. The rest, including the headline, the location framing, and the causal claim that the animal was hit by a vehicle, carry no attribution at all. There is no independent witness. There is no second veterinary facility. No transport authority speaks. The entire story is seen through the lens of the very agency that responded to it. That is a notable source structure, and it is also the familiar source structure of a great many transfer stories we read every day. A typical transfer item at peak season has exactly one source, usually unnamed, and the rest of the information is stitched together by the writer. After three republications it is no longer a rumour. It is a fact. By the fifth republication it is a statistic, because somebody has typed it into a spreadsheet. The transfer market is like a chess game in which everyone believes himself to be a grandmaster. Stopping at media criticism would make this story unremarkable. What interests me sits one layer deeper: what a wrong label produces inside a tactical analysis system. Picture a V-League club preparing for the next matchday. The analysis unit holds a dataset on the opponent's wide attacking patterns. Mixed into that dataset are a few items of contaminated origin, generated by articles that never discussed football but carried the football label. Each bad item tilts the picture slightly. Not enough for anyone to notice. But enough for one small decision to land in the wrong place, and in football the smallest error margin is the most expensive kind. Before trusting my eyes, I choose to trust structure. I have drawn colour-coded pitch maps for nearly every analysis piece since 2026, when I left the print newsroom. My first piece for a new football site dissected coach Park Hang-seo's 3-4-2-1 at the 2026 AFC U23 Championship. I used fourteen static frames and six passing patterns to show how the left-of-centre midfielders opened space for advancing full-backs. That article was shared more than twelve thousand times, faster than any record held by my old newsroom. The paper edition closed, but the tactical map began to open. What I took from that was not that visuals beat words. It was that when you are forced to draw, you are forced to choose. You cannot draw fourteen frames without deciding which ones matter. An automated tagging system never has to choose. It labels everything, including what should have been discarded at the door. In the summer of 2026, I and the numbers dived to the bottom of the V-League. With stadiums empty because of the pandemic, I spent six months analysing the 378 goals scored in the 2026 V-League season and built my own spreadsheet of danger zones. The surprise sat at Thống Nhất stadium: 68 per cent of goals came from the right wing, against a league-wide baseline of roughly 42 per cent. A deviation that large cannot be random. It points to the relationship between pitch width, stand configuration, and the habitual build-up direction of home sides. The point I want to stress is that the deviation means something because I know exactly where it came from. I counted every goal by hand, with a timestamp, a fixture name, a scorer name. If even a fraction of those 378 goals had come from a mislabelled source, the 68 per cent would move, and the conclusion about pitch width would lose its footing. Data never shouts, but it whispers loudly enough for anyone willing to listen. The problem is that it also whispers the wrong things at the same volume. There is another error class that Vietnamese football data people meet even more often than a wrong label: wrong entity resolution. Vietnamese names, structured as family name, middle name and given name, make collisions very common. Two players sharing a family name and a given name will be merged into a single node, and their entire statistical records blended together. A common name such as Nguyễn Quang Hải can appear at several clubs within one season, across several age groups, and nothing guarantees the system tells them apart. A young midfielder in the second division can be credited with the minutes of a national-team player. This damages individual records, and it damages every team-level index built on top of them. One layer deeper, wrong entities also hide structures we ought to be seeing. When a young player moves from club A to club B and returns to A two seasons later, the system usually records two independent transactions. Club A's domestic training index therefore looks better than reality, while that player's minutes are split across two unrelated profiles. The way big clubs use feeder networks to rotate domestic registration slots, and to turn talent from smaller leagues into satellite assets, is nearly invisible in that kind of data. Nobody hides it deliberately. The recording structure simply was not designed to see it. Something similar happens to tactical shifts that are read too quickly. When a back four keeps being punctured, a three-centre-back setup appears and is immediately described as a step forward in tactical thinking. Reading the data more carefully, I find a different story: goals conceded do not fall meaningfully, chances created drop, but the number of goals conceded from fixed patterns declines. What is being protected is not the goal, but the reputation of the decision maker. For a coach under pressure, reducing variance matters more than raising expected points. I am not saying a back three is wrong. I am saying it is not automatically progress. By the same logic, when a Gulf league buys a wave of stars past their peak, their metrics change function. Goals no longer serve to evaluate a player; they serve to sell tickets, broadcast rights and national image. In the spreadsheet, everything is still a goal. But the meaning left the building long ago. All of this leads back to the horse in Tláhuac. The failure of that system was not a lack of data. The system had plenty: enough to describe the incident, enough to assign a label, enough to extract entities. The failure was that nobody checked whether the material belonged to the field they were working in. In quality-control practice this is called a negative control: a case where the correct result is necessarily an absence. And we almost never run negative controls on Vietnamese football data. We check whether a story is hot. Whether others republished it. Whether it drew engagement. Whether the syntax parses. We rarely ask: is this football at all, and is this source actually standing behind what it says? That is the biggest blind spot, and it sits exactly where we are most confident. A busy analytics room will not stop to ask a question that sounds naive. A newsroom chasing traffic certainly will not. But mislabelled sports items do not do damage immediately. They do damage when the model is trained, when the entity graph is drawn, when a staffing decision rests on an index that was contaminated months earlier. A lesson from print: everything can die, only understanding survives. In print, I had to pass an editor, cite sources, and answer for every sentence inside a finite space. On digital platforms, nobody stops me, and nobody stops the wrong label either. That release is a gift, but it comes with no self-defence mechanism. That mechanism has to be built by the people doing the work, out of small and dry rules, such as requiring every old figure to be accompanied by a new one drawn from the last three matchdays. I have applied that rule to myself. If an analysis rests only on the summer of 2026, it is a history piece, not a forecast. The horse in Tláhuac will not appear on any football feed. It will be cared for, assessed, and the incident will slide out of the news cycle within days, exactly as the short life of an incident report suggests. Its value to us lies elsewhere: it is concrete evidence that a system can mislabel something without hesitation, and that no layer behind it is capable of catching the error. For those working with Vietnamese football data, the truly memorable content of that report is a void: nine of fifteen information points carry no source. That is the number I want to see for our own transfer stories, if anyone is willing to sit down and count. Tactics is a foreign language, and I have spent a lifetime translating it. The Tláhuac report reminded me that before translating, you must confirm you are reading the right language. Next time I open the V-League dataset at eleven at night, I will add one step to the process: check whether the material in front of me actually belongs to football. If that question makes me a minute slower, I will take the delay. The price of a wrong label is not paid overnight; it is paid when the season reaches its decisive phase, when every number has already been used to make a decision. And if fifteen information points about a match we are about to analyse also contain nine unsourced points, how different is our conclusion from a guess presented in a table?