Mislabeled in the Football Data Pipeline: 43 Data Points, Zero Passes

Tuesday night, I opened a file in my personal data store, waiting for pressing numbers from a Malaysia Super League match. The first line read: a 2,500-square-metre parts warehouse in Hanoi. No starting eleven. No PPDA. Not a single player's name. I sat with it for two hours, read all 43 information points, and asked myself one simple thing: how did a car product introduction end up inside my football data feed?

My first blog post was not about football. It was about the gap between two Johor centre-backs. In 2026 I spent three weeks re-watching Johor Darul Ta'zim against Kedah Darul Aman, counting every pressing sequence, and arrived at a PPDA of 14.2. The first draft ran 2,500 words, most of it speculation about player psychology. I cut nearly all of it and kept the data and the diagrams. A community page shared it; 12,000 reads in the first week. Since then I have followed one rule: numbers as the foundation, observation as the surface.

That rule forces me to look at the pipeline behind every analysis. Football data does not fall onto my desk by itself. It passes through automated crawlers, taggers, and domain classifiers. A classifier works on probability, not comprehension. It sees a familiar string, a brand name, a sentence shaped like sports copy, and it assigns a label. The file lands in the right drawer. Nobody rechecks.

I know the cost of not rechecking. In 2026, with stadiums shut, I rebuilt a home-advantage model and found home win rates across Europe's top five leagues falling from 46% to 39%. In 2026 I counted Messi caught offside seven times in the first half against Saudi Arabia, simply because the opposing back line stepped up together around the 52-metre mark. Both conclusions only held because the inputs were clean. In 2026, inside the data group for the 32-team Club World Cup in the United States, I built a fatigue coefficient from flight kilometres and rest days; the model called three of four quarter-finals correctly. One contaminated file, and the accuracy collapses.

That contaminated file looks like this: 43 information points, all belonging to Tasco Auto and Geely Vietnam. No team, no player, no competition, no transfer, no regulations. The entities include the Geely EX5 EM-i, the Geely EX2, Carpla Service, and VETC rescue. The quantified facts: more than 80 authorised service workshops, two parts warehouses totalling 2,500 square metres in Hanoi, a 1,000-square-metre southern warehouse expected to complete by the end of 2026, a 90% availability rate on common replacement parts, 1-3 days for internal supply, 15-18 days for urgent imports, a late-delivery compensation policy of 100,000 dong per day, a 0.5% referral payment on car value, and a 1% discount for a second car in the same household.

Two customer testimonials sit alongside: Mr Tran Hieu, an EX5 EM-i owner, and Mr Nguyen Dinh Khanh, an EX2 owner. The entire technical section is carried by a single company voice: Mr Nguyen Trung Kien, Director of After-sales Service.

When I ran the file through my analytical frame, nine dimensions returned N/A. Tactics: N/A. Transfers: N/A. League table: N/A. Regulations: N/A. Dressing room: N/A. Injury risk: N/A.

To an outsider, a run of N/A looks like failure. To me it is the hardest and most valuable part of the job. A match does not truly begin when the referee blows the whistle; it begins when a defender decides to leave his position. In a data pipeline, the notable event is a file leaving its proper drawer, not a file being present. Writing N/A is drawing the offside line. Without that line, everything downstream is wrong from the root.

Nine N/As are nine refusals to fill a gap with guesswork. Had I forced a formation onto this file, I would have produced something that looks like analysis and stands on nothing. That is how football data gets poisoned: by treating off-domain material as if it were on-domain, rather than by obvious fake news.

Set the wrong label aside, and the file still has a readable structure. It describes a chain from manufacturer to distributor, to service network, to owner community, and back through a referral loop. Placed beside a football club, the chain almost mirrors a fan-retention funnel: the club supplies the match, the service system supplies the audience, the community supplies itself, and the fans bring more fans.

One detail stopped me for a long time. Mr Nguyen Trung Kien says after-sales cannot be measured by the number of workshops or the scale of parts warehouses. Yet the same document, further up, is promoting more than 80 workshops and a 2,500-square-metre warehouse. The two messages push against each other. The spokesman is pre-empting a criticism: a network that is wide but thin. I noted it, because it is exactly the kind of fact the reader needs and the copywriter wants buried.

This is where I can connect back to football. The car industry self-reports a 90% parts availability rate, with no independent audit. Football self-reports distance covered. A midfielder runs 12 kilometres, a handsome effort metric, while most of that ground cuts no passing lane. Ineffective running still produces beautiful data. A self-reported 90% and a self-reported 12 kilometres are the same disease: a metric that measures activity, not impact.

Every diagram lies when you watch from the stands; the truth lies on the grass, where the gaps move. A data label is a diagram too. It draws the position the content is supposed to occupy. Only by reading the actual text do you discover the content left that position long ago.

The irony is that this contaminated file is more useful than many clean ones. To test a filter, you need a sample you know for certain is wrong. This case is that reference sample. It exposes a classifier reading surface keywords instead of meaning. And it reveals an ingrained habit in sports data: better to over-collect than to miss. Over-collecting flatters volume metrics. Missing invites questions. So nobody deletes.

Strip the wrong label, and one dated milestone remains worth tracking: the 1,000-square-metre southern warehouse, expected to complete by the end of 2026. That is a rare self-imposed constraint, with a date attached and the possibility of verification. At the end of 2026 I will reopen this file. If the warehouse is not finished, or if urgent import lead times still sit around 15-18 days, we will know how much of the after-sales capability story was surface only.

One task I leave for myself. How many other files in my data store carry a correct label over incorrect content, and have never been checked by anyone? I do not have the answer yet. But from now on, every time I open a file to analyse it, I will read the first line before I trust the label on top.