A file labelled football: 25 information points, none of them football

A data file carried the label “football”. Inside were 25 information points. The number of points naming a club, a player, a coach, a competition, a transfer fee or a single rule of the game: zero.

I opened that file on a mid-window morning, with an inbox full of player names and fee figures. My job in Kuala Lumpur is filtering: keep what has evidence, drop what has only emotion. This file had nothing to filter. It was about Kate Hudson and Danny Fujikawa, about a seven-year-old daughter named Rani, about Oliver Hudson, about Erinn Bartlett, and about a podcast called Sibling Revelry. The original report asked why the couple remains unmarried five years after their engagement. Source: The Express Tribune; the podcast episode aired on 8 September. On almost every information point, the source column carried a single word: None.

The entity map tells the whole story. Kate Hudson appeared in 12 information points, Danny Fujikawa in 7, Oliver Hudson in 4, Rani in 3, Erinn Bartlett in 1, Sibling Revelry in 1. Validated football entities: 0. The domain label read football; the actual content belonged to entertainment and celebrity lifestyle. One entertainment item had wandered into the football drawer, and it wandered in silently.

A pipeline that cannot say no

From the outside, a modern sports news system looks more like plumbing than like a newsroom. Wire copy flows in; an automated tagging layer sorts by keyword; an entity-extraction layer pulls out people and organisations; then the labelled content is pushed into analysis. Nobody reads all 25 information points before the label is attached. The label arrives first, the reader arrives later, and sometimes the reader never arrives.

For someone who filters content for a living, this is the soft spot of an entire vertical. In a transfer window, readers are not short of information; they are short of filters. They need to know which story carries a release-clause structure and which is merely an agent speaking. Yet the first filter in the system, the domain label, is the loosest one. A story about a wedding that has not happened slipped into the football drawer over a handful of overlapping keywords: engagement, wedding, family, daughter. The machine does not read meaning. The machine counts words.

The cost is not the misplaced article. The cost sits downstream in the aggregate statistics. If enough files of this kind enter the system, the article count in the football vertical rises, the frequency of a name rises, and the sentiment indices tracking the market drift with them. Nothing in the summary report will raise an alarm, because every figure still prints in the correct format. Based on my experience following matches in the Malaysia Super League and across the region, this class of error always begins with a data field that looks entirely ordinary.

The template fabricates; the model does not

My first blog page was not about football but about the gap between Johor’s two centre-backs. In 2026 I spent three weeks rewatching Johor Darul Ta’zim against Kedah Darul Aman, counting every pressing action. Johor’s average PPDA that season was 14.2, meaning opponents were allowed 14 passes before each active defensive action. That number only means something if I have defined what an active defensive action is. Remove the definition, and 14.2 still prints, still looks clean, and is still wrong.

That is what came to mind when I looked at the 25-point file. The analysis stage behind it is built to answer nine fixed dimensions: tactics, finance, results and public opinion, league landscape, rules and governance, dressing room, risk, media narrative, and industry transmission. Nine boxes. For a story about a wedding that has not happened, those nine boxes have nothing to fill. A model does not invent clubs when it lacks data; it invents clubs when the template forces it to fill nine boxes.

The trade taught me that effort metrics are the easiest numbers to lie with. Distance covered and sprint counts are packaged as measures of will, but running without purpose still produces pretty numbers. A midfielder who covers 12 km without cutting a single passing lane still ranks above one who covers 9 km and kills three counter-attacks. Running without purpose still produces pretty numbers, and a template forced to fill nine boxes produces a beautiful piece of analysis in exactly the same way: complete, fluent, and carrying no information at all.

Covid took away the stands but gave me back a way to measure home advantage without needing the sound of a crowd. In 2026 I pulled data from Europe’s five major leagues before and during the behind-closed-doors season. Average home win rate fell from 46 per cent in 2026-19 to 39 per cent in 2026-20 once matches were played in empty stadiums. Proactively pressing sides lost around 11 per cent of their effectiveness. That model held on one condition alone: the sample definition did not change. If half the rows in the dataset come from another domain, the comparison collapses at the first row, even though the chart still draws a very persuasive curve.

Before the ball was circulated, I had already seen three decoy runners and one real path. At the 2026 World Cup I sat for two days with the tape of Saudi Arabia’s 2-1 win over Argentina. Messi was caught offside seven times in the first half. Saudi Arabia’s defensive line held an average of 52 metres from goal and stepped up together under one synchronised rule: when the ball went into central areas, the whole back line moved as one. To count those seven offsides, I had to strip from the dataset every moment that was not a live phase of play. One extra count, and Messi’s offside rate in that match becomes a meaningless figure. Bad counting does not leave the sheet empty. It only makes the sheet lie.

In 2026, working in the data analysis group for the expanded 32-team Club World Cup in the United States, I built a logistic fatigue coefficient from kilometres flown, consecutive matches and pitch temperature. Major European sides lost roughly 18 per cent of their scoring efficiency when they had to fly over 4,000 km with fewer than three days between matches. The model called three of the four quarter-finals correctly. A colleague said football cannot be reduced to mathematics; I did not argue, I just printed the chart and pinned it to the board. But even that model stood on a condition identical to every model before it: the input dataset had to be football, all of it, and nothing but football.

The quiet risk

People fear transfer rumours because rumours are loud. A rumour has sources, credibility tiers, rebuttals, a day it finally dies. A mislabelled data file is silent. It sparks no argument, no lawsuit, no correction. It sits in storage, gets counted, gets archived, and flows into every downstream summary report.

The counter-intuitive point sits here: the most dangerous thing in a football data system is rarely false content; it is true content that belongs in another room. A celebrity report that is accurate to the last detail is still waste in the football drawer, and the reverse holds too. This kind of content is harder to catch than fake news, because there is nothing to catch. No number is wrong. There is only a number that does not belong here.

The correct handling in this case is refusal. No tactical analysis, no financial speculation, no assigning public pressure to a manager who does not exist. The correct conclusion takes the form of one short line: out of scope, insufficient information. For a writer, that is the hardest and least rewarded skill, like a defender holding the line by refusing to step out. Nobody remembers the challenge he did not make.

There is also a problem that sits outside every technical dimension. The original report named a seven-year-old child and private family matters. Whatever the domain label says, that file should be excluded from every storage, cross-checking and republication pipeline. A wrong label is a technical fault. Continuing to process such a file after you know what it is stops being a technical fault.

The remedy is simple and needs no sophisticated model. Add an entry gate: a file only enters the football drawer if it carries at least one validated football entity, even if that is just the name of a club. Track the share of information points with a blank source field; once that share passes 70 per cent, lower the confidence weight of the entire source instead of waiting for a human to notice. And audit the vertical to see how many similar files are lying quietly in place.

For me, the question that tests the next data batch is not what percentage the model predicted correctly. The question is this: if that file arrives again tomorrow, is anyone in the chain awake enough to say one short sentence, that this thing does not belong here. A match truly begins not when the referee blows the whistle, but when a defender decides to leave his position. Data pipelines work the same way: they only become trustworthy from the moment someone dares to stand still.