A "Football" Record With No Football: The Dirty Data Problem in Football Analytics

In the data batch I reviewed, one record carried the label "football". I opened it and found an administrative personnel notice from Pakistan: Captain (retd) Muhammad Sohail Chaudhry was appointed Inspector General of Police for the Islamabad Capital Territory, replacing Syed Ali Nasir Rizvi, who was moved to Director General of the National Cyber Crime Investigation Agency (NCCIA). The record contains no club. No player. No coach, no scoreline, no xG, no PPDA, not a single line about a pitch. Yet it sits inside a football dataset, still counted, still feeding a total that nobody in the processing chain was patient enough to open and verify.

What made me stop longer than the content was the mechanism that produced it. Four keywords triggered the classifier: "transfer", "Captain", "appointed" and "DG". In administrative English, "transfer" means reassigning an officer. In football English, "transfer" means moving a player. Two entirely different meanings, one identical string of characters, and an automated filter has no way to tell them apart if it only reads words. Muhammad Sohail Chaudhry is a BS-20 officer of the Police Service of Pakistan — a civil-service grade, not a team captain. To the labelling engine, the word "Captain" sitting next to the word "transfer" in the same paragraph was enough.

The error matters because of how the football analytics industry handles data. Most of the statistical material readers consume today does not travel straight from the pitch to the page. It passes through a chain: text harvesting, entity extraction, topic labelling, then distribution to news feeds. Every link has an error rate. That rate is usually small. That rate is also usually unmeasured. In Vietnam, most football readers encounter European football data through aggregator sites. That means the same faulty record can appear in several places at once, and its repetition ten times does not make it any more correct than once.

As an analyst working in Madrid and writing for Vietnamese readers, I sit at the end of that chain. I receive already-processed data, and I have to decide whether to trust it. Based on my experience following matches in La Liga and across European competitions, I have learned something uncomfortable: most mistakes in football analysis do not come from misreading a number. They come from reading the right number and placing it in the wrong spot.

In 2026, when PSG paid 222 million euros for Neymar, I wrote a piece dissecting their 4-3-3 with Neymar, Cavani and Mbappé. I used tracking data to show Neymar stretching defences and opening space for Cavani. The article drew attention. I ignored the midfield. PSG were knocked out in the Champions League round of 16 by Real Madrid. The data I used was not wrong. The conclusion I drew was, and that error came from checking only the parts that were easy to check.

A year later, at the 2026 World Cup, I predicted Spain would beat Russia 2-0 in the round of 16. Spain dominated possession and went out on penalties. I spent three weeks reviewing footage and found only five shots on target from the home side. Russia deliberately ceded the ball, collapsed into a 5-4-1 block and sealed every passing lane between the lines. Spain 2026: 75% of the time on the ball, and 75% of the pitch volume wasted. That lesson went beyond tactics; it belonged to how I read data.

Back to the Pakistan record. It belongs to the least-discussed error class in the industry: false positives at the labelling layer. The record carries enough metadata to pass the filter, but contains no football entity at all. No player, no club, no league, no governing body. If an analyst uses it in a calculation, he is counting ghosts.

The second class is the one I committed in Russia 2026: correct data, wrong context. Holding 75% possession is a correct fact. Concluding that the team with more possession is controlling the match is an inference. The gap between the two is my job. For Spain that year, I counted a large share of passes that merely circulated the ball between two centre-backs and a holding midfielder — passes that did not shift the opposing defensive block a single metre. I call them meaningless passes. A team can complete 900 passes and not one of them opens a gap.

The third class is more dangerous because it is quiet: home-made metrics, built on a small sample and then delivered as truth. In 2026, when football stopped for the pandemic, I lost my broadcast contract and retreated into data. I studied 500 matches from 2026 to 2026 and found the average home advantage was a 46% win rate. When football returned to empty stadiums, I collected 120 La Liga matches and saw that rate fall to 38%. Eight percentage points — enough to write an article and enough to be challenged if anyone asked about the margin of error. I published "The Crowd Is a Tactical Position" and deliberately stated the sample size in the opening paragraph. When the stands are empty, numbers have no cheering to hide behind. But they do not automatically become more trustworthy just because there is less noise.

There is one more error class in the football data ecosystem that the Pakistan record only reflects indirectly: errors at the event-definition layer. A shot from 25 metres blocked by a three-man wall and a tap-in from six metres into an empty net are both stored in the same data column called "shot" across many low-tier providers. High-quality xG models distinguish the two situations and output entirely different values. Low-tier models do not, and they still produce a number that looks highly professional. When I see a team with 18 shots but only 0.9 xG, the first thing I do is look up that provider's definition of a shot, not search for a tactical explanation. Most of the time, the problem lives at the definition layer, not the tactical one.

Now place the Pakistan record against a larger system. Suppose a news system's classifier has a 0.2% false-positive rate. Sounds tiny. But if that system processes 50,000 records a month, the absolute figure is 100 phantom records a month, 1,200 a year. And the error rate is not evenly distributed. It clusters, because the same keyword appears in the same kind of document. The "transfer" cluster drags in every personnel reassignment notice. The "Captain" cluster drags in every military and police bulletin. The "DG" cluster drags in every appointment of a director-general in any country.

That means the errors are not scattered like salt across a table. They pile up, and those piles overlap with the topics football media checks least. One stray record in an ocean of data is harmless. A hundred stray records of the same kind become a trend, and trends get cited.

One detail made this problem more complicated than I first assumed. Military and police keywords are not entirely foreign to football. In some South Asian football cultures, the national league system once operated with departmental teams — army, public works, police. Which means the classifier may still be right in many genuine cases, and those hits are precisely what keeps it from being fixed. A system that is wrong 0.2% of the time but wrong in exactly the same place is harder to detect than a system that is wrong 5% of the time at random. I do not have enough data to claim this is the root cause of the Pakistan record, so I leave it as a hypothesis awaiting verification. For an analyst, that is the most honest state a hypothesis can occupy.

The first reaction most people have to this story is: we need a better filter. I think that is the wrong diagnosis. The filter is not broken for lack of intelligence; it is broken because it was designed for a different objective. The football data supply chain pays for volume, not for accuracy. Nobody is rewarded for dropping a record. Nobody is penalised for adding one. In that environment, a classifier optimised to collect more will always beat a classifier optimised to be more correct. This is a structural problem, and it will persist as long as the metrics are article counts, record counts and page views.

There is one more blind spot I consider bigger than the filter itself: inside football, the word "transfer" already carries at least three meanings. Player transfers. Managerial changes. And the transfer of club ownership — something European investment groups do constantly, but under logic entirely different from buying a centre-back. All three share a keyword, a data layer and often a section. An article about an investment fund buying club shares can be pushed into the same stream as news of a player extending his contract, and neither is technically wrong. But they demand two different modes of analysis, and the reader receives a single flat feed.

Every tactical scheme is a puzzle, but the real puzzle lies where two schemes intersect. The same is true of data: the real error lies where two different labelling systems intersect on the same record, and no system takes responsibility for catching it.

The cost here is not one mislabelled article. It is the knock-on effect. When a phantom record enters an aggregate index, the index rises. When the index rises, an editor may write a piece about a rising trend in football governance problems. When that piece is published, it becomes a source for another record. The loop closes, and the root has long since disappeared. I do not treat this as a funny story about a stupid machine. I treat it as an accurate description of how dirty data reproduces.

Crisis does not ruin football; it strips off the makeup football has applied far too thickly. 2026 taught me that emotionally. The Pakistan record taught me the same thing technically: once the makeup is gone, people see things that were always there and were never inspected.

The process I now apply has four steps, and I am writing them down so readers can audit me. Every transfer record must identify a concrete player or coach entity; without a name, there is no analysis. Every number must carry its sample size and time range in the same sentence. Every conclusion about match control must place a meaningless-pass index next to the possession figure. And I keep a separate list of records whose only football signal is a keyword, to be reviewed manually before each transfer window.

A tactical analyst is like a storm chaser: the deeper into the eye, the clearer the system. But this storm lives on a hard drive, not on a pitch, and it does not rain — it only miscounts.

What I carry into next season is not a new model. It is an old habit: open the record, read the entity name, and only then trust the label. If, in the coming transfer window, someone asks me why the total number of football records in some source has fallen, I will answer that it has not fallen. It has simply stopped counting things that never belonged to it.