Dirty Data in Football: How a Casting Notice Ended Up in a Transfer Tracker

I was reading my transfer tracker on a summer evening, and on the seventh line there was a film project. A title, two leads, a director, two production companies, one executive-producing banner. No club. No player. No competition. No contract. And yet the row sat there wearing a "football" label, waiting to be counted into a heat index I had never defined.

Six years of tracking youth football taught me to live with data that arrives late, data that is missing, data distorted by a note-taker who does not understand positions. This error is a different species. It does not take information away from me. It hands me something that never existed.

Context: a casting notice

The original piece was film-industry news, published by an English-language daily in Pakistan, summarising the casting announcement for Still We Met, an original romantic comedy. Mary Beth Barone wrote it, loosely inspired by her own experience, and takes the lead. Joe Alwyn co-stars. Zackary Drucker directs, her narrative feature debut, after an Emmy nomination for This Is Me. Assemble Media and Irony Point produce together. Lena Dunham and Michael Cohen executive produce through the Good Thing Going banner. Production begins in the fall in New York. The plot follows a young woman at a crossroads who meets a charming British stranger, and one unforgettable night across New York City.

Reading through the twenty-eight information points, I counted the football entities: none. Clubs: none. Players: none. Coaches: none. Competitions: none. Governing bodies: none. The domain label attached to the article still read "football".

The pipeline I work with is simple. Stage one breaks the article into discrete information points. Stage two builds analysis out of them. Here, stage one extracted cleanly, except that two mandatory fields were left blank: time sensitivity and the entity list. Those two empty boxes were the only safety catch, and it had been removed. The "football" label travelled straight downstream.

The mechanism that lets a film slip through

Every football data system has an entrance wider than its exit. The entrance takes proper nouns, keywords, character strings, frequency of mention. The exit demands meaning: which player, where, how long left on the contract, what fee, who pays. The gap between the two doors is where dirty data lives.

I once built a spreadsheet of more than 1,400 data points across twenty-three national U19 matches, just to answer one question: where does this team create chances from. The answer was fourteen percent of shots from the central corridor, enough to say the attack leaned on crosses. Underneath the raw data, I found the first brick of a generation, on the precondition that every row belonged to a real match. A stray row is only contamination wearing the shape of data.

Three consequences follow, in ascending order of severity.

First, volume metrics get poisoned. Many transfer trackers score heat by counting articles that mention a name. A film project entering the dataset adds one heat unit to an entity that does not exist in football. One article is negligible. One transfer window is enough to produce a skewed model.

Second, string collisions in identifying fields. Film titles, character names and production banners can collide with club names or player nicknames in certain languages. An automated classifier compares strings; it does not read meaning. When strings collide, the wrong label is confirmed, and the cross-check fails exactly where it was needed.

Third, the part that interests me most: the error sits in the absence of a validation gate. An article can only be mislabelled if nobody asks the minimum question. That question is short: does the piece contain at least one verifiable football entity, a club, a player, a coach, a competition, or a governing body. If not, the article leaves the analytical chain.

Mid-window, the real signal sits in structure: release clauses, wage bills, years left on contracts, the movement of agents. These are countable, checkable and priced. A casting notice has no release clause, no wage bill, no football agent. It cannot pass a competent validation gate, but it passes a string matcher in milliseconds.

I once spent forty-five days in Qatar building a twelve-criteria scoring system for fourteen young midfielders. Enzo Fernandez stood out with 91.3 percent passing accuracy across five matches, at a moment when no newspaper had mentioned him. Seventy-two hours later, the media confirmed it, and a 121 million euro deal closed. A real signal needed forty-five days and twelve criteria to surface. A false one needed a single field left blank.

Uruguay taught me something similar on a different pitch. In 2026, after the World Cup group stage, I wrote that Mbappe would lift the trophy. By the quarter-final I sat down with the footage and counted nearly eight players always behind the ball. Pace was not broken; it was drawn into space defined by someone else. Uruguay does not build walls. They build manifestos about space. A validation gate does the same to junk data: it does not deny that the junk exists, it simply refuses to let it through.

The fortress lesson came during two years without crowds. Home win rates in the Bundesliga fell from 44.8 percent to 33.2 percent. In the V-League, away teams added twenty-six percent to expected goals per match. Home advantage used to be a fortress. The pandemic taught us a fortress is only a variable. A football domain label is the same: a variable, and variables need re-checking at every ingestion.

The counterintuitive angle

The counterintuitive part sits here: an error like this belongs to the reader more than to the machine.

The machine does exactly what it was told, match and label. Readers mislabel inside their own heads far more often. During a transfer window every supporter runs an emotional filter: the name they want to see gets its credibility raised automatically, the name they do not want gets lowered. False news travels faster than true news for one reason only, it satisfies a desire that already exists.

There is a more dangerous kind of wrong than being wrong outright: being partly wrong. An article that gets ninety percent of its facts right, and ten percent wrong, will never be discarded. It gets quoted, aggregated, passed on, and the wrong tenth outlives the useful ninth. The article in question is the easiest type to handle: wrong in its entire frame of reference. It is an ideal negative control for any validation gate.

I keep an uncomfortable habit: appointing a devil's advocate inside my own head, whose job is to attack my conclusions. On this one, the advocate has a real argument. If a pipeline handles thousands of articles a day, demanding manual checks is impossible on cost. True. But a validation gate does not need to read content. It needs to count entities. A countable condition, running automatically, is far cheaper than repairing a dataset already contaminated.

What remains

If a film announcement can sit inside my transfer tracker, how many other rows in that tracker are sitting in the wrong place because I never opened them? And when a system has never been caught miscounting, are we entitled to trust that it counted correctly, or only that it counted fast enough?