Mislabeling in Football Analytics Pipelines: When the Wrong Data Reaches the Pitch

In the coordination room of a football data analysis unit, a file tagged "football" passes the automatic gate and moves into deep review. The analyst opens it. No club. No player. No formation, no expected-goals figure, not one name that belongs to a pitch. The only things inside are the Brent crude price, the WTI crude price, the Strait of Hormuz, and the names of two heads of state. A document about energy markets had travelled almost the entire length of a pipeline built for football, and was stopped only at the final stage, by the analyst rather than by the system.

That moment has a referee's quality to it. And like every refereeing moment, it is worth nothing unless someone sits down and watches the replay.

I have followed football analytics systems for years, and I have learned something that sounds obvious and is almost never practised: the quality of a conclusion can never exceed the quality of its input. A model can compute thousands of variables. But if the label attached at the first step is wrong, everything after it is a chain of correct arithmetic standing on a crooked foundation.

The pipeline here has two stages. The first breaks down the text and assigns a domain label. The second applies a nine-dimension framework: tactics and technique; club finance and the transfer market; results and the opinion cycle; league landscape; rules and governance; management and the dressing room; risk profile; media narrative and expectation; and finally the transmission chain of the football industry. Every one of those dimensions requires pitch-grounded data: line-ups, pressing metrics, wage bills, standings, contracts, sanctions.

The input record labels an article about oil as "football." Not one of the nine dimensions can be filled with real data.

I know that feeling. In June 2026, during France against Australia, I told more than four million listeners that a ball touching the armpit could not be a handball offence. I was relying on a rule I had learned in 2026. A colleague corrected me live on air: since 2026, the armpit area had been included in the definition of handball. I was wrong. The editorial desk had to issue a correction. Since that day I have understood that a conclusion is only trustworthy when the person making it publishes both the conditions under which the data was gathered and the date on which the rule took effect. The law is not in memory; it is in the data.

The actual content of the record belongs to energy markets. Brent crude fell $1.30 to $105.35 a barrel. US WTI fell $1.25 to $92.49 a barrel. OPEC cut its demand-growth forecast to 400,000 barrels per day. Shipping and insurance costs for goods moving through the Gulf exceeded one third of the commodity's value. Attacks on tankers in the Strait of Hormuz and a drone strike on Riyadh international airport sit in the same record. On the political side, the United States lifted sanctions on Russian fuel, and a bipartisan group of lawmakers objected. President Donald Trump and President Vladimir Putin are named. Two analysts are quoted: Tim Waterer of KCM Trade and Haitham Al Ghais, Secretary General of OPEC. A financial institution, ING, expressed considerable scepticism about the direction of oil prices.

Not one entity on that list belongs to football.

I applied the nine-dimension framework. All nine returned the same sentence: insufficient information, cannot assess. The tactical dimension is empty because there is no line-up, no playing style, no expected-goals figure. The finance dimension is empty because there is no club revenue structure, no wage bill, no transfer amortisation. The results dimension is empty because there is no table, no form, no fixture list. The rules dimension is empty because there is no financial fair play, no player-registration rule, no sanction. The dressing-room dimension is empty because there is no owner, no sporting director, no captain.

What stands out is that the record does contain financial data, except that it is a commodity price rather than a club balance sheet. The record does contain risk, except that it is geopolitical supply risk rather than injury risk or suspension risk. The record does contain public pressure, except that it is market and lawmaker pressure rather than managerial sack pressure.

This is where I want to slow down, because it is the heart of the problem. A coincidence of vocabulary is not a similarity of domain. Supply and demand in the oil market and supply and demand on a pitch share two words, but not a single unit of measurement. Pressure in a trading room and pressure in a stand share one word, but not one scale. When a system matches keywords only, it will mislabel. And worse, it will mislabel with confidence.

If the analyst at the second stage chose to "do their best" and write a football analysis out of oil data, the error would not stop at one article. It would enter the training set. It would become precedent. A bad record that slips through the gate today will make the gate looser tomorrow.

There is one more detail worth naming, because it is the kind of error I know best: the record carries no publication date. In my trade, a number without a date is a number that cannot yet be used. What does Brent at $105.35 a barrel tell us if we do not know which trading session it belongs to? What does a demand forecast of 400,000 barrels per day mean if we do not know whether it was published before or after the strike on Riyadh airport? The timing of information matters as much as the information itself.

For this record, I rated the value on four measures. Sporting value: one out of five. Industry value: one out of five. Timeliness value: three out of five, because the record is timely, but timely for energy markets. Reference value for a football readership: one out of five.

I once built a personal database from 523 matches in La Liga and the Champions League, starting in August 2026 and ending when the pandemic stopped football in March 2026. I logged every video-review decision, error code, timestamp, distance and ball speed. The result showed that 74% of wrong offside decisions were contested with an average delay of 47 seconds. I published a 48-page report and proposed a 30-second limit on each review. The Valencia football federation invited me to advise on reforming the process. That project taught me something this record repeats: a dataset is only trustworthy when every row in it can be traced back to its origin.

My proposal for this case has four verifiable steps. The first gate checks entity type before it checks content: if a file contains no club, no competition and no player, it does not belong in a football pipeline. The next gate separates "no data" from "zero data", so that a dimension that cannot be assessed is marked empty instead of being filled with a guess. Alongside that, the system must keep the original labelling trace so the source of the error can be traced back. And no less important is public disclosure when an error is found, because an error that is published is an error half corrected.

We tend to treat a system that always produces an answer as a strong system. In football we do the same. We reward the referee who blows, and we distrust the referee who does not. But a good referee is not the one who blows most. It is the one who blows correctly, and who knows when there is nothing to blow for. The "insufficient information" result in this record is a correct decision, not a technical failure. Referees do not need protecting. They need to be understood through correct numbers.

The frightening side is the opposite: a system that never returns an empty result. Such a system will always find an answer, even when the answer is built out of nothing. In football we have seen the price of decisions manufactured for completeness. A penalty that does not exist can still be given, if the referee believes that silence is a failure. And a mislabel that is never caught can still enter a report, if the writer believes that an empty field is a failure.

One match is only a story. Five hundred matches are the law. One mislabelled record is only an incident. But a pipeline with no domain gate will produce thousands of such incidents, and then we will call them data.

At 67, I do not need to remember everything. I need to know how to find what is right. And finding what is right, in this case, begins by admitting that a file about oil does not belong on a pitch, whatever label is stuck on it.