A Fuel-Price Bulletin Bearing a Tennis Label: The Discipline of Data Tagging in Sport

The effective date printed on the document was 24 September 2026. The label our system assigned to it was “tennis.”

I opened the record and read it. Fourteen information points. Not one line mentioned a player, a tournament, a surface, a ranking, or a single game. The only thing present was fuel pricing: high-speed diesel down 4.21 rupees to 414.75 rupees per litre; petrol down 1.93 rupees to 390.12 rupees per litre. Elsewhere in the text, Brent crude rose 1.84 dollars to 101.09 dollars a barrel, and WTI rose 0.69 dollars to 91.21 dollars. The named institution was the Petroleum Division. The named person was US President Donald Trump, with a warning directed at Iran. And one further subject — the most valuable of all — had been truncated out of the document entirely.

That is the whole content. And that is the whole problem.

Work with data long enough and you learn something uncomfortable: the most serious errors rarely come from wrong numbers. They come from labels placed in the wrong place. A spreadsheet can add and subtract perfectly down to the last unit. A pipeline can run smoothly across thousands of records. And all of it can still be answering a question that has nothing to do with the thing you are trying to establish.

Data never rushes. It is the people in a hurry who get it wrong.

I have covered sport for twenty-five years, most of it spent closer to spreadsheets than to stands. My job is to turn a match into a verifiable file: who did what, at what moment, at what probability, and where the margin of error sits. In a courtroom, the order of presentation is everything. In a spreadsheet, the same is true.

That is why I regard labelling as the single most important step in any sports data system. A record tagged “tennis” travels down a tennis pipeline. It gets checked against first-serve points won, break-point conversion, return points won. It is forwarded to the people who analyse tactics, surfaces and calendars. None of them has any reason to doubt the label, because the label is the reason they are in the room.

In 2026, mid-way through the V-League season, I published the first series applying expected goals to Vietnamese football. Hai Phong met SLNA at Lach Tray; the hosts generated 1.92 xG and lost 0-1. The press called it decline. I called it random injustice: the opposing goalkeeper saved 11 shots, 3.8 times the average. The piece was ridiculed for two weeks, until Hai Phong’s head coach cited my numbers in a press conference.

The lesson I kept from that day was not about xG. It was this: had I mislabelled that match at the outset, the entire chain of reasoning behind it would have collapsed — and it would have collapsed in a way that is very hard to detect.

Back to this morning’s document. I checked it layer by layer.

The first layer is arithmetic. The old diesel price was 418.96 rupees, the new one 414.75 — a difference of exactly 4.21. The old petrol price was 392.05, the new one 390.12 — a difference of exactly 1.93. Both subtractions match the figures in the headline precisely. The document is internally sound.

The second layer is cadence. The previous review recorded cuts of 3.12 rupees on diesel and 1.70 rupees on petrol. This one records 4.21 and 1.93. Two consecutive cycles, same document type, same issuing body. This is a recurring news beat, not an isolated event. In my experience, recurring beats are the records most prone to mislabelling, because they resemble each other so closely that a system treats them as one continuous stream rather than as separate documents.

The third layer is the damage. Two of the fourteen information points have lost their subject and a proper noun. One reads as a broken sentence, entirely missing its actor. Another exposes a fragment of a name that was cut off. These are the fingerprints of a text-extraction failure. But the consequence mirrors a human error: a system that reads a subject-less sentence will not raise an alarm. It will learn a false pattern.

The fourth layer is the blank fields. Three mandatory fields of the first processing stage — entities involved, time sensitivity, source quality — were left empty. The entities field contains an instruction rather than a result. The time-sensitivity field states that it was not assessed. The source-quality field states that it should be inferred from the input data.

This is the part that made me stop longest. Had the entities field been filled in correctly, it would have contained two words: Petroleum Division. Place those two words beside the label “tennis” and you have an instant disqualifying signal. The safety valve that should have blocked this error was precisely the valve left unset.

A label is not evidence. A mislabelling system can run smoothly across thousands of records before anyone bothers to open one and read it.

There is one further detail about the source structure. The text refers to Platts rates, premiums and incidentals — that is, an import-parity pricing formula, something a specialist energy desk uses fluently and a general-assignment desk does not. That implies the outlet has its own energy desk, and very likely its own sports desk, both feeding a single wire. When two desks feed one stream, cross-contamination is a structural risk, not a random one.

I have met this exact mechanism at another scale. In June 2026, before Germany faced South Korea in the World Cup group stage, I published an analysis showing Germany’s pressing coefficient had fallen from 8.1 PPDA in 2026 to 12.6, with average distance covered down 6.2 km per match. Germany collapsed in my spreadsheet before it collapsed on the pitch. On the pitch: 74 per cent possession, a 0-2 defeat, elimination in the group stage.

What the two stories share is not their conclusions. It is this: in both cases the warning signal was already sitting inside the data, and what decided the outcome was whether anyone bothered to read it. In Germany’s case, I read it. In this fuel-price document, the system did not.

One more detail I recorded without drawing a conclusion. On the same day of publication, Brent printed at 101.09 dollars a barrel, up almost 1.85 per cent, while the domestic retail price was cut. Two movements in opposite directions. For an energy analyst that is a real question: a lag in the pricing window, or a subsidy decision? For me it is only a reminder that every field has its own questions, and no field should be answering another field’s.

Intuition suggests that if a record carries a “tennis” label, someone must have checked it. That intuition fails across most modern data systems. The label is generated at the first stage, usually by a machine, and from then on it is treated as an established fact. Nobody re-examines something already deemed correct.

The worrying part is not the scale of this error. The scale here is total, and therefore easy to see. The worrying part is the gap it exposes: a record whose label is correct but whose body drifts into another subject will not be caught. That error class is far subtler. It produces analyses that read perfectly well, with numbers and charts, and are wrong from the foundation up.

People remember results. I remember the conditions that produced them.

There is one more temptation worth naming. When an analysis template demands nine dimensions, and the input contains zero relevant information, the pressure to fill the template is enormous. The inexperienced writer fills it. The disciplined writer returns an empty result. I choose the second, even when the second looks less impressive.

Every shot is a hypothesis. xG is how we test it. A fuel-price bulletin is the same: it is a hypothesis belonging to the energy field, and no tennis test can be applied to it.

This document will be routed to its correct pipeline and removed from every downstream tennis product. But its real value is not in being discarded. It is in showing us where the pipeline is blind, at the exact moment when that blindness is still cheap to fix. Insufficient evidence is always a valid answer — and to me, it is the most honest one a spreadsheet can give.