47 Data Points, Zero Serves: When a Sports Analytics Pipeline Mislabels Its Own Source

I opened the file that morning and the classification line at the top read a single word: tennis. Underneath were forty-seven numbered information points. I read them slowly, the way I read a post-match statistical sheet: label first, numbers second, always asking which lines are facts and which are opinions.

Point seven described an electricity distribution company. Point fifteen described an investment deal valued at 1.77 billion US dollars. Point thirty-five described transmission and distribution loss ratios. Point forty-seven stopped mid-sentence. Not one serve. Not one set. Not one tournament, ranking table, or tennis governing body.

I have spent my working life saying "not enough data". I was not used to saying "entirely the wrong domain".

Before I open a dataset, I ask where it was born. Three checks decide whether a source is allowed near any model I build.

The first is an entity check. Here, the file named distribution companies, a national energy regulator, a privatisation commission, and a foreign strategic investor. Every one of them is real. None belongs to the tennis ecosystem.

The second is a unit check. Units are the fingerprint of a discipline. This file carried electricity tariffs, deal values in billions of dollars, grid loss percentages and bill-recovery rates. Tennis runs on first-serve percentage, return points won, break-point conversion, winner-to-unforced-error ratio. Those two unit systems cannot be mixed without corrupting both.

The third is a grammar check. Every field speaks its own language. Power-sector documents talk about multi-year tariff control periods and circular debt between generators, distributors and the state. Tennis talks about surface, streaks and deciding points. This file spoke fluent power-sector grammar and could not form a single sentence in the grammar of tennis.

All three checks returned the same verdict. The label was wrong. The true domain was energy, utilities and investment policy. Not sport.

I have met this failure mode before, only smaller and harder to catch. One season a data provider silently changed its definition of possession mid-competition, and a club's share dropped from 58 to 44 per cent without changing how it played. Expected-goals models trained on one event taxonomy get applied to leagues using another, so the same shot is worth 0.08 to one model and 0.21 to the other. Offside lines drawn to the millimetre rest on calibration files that are only as good as the person who measured the pitch.

A number that is correct in the wrong domain sings in almost exactly the same voice as a number that is correct in the right one. Numbers whisper. Whoever listens closely enough hears an entire match. But you have to know which room you are standing in.

A missing record stops a system and forces a question. A mislabelled record lets it run on, confident, and produce conclusions. That is why mislabelling is more dangerous than absence.

The instinctive fix is to upgrade the classifier. I do not believe that is the right repair, and this file supports my scepticism. The first problem is the missing provenance field: not one of the forty-seven points said whether it came from a regulation, a prospectus, a report or the author's own reasoning. A system that never records where its inputs came from can never detect its own mistakes. It can only be wrong fluently. The second problem is the ratio of opinion to fact. Most points were judgements, not verifiable data, and the source itself is classified as commentary, which means it arrived with an editorial slant built in. The third problem is truncation: point forty-seven ends mid-clause, and the part lost in extraction is usually the part that mattered.

I could have turned this file into a tennis column. Rename a few entities, attach a few percentages to a player, add a line about pressure at a deciding point, and no reader would have noticed. But I would not have been able to defend a single sentence of it.

Years ago I refused a commission on crowdless football because I wanted three more weeks of data. Refusing one article is easier than refusing nine analytical categories I know cannot be applied. Misreading one variable is like losing your bearings for a whole season.

So instead of inventing a serve, I did what an analyst should: I marked the file out of domain, logged the evidence, and routed it to the right desk. To be fair to the source, it does contain a genuine analytical story about how regulatory uncertainty deters foreign capital from electricity infrastructure. I am not qualified to draw conclusions about it, and knowing which questions are not mine to answer is itself a form of knowledge.

The signal I will track in the next batches is not a match. It is the labelling log. If every record must carry four fields — domain, primary entity, unit of measure, and provenance — then the mislabelling rate becomes a measurable indicator. I will also count truncated fields and single-source items. Added together, those three numbers tell you whether a pipeline is healthy or ill.

In tennis I watch first-serve points won to know whether a player is holding rhythm. In data, the equivalent is the share of records that state their origin. The evidence currently available gives me one conclusion, and it has nothing to do with any surface: in those forty-seven points, one of the most important fields was left empty from the start.