A Ghost in the Data Pipeline: When a Pakistani Manufacturing Bulletin Wore Tennis Clothing

On a Wednesday morning, I opened the automated classification inbox and found a headline sitting among Davis Cup bulletins: "Jul LSM grows 3.03pc YoY, 9.51pc MoM." The label beside it read a single word — "tennis." I read it twice. No player. No court. Only a 3.03% figure standing there, cold and inert, like a pebble on the desk of a man long past believing in coincidence.

Thirty-eight years of staring at data tables taught me one thing: most errors do not come from miscalculation, but from mislabelling. A Pakistan Bureau of Statistics report on Large Scale Manufacturing had flowed into the very pipeline I use to track tennis. And what kept me sitting there longer than usual was this: at the numeric layer, nothing was wrong.

Context

To understand why this incident deserves writing about, I should explain how my pipeline works. Every day, the system ingests thousands of documents — journalism, press releases, statistical bulletins. The first layer classifies them by domain: tennis, football, athletics, economics. The second layer extracts entities: player names, tournaments, governing bodies. Only when a document passes both layers is it permitted into deep analysis.

This document passed. The label said "tennis." But the entity field was empty — the system returned the original instruction string, meaning it found no person and no organisation to assign. A document labelled tennis with not a single human name inside. To me, that is a clearer signal than any number.

Its actual content concerned Pakistan's Quantum Index of Manufacturing (QIM), rising from 115.62 points to 119.13 points year-on-year, and from 108.78 points to 119.13 points month-on-month. That is a story about automobiles, textiles, pharmaceuticals, chemicals, leather, furniture. Not one word about tennis.

When the stands are empty, numbers begin to learn how to sing. But this time, they sang a song that belonged to no arena at all.

Core Analysis

The interesting part: the data at the first layer was perfectly consistent. Divide 119.13 by 115.62 and you get 1.03035 — exactly 3.03% year-on-year. Divide 119.13 by 108.78 and you get 1.09515 — exactly 9.51% month-on-month. These two figures reconcile to the last decimal. Arithmetically, this is clean data.

But descending into individual sectors, the picture cracked. Automobiles appeared twice with two different growth figures — 57.01% and 57.77% — with no distinguishing time basis. Furniture twice: 22.69% and 10.10%. Chemicals twice: 0.25% and 0.50%. Tobacco twice: 35.82% and 0.55%. And a corrupted string in the middle: "non-metallic mineral products posted a growth of 6.52 percent 4.25 percent" — two figures jammed together, with no way to tell which is the growth rate and which the contribution.

Then I noticed the small-magnitude cluster: 0.01%, 0.04%, 0.11%, 0.18%, 0.21%, 0.27%. In a month where the headline index rose 3.03%, no single sector could have grown by only 0.01%. Those values are almost certainly weighted contributions to overall growth, conflated by the extraction layer with per-sector growth rates.

The crux: the contaminating agent is a word, not a number.

In the sector list, one line made me stop: "other manufacturing (football)" declining 0.22% year-on-year. The word "football" sat inside parentheses. And nearby, "wearing apparel" rising 3.87%. Pakistan is one of the world's major sports-goods manufacturing hubs. Those two lines, together, may be the only fragment touching the world of sport — and only at the farthest edge of the supply chain.

I believe that word "football" triggered the classifier. A single sports keyword inside an economics document was enough to mislabel the whole file. Data analysts call this a substring collision — the system does not understand meaning, it only matches strings.

On an Anfield night, I stopped counting numbers to listen to the ghosts whisper. But in this case, the ghost was not at Anfield. It was inside the classifier itself.

Contrarian Angle

Most people, seeing a wrong-domain document enter a system, blame the extraction layer. I argue the opposite. The extraction layer did the right thing: it invented no player, assigned no tournament to a document with no tournament. It returned emptiness. In a world where language models will fabricate anything to fill an empty field, a system daring to return "nothing" deserves praise.

The fault lies in the classification layer, and it lies there for a very human reason: we teach machines to find keywords instead of teaching them to find meaning. An economics document mentioning football in parentheses will always beat a keyword filter. The problem is not a weak algorithm. The problem is that we ask the wrong question: "does this document contain a sports keyword?" instead of "does this document concern a person, an event, a sports organisation?"

There is a greater risk few notice. If a document like this enters a "tennis news coverage" index, it will contribute one unit to the total — not because there is tennis news, but because there is a wrong label. This kind of junk number does not vanish. It accumulates. And months later, a young analyst opens the aggregate, sees coverage rising, and writes a piece on "the rise of tennis in South Asia." That entire chain of reasoning begins with one word — "football" — in parentheses.

Every dataset is a garden — the farmer sows questions, the harvest returns contracts. But a garden is only good when the farmer knows what he is sowing. Sow an economics seed in a tennis bed and the harvest will be a pile of untestable theory.

What I Might Be Wrong About

The document does not name its publishing outlet. I could only identify the primary data source, the Pakistan Bureau of Statistics; the secondary publisher is unknown, so I cannot assess its editorial standards. The 35.82% in one place and 0.55% in another — I infer the first is a fiscal-year cumulative figure and the second a single-month figure, but that is my inference, not something the document states. The label "July 2026-27 period" is ambiguous: it might mean a full fiscal year, or only July 2026, the first month of the fiscal year. If I have read this wrongly, the whole time frame is off by a factor of twelve.

And I must remind myself: the only link between this document and sport — apparel up 3.87%, other manufacturing down 0.22% — is far too attenuated to be used for anything. Pakistan manufactures sports goods, yes. But not one word in the document mentions rackets, balls, or tennis equipment. I am too old to believe in miracles, but young enough to know which miracles can be measured — and this is not such a miracle.

Takeaway

I will draw no tennis conclusion from this document, because there is simply nothing to draw. What must be done is to stop it before it flows into any model, relabel it correctly — macroeconomics, industrial policy — and then audit how many other documents are carrying similar wrong labels.

There are things data never touches — like the way a stadium breathes. And there are things a data pipeline should never be permitted to touch: domains that do not belong to it. An honest system is not one that never errs. It is one that dares to say "I have nothing here" when there truly is nothing.

A minimum threshold of valid entities before a document enters deep analysis is not a minor technical detail. It is the fence that keeps tennis stories forever tennis stories — rather than the story of some Punjab factory wearing the costume of a match that never happened.