Empty Data, Filled Template: Football Analytics' Most Dangerous Output Is the Missing Number

Empty Data, Filled Template

Hook

A report landed on my desk last week. Nine sections. Thirty-three tables. Every cell populated. Actual facts: zero. The file carried a "Football Domain" label and an "Overall Risk Rating: High," but the list of information points the entire structure was supposed to rest on was empty. The system that produced it had failed to capture the article's title, its source, its body text — and had gone ahead and built nine analytical dimensions anyway. In forty years I have read a great many wrong reports. This was the first that was not wrong. It was hollow, and it looked complete.

Context

Sports data exists to turn raw events into something a decision can stand on. In the 2026–99 Premier League season I worked as a match coder for a London data firm, logging the origin and destination of every pass by hand. I built an 18-zone grid on graph paper and used it to reverse-engineer the width in Manchester United's 4-4-2 — width that came from Gary Neville and Denis Irwin pushing to the touchline, not from Beckham and Giggs hugging it.

For 2026–05 I tracked all 38 of Chelsea's league matches myself, measuring the distance between the back four and the midfield line: roughly 22 metres with the ball, compressing to 14 without it. The output was 15 goals conceded, 24 clean sheets, 95 points and the title. That season got filed as boring. Strip the badge and the noise, and what remained was a mechanism, not a mood.

The work is automated now. A top-flight match yields well over a thousand events, and in several leagues those events travel from the stadium straight into betting markets. What changed is speed. What changed with it is that the pipeline no longer has anywhere to write "I don't know."

In 2026 a Mumbai website asked me to audit the inaugural ISL's marquee signings. I built aging curves from ten seasons of Serie A minutes and predicted that Alessandro Del Piero (39, Delhi Dynamos), Marco Materazzi (41, Chennaiyin) and Robert Pires (41, FC Goa) would each finish under 1,200 league minutes. Del Piero stopped at 10 appearances, Pires at 8. Every number in that piece carried the season it came from. The price tag arrives before the player; the audit arrives after — and an audit is worth something only if it declares its sample.

Core

The anatomy of that report's failure is legible. Metadata survived; the body died. The file said "football" but named no club, no player, no match, no date. The system caught the shell of the page and dropped the meat inside it. The error sits in ingestion, not modelling: the page was behind a paywall, or it was a video or live-blog shell, or the source language and the extractor's language never met.

The deeper hole is in the schema itself. The instruction for extracting entities read: "identify from the information points above." The information points list was empty. The instruction became vacuous against its own premise. A field that should have fallen back to the article body had never been taught to fall back at all. A good data system designs its empty spaces. This one was covering them with "N/A."

And this is where my real objection starts. Thirty-three entries in that report read "N/A – insufficient information." That is honesty, and I respect it — it is far better than manufactured "hidden information" invented to fill a template. But N/A is not a negative finding; it is a refusal. Downstream aggregation, ranking and alerting layers cannot read that distinction. To them, N/A is a value. The disclaimer printed at the foot of the report — that N/A rows must not be read as conclusions about any club — is itself the proof of how real the hazard is.

One more thing. The highest rating in the document, "High," was assigned to the pipeline's own internal data gap, not to any club, coach or competition. Technically correct, and quietly misleading. On whatever dashboard that report feeds, "High" always means risk on the pitch. Same word, two meanings.

I started counting touches because nobody else was counting. One rule came out of that habit: if I cannot re-derive a number, I do not publish it. Could a single one of those thirty-three numbers be re-derived? No. It went out anyway.

Contrarian

The reflex worry about football data is the wrong number — a bloated xG model, a misclassified shot, broken tracking. I would argue the wrong number is the least dangerous thing in the stack, because it gets caught. A wrong number starts an argument; an argument forces a re-check; a re-check produces a correction.

The dangerous object is a number that is absent while appearing present. Every system leaks somewhere. Find the leak and you have found the system — this report's leak was that eight of its nine dimensions had no subject at all, and a headline rating was printed regardless.

Live betting feeds are the fastest illustration. A feed has no N/A field. When it drops, the market does not stop; it keeps pricing on the last tick, and the last tick is the price of an older truth rather than of the match in front of it. That is the business model in one line: pausing is expensive, being wrong is cheap. My own instinct runs the other way. I file the day after, with a re-watch and a log, because arriving late with a count beats arriving first with a guess.

Takeaway

Next time a clean, complete table appears on your screen during a match, ask one question: which event produced its first line? If the answer is "the feed," ask the next one — where was the person who coded that event sitting? The day a system learns to refuse an answer, trust becomes real. Until then, the fullest table on the page is the one to interrogate first.