An Empty Analysis Sheet: When Sports Data Work Has to Learn to Say “Not Enough Information”

2:47 a.m., Brisbane. The third night of the week, and I was sitting with an esports analysis document sent over by the editorial desk. It had nine sections, framed as neatly as a blueprint: patch and meta, tournament format, roster and players, regional picture, club finance, competitive-rules compliance, risk profile, public narrative, and industry transmission chain. I scrolled down. First line: “Game: N/A”. Second line: “Version: N/A”. By the ninth line it was still N/A. Only one cell carried content, and that content was a label: “esports”.

I am used to tables that open with an anomalous number. One season I spent nearly a month rewatching 19 Melbourne City match tapes because one xG column refused to match what my eyes had seen. That night was different. No anomalous number. Nothing to cross-examine. The whole sheet told me exactly one thing: the information had not been retrieved, and the people behind it had recorded that fact instead of inventing a story to fill the space.

What actually happened is the part rarely written down.

A sports-data analysis pipeline usually runs on two layers. Layer one reads sources — articles, press releases, results tables, footage — and extracts information points: verifiable statements such as who moved to which team, for what fee, which patch changed which stat, who was suspended for how many matches. Layer two takes those points and builds the analysis. Layer two never reads the original article. It lives entirely on what layer one brings home.

That night, layer one returned an empty list. The skeleton was correct: correct field names, correct domain label, blanks marked according to convention. But inside there was not a single statement. In other words, the system finished the template and failed at the retrieval. In this trade those are very different failures. A broken template is a half-day fix. Missing content means going back upstream: JavaScript-rendered pages, image infographics with no text, paywalls, or simply a link that returned an error page.

What stands out is that the “esports” label survived intact. That label was almost certainly assigned from metadata — a section, a URL path, a vertical’s name — rather than from article content. A system classified the topic correctly while failing to read a single word of the article. For anyone who works with data, that is the most uncomfortable kind of failure: everything looks fine until you open it.

Why did all nine sections collapse at once? Because every section needs at least one proper noun.

The game title is the first hard gate. Publisher cadence varies too widely to generalise: Riot patches every two weeks, Valve ships a major update every few months, Tencent runs on seasons. How metrics are defined, how tournaments are structured, how money is made — all of it is tied to a specific title and does not transfer. Without a title there is no patch, no meta, nothing.

Tournament format is the next gate. The gap in upset probability between a single-elimination BO1 bracket and a double-elimination BO5 bracket is one of the strongest explanatory variables in any esports outcome model. Without format, every probabilistic statement is baseless, including the ones that sound obvious.

The player section is the one I regret most seeing empty. The three most useful early-warning signals in player analysis all require a specific name: the reflex-decline curve in entry or carry roles, a history of wrist and tendon injuries, and the contract-year effect. Without a name, all three become meaningless.

The regional picture works the same way. The same region can sit at entirely different tiers depending on the title: strong in one MOBA, a wildcard at best in a shooter, or in a different MOBA altogether. Generalising across titles is a methodological error, not a nuance.

Club finance needs numbers. The sponsor-concentration threshold — a single sponsor accounting for more than half of revenue — is a familiar red flag, but you need a revenue table before you can say anything. Rules compliance is harder still: it depends on jurisdiction, and jurisdiction differs between publisher rules, league rules, third-party organiser rules and national policy. This is the section most sensitive to missing metadata, because it lives not on article content but on names and territories.

The public-narrative section needs a subject and a claimed outcome, so that heat can be measured against fundamentals. The industry-transmission section is the most context-hungry of all: it needs the publisher, the title, the region, the platform and the tournament simultaneously. Lose one element and it degrades badly; lose all five and it stops working.

But one detail is worth more than the nine sections combined, and it sits in the skeleton rather than the content. Two fields in the pipeline are designed to depend on each other: the “entities involved” field must be derived from the “information points” above it, and the “source quality” field must be judged from “source fields” — while both of those upstream fields are empty. At that point the error is no longer a retrieval error; it is a design error. A system has placed itself in a position it cannot answer from, and then returned N/A like a confession.

This is where I want to linger. A risk table full of N/A is very easily read as a clean table. A reader skims it, sees no red flags raised, and assumes nothing is wrong. But silence is not safety. In sports data, the three highest-consequence risk categories — unpaid wages, match-fixing, injury — are precisely the ones where an absent signal usually reflects missing data rather than a healthy condition.

I learned that the expensive way. In 2026 I wrote a piece criticising Jamie Maclaren for scoring only 8 goals in a season with an xG of 14.2. My editor struck out nearly all the numbers on the grounds that “nobody understands this”. A month later I sat through 19 match tapes looking for which shots genuinely deserved to count as clear chances — and realised I had been right about the numbers and wrong about the telling. Every number has a story; my job is simply not to ruin it.

There is a counter-intuitive angle here, and it has nothing to do with esports.

The industry’s problem is not empty data. The problem is the pressure to fill it. A table with team names, player names and a few attractive figures gets shared within fifteen minutes. An N/A table gets dismissed as useless, even as a sign of laziness. Yet the N/A table is the honest one, and it is the one that protects readers from an analysis that reads beautifully and is not true.

I have seen that mechanism run the other way. In 2026, writing about France against Argentina in the World Cup round of 16, I was absorbed by Kylian Mbappe’s top speed of 37.6 km/h on the decisive assist. None of my pressing or xG metrics could explain the raw beauty of that acceleration. I stayed up two nights breaking down frame after frame and understood: data measures what it measures, and it does not measure what makes people love football. That gave me no licence to invent a metric. It only gave me licence to write one more sentence.

In the summer of 2026, when competitions froze, I rebuilt Andrew Robertson’s movement table from Liverpool against Barcelona: 12.4 km, of which 2.1 km was sprinting. There was no new match, but old data could still be verified, still cross-examined against footage. An empty table has no such capacity. It cannot grow old, and it cannot become correct.

In this particular case the temptation was greater. Pick a trending title, attach a familiar Vietnamese team, add a few names and a plausible Pick/Ban rate, and a very smooth analysis writes itself. Nobody checks, because everybody is busy. Doing that turns an honest process into a machine for manufacturing illusion. And in an industry where young Southeast Asian rosters live on their credibility with sponsors, data illusion is not a small matter.

One more thing: correlation is not causation, and an empty sample is not a bad sample. If a piece about a young player cites no metrics, the correct conclusion is “not yet assessable”, not “this player has not proven anything”. The two sentences look almost identical on paper and differ by a world in consequence.

That night I did not edit the analysis. I added a line at the top of the file: refuse to run layer two until at least one verifiable statement exists. Then I proposed three small things: a hard gate at the game-title stage, a new measure called the emptiness rate — records with no information points per thousand fetches — and a rule that any all-N/A risk table must carry a warning that silence does not mean safety.

Those three things do not make the writing better. They only make it more correct. When the numbers speak, the stadium must learn to be quiet — but when the numbers fall silent, it is the writer who must speak, and the most honest thing to say then is the hardest sentence of all: I do not have enough data.

At 39 I have learned that data also hurts when it is twisted. The question I carried back to Brisbane that night was not which team is stronger. It was this: if a machine has learned to say “I don’t know”, can the people running it learn the same before they hit publish?