Nine Analytical Sections, Not a Single Data Cell: How Sports Analytics Fools Itself

A document of nearly three thousand words was sitting on my desk in Liverpool on a winter morning. It had all nine sections: technical and tactical analysis, data and form, tournament system and schedule, the tour landscape, rules and governance, team and player management, risk, media narrative and expectation, and finally the industry transmission chain. Every section had a table. Every table had column headers. And across more than forty cells, not one contained a real value.

Some cells said N/A. Some said “insufficient information, cannot assess”. Some said “not assessed at Stage 1”. One or two even carried a confidence label: Confidence — High. A document that could still grade its own confidence as high for knowing nothing.

It was right in a way that leaves no room for argument. The document had no original article title, no source name, no list of information points, no resolved entities. A nine-layer analytical engine was set running on an empty input, and it returned exactly what an empty input must return: an empty result. No player was invented. No first-serve percentage was fabricated. No draw was sketched.

Ten years ago, I would have filled those cells in.

I would have picked a name that was being talked about, built a plausible-looking analytical frame, dropped in a few numbers about first-serve percentage and return points won, and written a conclusion soft enough that nobody could argue with it. Nobody checks. Nobody has time to check. And the piece would have looked exactly like every other piece on the front page.

In 2026 I was twenty-three, an intern at a sports analytics firm in Liverpool, assigned to log the entire knockout phase of the World Cup in Russia. On 1 July, Spain played Russia at Luzhniki. Spain had 71.4% possession and made 1,029 passes, and I predicted they would win. They lost 3-4 on penalties after a 1-1 draw over 120 minutes.

I sat with that data for a week. What I found was not a small error. Across 120 minutes, Spain generated 0.9 expected goals. Nine-tenths of a goal, from one thousand and twenty-nine passes. On average, each pass contributed less than one-thousandth of a goal. Those passes were real, they were accurate, they were replayed beautifully on television, and almost all of them went sideways.

From that day I began every piece with xG and actual chance creation instead of retelling how the game felt. And I learned something I have had to repeat to myself every month since: possession is what the opponent allows you to have, not what you take. When a side drops deep and seals the box, it is handing you the ball in the area it does not need to defend. You accept the gift, you keep it, and the stat sheet records you as the team controlling the match.

I do not trust a single number, but I trust the story it tells after I have interrogated it three times. The first pass is the raw value. The second is the context of the season, the surface, the rhythm of the tournament. The third is the hardest question: if I swap the subject, does this number still stand?

That empty document operates on the same principle, only one layer lower. In our system every conclusion must cite a numbered information point from the source article. No information points means no conclusions. No conclusions means no article. It is a technical contract, and the contract was honoured.

What bothered me was something else. The document was still handsome. It had section headers, tables, hierarchy, a comprehensive judgment, a five-star information value rating. It had the shape of an authoritative text. And in my trade, shape is often read instead of substance, because shape takes three seconds to recognise and substance takes thirty minutes to verify.

I have seen the consequences of that many times, and most clearly in the summer of 2026.

When the pandemic emptied the stands, I was working as a data analyst for a tactical consultancy. The Merseyside derby in June 2026 finished Liverpool 0-0 Everton. I compared Liverpool's PPDA — the number of opponent passes a side allows before committing a defensive action, lower being more aggressive — before and after crowds. The figure moved from 9.8 to 11.5. The forward line was pressing far less effectively, with the same opponent and the same system.

The home side's high-intensity running distance fell 4.3% in a crowdless environment. I wrote the report and argued that the crowd is a physical variable, not an emotional one. The empty stands taught me something brutal: noise never appears in a spreadsheet, but it is always present in every heartbeat.

But I have to state plainly the part I could not do. The 2026-20 season resumed after a three-month break, June weather is not December weather, the calendar was compressed to a density never seen before, and every team changed its rotation. I could not separate that 4.3% into a crowd component and a schedule component. I tried, and I failed. Error is the least likeable friend I have, but it is the only one in the meeting room that never lies to me.

The third lesson came from Leicester City in 2026. I was assigned to analyse their fifteen-match collapse after the FA Cup triumph. Seven centre-backs injured. Jonny Evans out for twelve matches. The side's expected goals conceded rose 24% over that stretch. The explanation available in the meeting room was bad luck, and I refused it — not out of stubbornness, but because bad luck is not a variable you can forecast.

I went into the centre-backs' running distances. An average of 8.2 km per match. But when two matches were fewer than 72 hours apart, that figure fell 12%. Not because they were lazy. Because they were no longer capable of running more. A centre-back running 12% less arrives half a second late, and half a second at that level is a conceded goal.

A chain of injuries is not a curse; it is a map that exposes the depth of a system being eroded. I proposed an index called expected injury load, and the firm adopted it. My work shifted from research to strategic advisory. But what I carried away was not the index. It was the principle: when a run of bad results appears, the first question must be the fixture list, not the identity of the person who made the mistake.

Those three stories share one structure, and that structure applies intact to tennis.

A player wins 78% of service points on grass and 62% on clay. Put those two figures side by side without recording the surface and you have manufactured a false paradox. A return-points-won rate dropping 5% after an indoor hard-court event does not mean the player has declined; it means the ball travels faster and reaction time is shorter. Rafael Nadal's fourteen Roland Garros titles only mean something attached to the words red clay; remove those words and the number becomes an unverifiable legend.

Break-point conversion is the most misread metric in tennis. A player converting 45% from a sample of twenty opportunities and a player converting 38% from a sample of one hundred do not belong to the same data class. But on the news board they sit side by side, in the same font size.

And here is where that empty document returns. Form is a short memory, and it took me years not to confuse it with substance. The form column in any match preview is usually built from the last three matches, unweighted for opponent quality, undifferentiated by home or away, undated. Those three matches may span two surfaces and a twelve-hour flight.

Hand a language model that nine-section frame and one player's name and it will fill it in four seconds. It will write about surface adaptability, about clutch points, about generational landscape, about injury risk, about the industry transmission chain. The prose will flow. The structure will be immaculate. And roughly thirty per cent of the content will be wrong, in a way nobody catches because no cell records a source.

That is the real risk. An empty document is rejected at the first gate. A fluent document gets cited, reposted, used as the basis for a scouting decision, and three months later nobody remembers where it came from.

My first instinct was to point at the analyst. That instinct was wrong. The analyst in this case is blameless, because there was nothing to analyse. But if I point at the tool instead, I am equally wrong. The tool did exactly what it was programmed to do.

The fracture is in the incentive structure. A pipeline is designed to always emit a document, because an empty document is treated as a process failure rather than as information about the source. Editors need a piece of a certain length. The system needs an output to pass downstream. Nobody is rewarded for blocking a piece that has no data; people are rewarded for delivering on time. And when the reward sits in the delivery format, the format gets delivered and the substance is left to sort itself out.

But I do not want to turn honesty into a shield. A null result is only worth something if it triggers a repair. If it is filed away, marked complete, and nobody re-runs the input stage, then it is just a politer way of producing nothing. Honesty without a feedback loop is silence with formatting.

And there is a limit I have to concede myself. I can demonstrate that a report with no data is worthless. I cannot demonstrate that a report with data is valuable. Those are different claims, and I once conflated them in a 2026 piece about a mid-table side's away run. I drew a fitness conclusion from running-distance data when what I had actually measured was that the team had faced three consecutive high-pressing opponents. Correlation is not causation, and I got it wrong. I corrected it publicly three weeks later, and that was the only time I wished I had done so sooner.

So what to watch in the next cycle.

Three signals. First, the title field and the source field must become mandatory and non-null, because the title is the only thing still analysable when the body goes missing. Second, a publication timestamp must sit in the input schema, because any form analysis without a date cannot be positioned on the season's axis. Third, there must be an automated guard that rejects any output with zero information points before it reaches the publishing queue.

Every match is a hypothesis. I only write the piece when I have enough data to disprove myself.

Next time you open a match preview with all nine sections, count how many cells actually contain a number that could be proven wrong. The number of falsifiable cells is the only real measure of an analysis. Everything else, however well presented, is the shape of an authoritative text waiting for someone to fill it in.