A Full Spreadsheet With Nothing In It: The Silent Bug Spreading Through Football Analysis

02:47 in the morning. Saigon. Rain. I open the match report that the data collection system has sent back. Nine sections. Every section carries a heading, a frame, a status line, a cell waiting for a figure. Nothing is blank. Nothing is flagged. And not one section contains a single usable piece of information.

Article title: unidentified. Source: unidentified. Article type: unclassified. List of information points: empty. Time sensitivity: not assessed. The only field left standing is the domain label: football. That is everything I received — a record that passed every structural check, complete in form, hollow in content, drifting onward into the analysis layer behind it as though it had just proved something.

What kept me at the screen for another forty minutes was not the technical fault. It was its position in the chain. A file like that, if I did not stop it, would be processed further. It would be scored. It would be ranked. It would be quoted in another bulletin, and then another, until the original emptiness was wearing the coat of a professional conclusion.

Vietnamese football runs on a great many templates, and almost nobody checks which of them are empty.

Context: a template can always be filled with zero

I entered the trade in 2026, after graduating from a journalism academy, writing for a football newspaper in my mid-twenties, then spending a few years in Madrid covering European football for an international sports title. It took many years before I understood that the hardest part of analysis is not getting the calculation right. It is saying out loud where you do not know.

My current workflow splits every football fact into nine dimensions. Tactics and technique. Club finance and the transfer market. Results and the opinion cycle. League landscape and team positioning. Rules and governance. Management and the dressing room. Risk profile. Media and expectation. And finally industry transmission. Nine dimensions, each with its own table, its own comparison row, its own warning flags, its own conclusions. Outsiders look at it and assume it is an automated machine. Not quite. It is a ritual, and a ritual can be performed with nobody actually present.

The problem sits in one very small design detail: a template can always be filled with zero and remain valid. “Insufficient information, cannot assess” is a structurally correct answer, and because it is correct it is dangerous. Once every cell is filled, the reader's eye stops distinguishing conclusions from blanks.

Based on my experience watching matches in the V-League, this happens every week. A club receives a ten-page opponent report, every heading present, every chart drawn, yet the section on the opponent's weaknesses repeats only what any television viewer already saw. A transfer item arrives with name, age, position and estimated value filled in, and not one line of verifiable sourcing. A betting preview carries probabilities, prices and a verdict, with no model standing behind the probability at all. To me these are the same defect. The only difference is that it never goes red.

The xG shock at Hang Day

In 2026 I watched Hanoi FC host Quang Nam FC at Hang Day Stadium. That night cost me 180 million dong. Hanoi FC, with Nguyen Van Quyet in the side, took 17 shots and their xG reached 2.87. Quang Nam FC managed exactly two attempts, xG 0.94, and the match ended 1-1.

The usual response is to blame luck. I did something else: I sat down and went through every shot of 112 V-League matches from round one to round fourteen that season, calculating xG by hand, attempt by attempt. The result: Hanoi FC were creating chances at a top-of-the-table rate but converting 23 percent below the league average. That was a signal, not a verdict.

The xG shock at Hang Day turned me from a man who watched football into a man who read data.

The three-thousand-word analysis I published at the time was laughed at by the media. A month later, that same dataset called Hanoi FC's run of four straight defeats before it happened. I am not retelling this to praise myself. I am retelling it because its structure is identical to the structure of that empty report: a top layer packed with signal — scoreline, momentum, crowd noise — and a bottom layer with no evidence in it at all. The only difference is that in 2026 I was still curious enough to go looking for the bottom layer. This time the system went looking for me, and handed back a blank page, neatly framed.

Kazan does not take revenge

2026, the World Cup in Russia. Before the group stage I went through Germany's pressing data, the reigning champions since the final of 13 July 2026 at the Maracana. Their average distance covered had fallen 12.3 percent against four years earlier. Their PPDA — the passes an opponent is allowed before a defensive action — had risen from 8.2 to 11.7. That number says something very simple: Germany were letting opponents keep the ball longer before challenging.

I published a forecast that Germany would go out in the group stage. Hundreds of mocking replies arrived, most of them from people who never finished the methodology paragraph.

On the night of 27 June 2026 in Kazan, Germany lost 0-2 to South Korea. Kim Young-gwon scored in the 90+3rd minute, Son Heung-min in the 90+6th. Across the match Germany's xG was a mere 0.41, and their last six shots all struck a defender. Kazan does not take revenge; Kazan simply keeps the books and waits for me to get my arithmetic wrong.

But here is the part I want to keep. After Kazan I was not pleased. I realised I had been right using a model whose input data I had never verified: the distance figures came from two different providers, measured under different definitions, and I had stitched them together without reconciliation. I was right because my conclusion was coarse enough to survive the error. Had I concluded anything more precise, the error would have shown.

A correct model built on unverified input data is still just a handsome report.

Empty stadiums and the price of a coefficient

In May 2026, with world football frozen, the Bundesliga returned on 16 May in empty grounds. I checked the first 28 matches back: the home side won only five, 17.8 percent, against a historical home-win rate around 42 percent for the league.

My betting model at the time multiplied a home factor of 1.32. In one week I lost 40 million dong. I immediately went back through 200 Bundesliga matches from that season and found the fracture: home teams still pushed high the way they did with a crowd, but their real xG fell 0.45 per match. No roar behind the goal, no invisible pressure from four sides, and the home advantage every model treats as a constant turned out to be a variable dependent on the crowd. Within 72 hours I wrote the piece “Home is no longer an advantage” and rebuilt the whole system, adding what I call the context coefficient: xG, PPDA and outcome probabilities adjusted for empty stands, weather and travel distance.

What I carried out of that week was not a new model. It was a new habit: before trusting any coefficient, I have to know the conditions in which it was measured.

Three fractures, one common denominator

Those three stories differ in scale: one V-League match, one World Cup, one Bundesliga week. They share one shape. In all three, the top layer — scoreline, spreadsheets, public opinion, report template — was full. And in all three, the bottom layer — traceable evidence — was the thing I nearly failed to check.

The 02:47 report is the extreme version of that shape. It does not pretend to hold conclusions. It prints one sentence per dimension: insufficient information, cannot assess. Technically, that is the most honest thing a system can do. Set beside the hundreds of other reports circulating in the market — reports that always have a conclusion, always a score, always a forecast — that honesty becomes the exception, and exceptions get filtered out of the workflow.

This is where I have to be clear about my own trade. For many years I sold certainty. Not because I believed in it, but because the market bought it. A client paying for a forecast with a clear price will not be satisfied with cannot yet be assessed, more fitness data required. Belief is a nuisance variable; run the emotional regression before you place the bet. I had to write that line on paper and stick it to the wall, because I was the first person to forget it.

The counterintuitive angle: the fault is not in the machine

The first reaction most people have to this story is to blame technology. A system that returns an empty report is a broken system. Technically true, causally wrong.

The cause sits in the incentive structure. Nobody pays for I do not know. A scout who files a report stating insufficient data to conclude on this player is marked down as incompetent, while the one who files ten pages of opinions — most of them untraceable to any source — is praised as professional. Rewards flow toward completeness, not accuracy. And when rewards flow toward completeness, systems learn to manufacture completeness.

That is also why I distrust simple explanations for defeats. A team loses control of midfield, and the verdict is that the midfield is weak. Correlation is not causation. A V-League club receives an empty report, makes a bad decision, and people say the report was poor. Possibly. But both may be symptoms of the same way of working, one in which the verification step does not exist.

There is one more class of error I have to name: a risk profile that cannot be rated is not a low-risk profile. The absence of evidence of risk is an absence of information, not evidence of the absence of risk. I have seen analysis rooms present boards with a clean risk table, every cell green, where every green cell was really an empty cell that had been coloured in.

The most frightening layer of analysis is not the one that calculates wrongly. It is the one that never calculated, and still presents as though it did.

What I carry into the next round

Out of that incident I set a hard gate for my own workflow. No analysis layer is allowed to run unless the input layer contains at least one information point, one named entity and one time anchor. When the gate fails, the system does not return a report. It returns a single status code: insufficient data. Ugly, blunt, but readable.

And I have started tracking a new signal week to week: the share of reports that are field-complete but content-empty. If that rate is a fraction of one percent in a batch, the problem is one stray article. If it crosses one percent, the problem is in the collection layer, and every ranking, every score and every aggregate built on top of it must be marked unverified.

The next round's job is not to ask which model is right. The next round's job is to ask: what share of the conclusions in my hands can be traced back to a source, a date, an entity, a calculation? The share that cannot be traced should be deleted from the report, not bolded in it. The crowd leaves, the model breaks, and I learn to listen to the breathing of an empty stand.