When the Model Returns Zero: Lessons from an Empty Spreadsheet

3:47 a.m. in Shanghai. Before the most important knockout match of the week, I opened the spreadsheet I know by heart, typed the command to run the model, and waited. On the screen, the first line came back blank. No xG. No PPDA. No chance-conversion metrics. Not a single number. Only the line that any data analyst fears more than a wrong prediction: insufficient information to assess.

I sat staring at that empty spreadsheet longer than I needed to. Twenty-eight years in the trade, eight World Cups, eight Olympic Games, thousands of matches through my hands, and this was the first time I realised I had never prepared for the simplest scenario of all: when the data says nothing at all.

People imagine my job is to sit before a screen and wait for the golden number to appear. The truth is the opposite. Most of the time, I am waiting for a refusal. A decent model is not one that always has an answer; it is one that knows how to stay silent when it knows nothing.

Context: when football believes data can never fall mute

Fifteen years ago, when I started bringing advanced metrics into analysis for the Chinese market, people looked at me like a man selling charms. xG was unfamiliar then. Anyone who mentioned expected goals was treated as someone dodging reality, someone afraid to say plainly which team was stronger.

Then the data revolution arrived. Cameras tracking every footstep. Sensors inside the ball. Hundreds of thousands of data points per match. Clubs hiring whole analytics departments. Betting platforms building machine-learning models. And a new belief was born, no less dangerous than the old one: that data always has an answer, if only you gather enough of it and make it clever enough.

That belief is half right. The right half: data shows us what the naked eye misses. The wrong half: data, however abundant, still has gaps that cannot be filled. And my trade, for years, has been the trade of people who live inside those gaps without admitting it.

I remember 2026, when I worked as a senior expert for a new sports platform. Before Shanghai SIPG met Shandong Luneng on matchday 18 of the Chinese Super League, I published an xG-based analysis: SIPG had an xG of 2.8 against their opponents' 0.4. I predicted a 3-1 win, while the traditional pundits all picked a draw. The final score was 3-1. The piece drew 50,000 views within 24 hours.

But I do not tell that story to boast. I tell it to say that the success itself taught me the opposite lesson. Because right after, I abandoned the series to chase a basketball betting model, then abandoned that too, then changed direction again. My editor was furious. But what unsettled me more than his anger was a question: if my xG model was right, was it right because it was good, or because I was lucky?

That is the question I carried for years, from Belgrade to Shanghai, from small newsrooms to packed stadiums. And it has only grown heavier, never lighter, each time I see someone too confident about a single number.

The core: an anatomy of a null result

This is where I want to linger longest, because it is what almost no one writes about.

A model returning a null result is not the same as a model returning a neutral result. These are two entirely different things, and confusing them is quietly poisoning the whole football-analysis industry.

A neutral result means: I have data, I analyse it, and the conclusion tilts toward balance — two even teams, a high chance of a draw. That is still an answer. A null result means: I do not have enough data to say anything at all, not even that the two teams are level.

In my spreadsheet that night, every cell was empty. No match name. No team names. No information on squad, form, or head-to-head history. Not one data point to hold on to. That was not a balanced result. It was a result that did not exist.

And here is the strange thing about my trade: a null result is far more awkward than a wrong one, because it gives you nothing to fix.

When a model predicts wrongly, you can open the source code, add variables, adjust weights, re-test. You have an object to argue with. But when a model returns zero, you have nothing at all. Nothing to fix, nothing to defend, nothing to sell. You have only silence.

I have been through worse. At the 2026 World Cup, after a model built on PPDA and defensive height correctly predicted South Korea beating Germany 2-0 — the match in which Son Heung-min sealed the win — I tweeted urging people to bet along. I was right, and I believed in myself.

Then came the round of 16. My model believed Brazil would beat Belgium, because Brazil's defensive xG was better. I said so live on air. Brazil lost 1-2, with Kevin De Bruyne scoring a goal my model had never accounted for. Many clients lost money because they listened to me.

I argued fiercely on social media with a colleague. Then I spent three weeks rewriting the code, adding tournament variables and a randomness factor. But in those three weeks I learned something no spreadsheet could teach: my confidence was not an indicator of the model's quality; it was an indicator of my ego's.

The null result that night, then, was not a technical glitch. It was an ethics test.

Because this is what my industry does not want to say out loud: when a model returns zero, the greatest pressure is not to fix the model. The greatest pressure is to invent an answer.

You have a deadline. You have an editor waiting. You have a newsroom needing 800 words for the pre-match column. And you have an empty spreadsheet. The first option, honest but unsellable: write that you do not know. The second, seductive and sellable: gather a few numbers from elsewhere, stitch them into a plausible-sounding story, and call it analysis.

I have seen the second option chosen thousands of times. I have chosen it a few times myself. And each time, I told myself I was merely filling gaps with experience. But experience is not data. Experience is a bias dressed up in years.

Let me put this in the language of the trade. In statistics there is a concept called missing data. And there is an iron rule: how you handle missing data determines your result more than the data you have. If you replace missing data with the mean, you are silently declaring that the unknown resembles the average. If you replace it with zero, you are declaring that the unknown is nothingness. Both are lies disguised as method.

And in football, missing data is not the exception. It is the rule. A player is injured — you do not know how far he has recovered, because the club only releases information favourable to its share price. A manager changes tactics — you do not know what he is hiding for the big match. A national team gathers for three days — you do not know the chemistry in the dressing room.

I once sat beside a young analyst at a major tournament. He had a beautiful model, compact, running smoothly. Before the match, he asked what I thought of his prediction. I looked at the spreadsheet and saw a hole: the model had no variable for an unannounced starting lineup. I asked how he handled it. He said: I use the projected lineup. That was the moment I understood the problem. A projected lineup is an assumption, but it had been fed into the model as a fact, and the output would carry the precision of a carefully computed number.

Yet every day, hundreds of analyses are published in the tone of someone holding absolute answers. I read them and ask myself: where do these numbers come from? The answer is usually: from another model also filling gaps, from another analyst also inventing certainty.

This is why I say: data disappearing is not data loss — it is a kind of data. The absence of information, if you are willing to read it, is a message too. A club hiding its star's injury before a derby? That is data. A manager refusing to answer a question about the lineup? That is data. A model returning zero because its sources collapsed? That too is data — data about the very system trying to sell you certainty.

There is a subtler trap I once fell into: worshipping complexity to disguise vagueness. When I do not understand something, the reflex of a man living between two cultures is to add layers, add angles, add variables. More layers make you harder to challenge — but also further from the actual match. I learned that each piece should carry one central question. If more than three variables circle it, I must split it or say plainly: I have not solved this part.

And here the Vietnam-China gap I live inside becomes a lens, not a lectern. Data migrates between the two markets in very different ways. In China, people are addicted to absolute metrics, believing a bigger number is a truer one. In Vietnam, people still favour the eye and the feel of the match, sometimes distrusting any number from outside. Both extremes are ways of hiding. The number-worshipper forgets that numbers are made by people. The eye-worshipper forgets that eyes deceive too. And neither will learn to say: not enough information.

The contrarian angle: those who hide inside the word random

But I must argue against myself, because that is the duty of a man who has watched too many league tables collapse.

There is a temptation opposite to inventing answers: the temptation to hide inside the word random. When the model fails, instead of admitting a lack of data, we blame chance. When a prediction misses, we say: football is like that, there is always an element of surprise.

I believe football stopped rolling the old way in 2026. I believe randomness never takes a lunch break. But that belief easily becomes a shield. Every failure is blamed on noise, and so no one has to be responsible for anything.

That is why I set myself a rule: every time I write the word random, I must ask how many intervening variables I have ruled out. If I have ruled out none, I am not allowed to use the word. Because random is not an explanation. It is the name we give to what we do not yet understand.

And here is the most beautiful paradox of this trade: the null result and randomness are two faces of the same coin — both are admissions that we do not control what we think we control. The honest person recognises it and stays silent. The arrogant one fills it with story.

That night, before the empty spreadsheet, I could have chosen to fill it. I had enough material to invent a very convincing analysis: a few form numbers, a little head-to-head history, a firm concluding line. No one could verify it. No one could challenge it. And perhaps that piece would have been shared more than this honest one.

But if I had done that, I would have betrayed the very line I keep hanging on my office wall: all models are wrong, but a few are wrong in a useful way.

The takeaway: learning to say I do not know

People say I am good at predicting. Wrong. I am only good at saying I do not know at the right moment.

That is the hardest skill in this trade, harder than building a model, harder than reading xG. Because it demands you stand before a newsroom waiting for copy, before a client waiting for a prediction, before a reader waiting for an answer, and say: not enough information to assess.

The null result that night taught me something twenty-eight years had not. A good model is not one that always has an answer. A good model is one that knows its own boundary. And that boundary, in football as in everything else, usually lies exactly where we most want to cross it.

I still keep that empty spreadsheet on my machine. I do not delete it. Whenever I grow too confident about a number, I open it and look. It reminds me: every spreadsheet is a meditation, except that when you finish, you lose money.

The match that night went ahead anyway. I did not write a prediction. I just watched, for the first time in years, as a pure spectator, with no model in my head, no number to defend. And when the match ended, I understood that what I had just done was not abandoning data. It was the only way to respect data: to know when it is silent, and to be silent with it.

Perhaps that is the signal I will track in the next round. Not a team, not a player, but a question: in how many pre-match analyses does anyone dare to write two words — not known?