The Empty Pipeline: Lessons from Football Models That Had Nothing Left to Say

Three seventeen in the morning, Shanghai. I sat in front of three screens: two running data tables, one showing the match. By the seventy-eighth minute, the left screen — the one I trusted most, the one I had built over four years — showed only a single line: N/A. No red warning. No beep. My data pipeline had gone empty at some point, and it had not bothered to tell me.

I remember that feeling more than the match that night. A model does not collapse because it is wrong. It collapses because it goes silent. The numbers are still there, the charts are still there, the sense of control is still there — but inside, there is nothing left to interrogate.

The next morning, I opened the entire system log. Three weeks earlier, an injury data field had stopped updating. Two weeks earlier, a youth scouting feed had lost connection. One week earlier, a tournament variable had frozen. None of them threw an error. My model ran smoothly on an empty shell, and I read it like scripture.

To understand why this story matters more than a technical glitch, we need to go back to 2026, when I was thirty-five, working as a senior analyst for a young sports platform in China.

Before the Shanghai SIPG versus Shandong Luneng match on matchday 18 of the Chinese Super League, I published an analysis using the expected goals model — xG. The number I put out was shocking: SIPG had an xG of 2.8, their opponent only 0.4. I predicted 3-1. Every traditional expert in the newsroom picked a draw, because Shandong had a stable defensive run and a decent head-to-head record. The final score: 3-1. My article reached fifty thousand views in twenty-four hours.

That night I thought I held the key. I did not know that what I held was only one fragment of a much larger machine, and that the machine could stop feeding itself at any moment without making a sound.

At the 2026 World Cup, a betting company hired me as chief analyst. My new model rested on two variables: PPDA — passes allowed per defensive action — and the average height of the defensive line. It predicted South Korea to beat Germany 2-0 in the group stage. I posted a tweet urging everyone to bet accordingly. The result: 2-0 to South Korea, with Kim Young-gwon and Son Heung-min scoring late. Germany were eliminated in the group stage.

The whole internet called me a prophet. I believed it. That was the second mistake, and it cost more than the first.

In the round of sixteen, the model said Brazil would beat Belgium, because Brazil's defensive xG was better and Belgium's defence was undervalued in counter-attacking situations. I stated it live on air, in front of hundreds of thousands of viewers. Brazil lost 1-2. Fernandinho scored an own goal, Kevin De Bruyne struck from distance, and Renato Augusto managed only a consolation. Many clients lost money because they listened to me.

I argued fiercely online with a colleague who had warned that my model lacked a big-match psychology variable. Then I spent three weeks rewriting the code, adding a tournament variable and a random component. Those three weeks taught me more than the previous four years.

But it was only on that night in Shanghai, when the screen showed nothing but N/A, that I understood the root of the problem.

That incident was not about wrong data. It was about data disappearing, and disappearing silently. I call it the empty pipeline. A system can keep running, keep outputting numbers, keep drawing beautiful charts — while its input has long since run dry. And the most dangerous part: the reader cannot distinguish a full model from an empty one, because both present themselves in the same format.

In football analytics, we talk endlessly about wrong data, noisy data, biased data. Few talk about empty data. An injury field that stops updating does not produce a wrong value — it produces a gap, and the model fills that gap with old assumptions. A scouting feed that loses connection does not skew the number — it freezes the number, and the freeze is mistaken for stability. A frozen tournament variable does not break the formula — it makes the formula apply to the wrong context, and wrong context is the hardest error to detect.

I used to think the enemy of a model was randomness. That is wrong. The enemy of a model is silence. Randomness still leaves a trace — it produces error, and error is measurable. Silence leaves no trace, because it wears the coat of normality.

This is where we must distinguish two things the entire sports industry keeps conflating: data and evidence.

Data is raw material. Evidence is data that has been interrogated and still stands after being questioned. An xG of 2.8 is data. It only becomes evidence when you can answer: where did this number come from, how many matches were sampled, who recorded it, and does it still hold if the recorder changes. In my 2026 SIPG analysis, the 2.8 held up because I personally checked every situation, cross-referencing shot location and defensive pressure. Had the shot-location feed frozen that day, I would have had a beautiful number and an empty conclusion.

Modern sports analytics has been swept into a race for complexity. Everyone wants more variables, more layers, more machine learning. But every added variable is a new gateway for empty data to slip in. A model with thirty variables has thirty chances to be poisoned by silence. And because complex models are harder to verify, the person operating them often hides their ambiguity behind a technical performance.

All models are wrong, but a few are wrong in useful ways. I have said that for years. But only when my own pipeline went empty did I manage to add a second clause: a few are wrong in useful ways, while most simply stay silent in a harmless manner until someone loses money.

I want to go deeper into two matches that shaped how I see models: South Korea versus Germany and Brazil versus Belgium, both at the 2026 World Cup.

In South Korea versus Germany, my model was right. But later, looking back calmly, I had to admit something uncomfortable: being right does not mean being good. The model predicted based on PPDA and defensive height. South Korea won because Germany were complacent, because they pushed forward chasing a goal and were caught on the counter, because of a corner and a breakaway by Son Heung-min into an empty net. Among those causes, my model grasped only a small part. I was right for a different reason than the one I thought. That is the most dangerous kind of right, because it builds confidence in the wrong place.

In Brazil versus Belgium, my model was wrong. But wrong why? The model looked at defensive xG and concluded Brazil were stronger. In reality, Belgium executed one plan: concede possession, wait to counter, and punish Brazil with a set piece and a long shot. My model had no variable for tactical intent. It did not know that a team can deliberately choose weakness in order to win. It only measured what had happened, not what the team intended to do.

Two matches, one right and one wrong, yet both exposed the same flaw: my model measured the past and assumed the future would resemble it. Football does not work that way. xG does not score goals, but it makes people argue more than the actual ball does. And in those arguments, most of the time we are arguing about a number, not about the match.

After the 2026 World Cup, I rewrote the code, adding a tournament variable and a random component. But adding a random variable did not solve the root problem. It only made the model more modest in how it presented itself. What I actually needed was not more noise, but the ability to detect when it had nothing left to say.

A good model must have a mechanism to flag its own emptiness. When injury data stops updating, the model must downgrade its confidence, not hold it steady. When the scouting feed loses connection, the model must mark the gap, not fill it with an old value. When the tournament variable freezes, the model must stop applying it, not keep running. Sounds simple. But in reality, almost every sports analytics system I have touched lacks that mechanism, because it runs against the instinct of the profession: we want the model to always have an answer.

That is the trap of the trade. Clients pay to hear predictions. Editors need a number for the headline. Bosses need a chart for the presentation. Nobody pays for a line that says N/A. So when the data goes empty, the practitioner is pushed into a choice: admit they do not know, or present the empty as if it were full. Most choose the second. I once chose the second too, and that is why I am writing this.

There is a concept in search that I find applies to football analytics: information gain. An analysis has value only when it says something the reader did not already know. If it merely repeats the score, repeats the table, repeats what everyone can see, it has no information gain, however long it is. The empty pipeline is the enemy of information gain, because it makes the writer believe they are analysing when in fact they are decorating a gap.

Looking back over my career, I see I passed through four stages of belief.

The first stage, I believed data was truth. Right numbers meant right conclusions. This stage produced the 2026 SIPG analysis.

The second stage, I believed a more complex model would be more right. This stage produced the Brazil-Belgium failure and three weeks of rewriting code.

The third stage, I believed everything was random. This stage was the most dangerous, because it gave me licence to take no responsibility. When a prediction failed, I blamed noise. When an analysis was useless, I said football was inherently unpredictable.

The fourth stage, the present one, I believe data is a witness, not a judge. A witness can lie, can err, can fall silent, and my job is to interrogate it — not to worship it.

Data disappearing is not a loss of data — it is a type of data. The absence of a number is itself information. A player with no public injury data says something. A club that does not disclose recovery timelines is itself a signal. But to read that signal, the analyst must accept that silence does not equal normality.

I have watched a great many matches involving Chinese Super League clubs, and what caught my attention was not what was published, but what was left blank. A team winning three in a row but not disclosing the status of a key defender. A club spending big in the transfer window but silent about its wage structure. Those gaps often forecast more than the highlighted numbers. And when a model ignores the gaps, it places itself in an empty state without knowing.

Here I must address a trap I nearly fell into: using randomness as a shield.

Football stopped rolling in 2026, but randomness has never taken a lunch break. I love that line, and because I love it, I have misused it many times. After the pandemic, the schedule was scrambled, crowds vanished, and old models failed in unison. My colleagues and I drew a conclusion that sounded profound: everything is random, models are an illusion. That conclusion is partly true, but it became an excuse so that no one had to take responsibility for their predictions.

The problem is this: randomness is something to be proven, not something to be invoked. Before calling a result random, I must rule out the intervening variables. I must check whether the failure came from empty data, from a frozen variable, or from a genuine shift in context. If I have not ruled out a single variable, I am not allowed to use the word random. This is the discipline I set for myself after that night in Shanghai, and it is far harder than saying a line of philosophy that sounds clever.

At the same time, I realised I had to be careful of another trap: worshipping complexity to conceal ambiguity. Living between Vietnamese and Chinese football, and having changed research directions several times, I have a habit of seeing every problem from many sides. That habit is useful, but it also makes it easy to write multi-layered pieces that ultimately answer no question. An analysis with five variables swirling around it is usually a sign that the writer has not found the central question. With this piece, I forced myself to keep only one question: when data goes empty, what must a model do? Everything else is secondary.

In modern football, empty pipelines appear in places few expect. Look at the injury story. Clubs disclose only information that benefits their value. A minor injury is inflated into a major one to explain a defeat. A serious injury is hidden to protect a player's transfer-market price. Fans and media are deliberately kept in the dark, and that darkness seeps into every model that predicts line-ups. When I build a line-up prediction model, I do not lack algorithms. I lack the truth about who can still play. That is an empty pipeline built on purpose, and no algorithm compensates for it.

The same happens in youth development. Many former stars open academies, advertise heavily, but most are commercial operations tied to personal image rather than a structured training programme. Data on these academies is usually empty: no output figures, no coach-development pathway, no independent evaluation. An analytics platform relying on such sources will paint a beautiful picture of youth development with no basis. Investment in systematic grassroots coach education, the thing that truly creates long-term value, is left blank — a gap far larger than what the media cares about.

Looking across Asian football broadly, I see a paradox. Leagues have more data than ever, but data quality does not rise correspondingly. Positional data, running data, passing data are collected in bulk, yet few check whether they still hold in a new context. A model trained on last season's data may be entirely useless this season, because the squad has changed, the tactics have changed, and most importantly, opponents have learned to counter it. Data obsolescence is a silent form of empty pipeline: the data is still there, but it no longer describes reality.

Every spreadsheet is a meditation, except that when the meditation ends you have lost money. I learned this after years in front of a screen. The stillness of an analyst is not serenity, but focus on the right question. And the right question in this trade is rarely about the number. It is about the conditions that produced the number.

Back to that night in Shanghai. After discovering the empty pipeline, I did not rewrite the model immediately. I did something else: I listed every data source and marked the last update time of each. The result chilled me. Some sources had stopped updating three months earlier without my knowing. I had issued dozens of predictions in that period, each based on partially dead data. Not one of those predictions flagged itself as unreliable. They all looked normal.

Since then, I have added a mandatory step to the workflow: check pipeline health before analysing. Every data source must carry a last-update date, and if it exceeds the threshold, it is removed from the model rather than used with an old value. This is dry discipline, with nothing glamorous to show off, but it has saved me from more mistakes than any sophisticated algorithm.

I also learned to distinguish two kinds of silence. The first is technical silence: a broken feed, a severed connection, a frozen field. This can be fixed with checking discipline. The second is strategic silence: someone deliberately hiding information because it is unfavourable to them. This is more dangerous, because there is no technical error to detect. Dealing with it requires something else: reading the motive of the data provider. A club silent about an injury may be silent because it does not know, or because it knows and does not want to say. The two situations call for two different responses, and the skilled analyst is the one who can tell them apart.

This is the counter-intuitive part I want to stress, because it runs against the prevailing belief in the industry.

The prevailing belief is that the biggest problem in football analytics is a lack of data. In my view, the biggest problem is a surplus of dead data and a shortage of mechanisms to detect it. Sports platforms are racing to collect ever more metrics, but almost no one invests proportionally in checking whether the metrics are alive. The result is a paradox: the more data, the more easily a model is poisoned by silence, because there are more gateways for gaps to slip in.

The second counter-intuitive point: the accuracy of a prediction does not measure the quality of a model. The South Korea-Germany match gave me a correct prediction I did not deserve. Had I taken it as proof of my brilliance, I would have deceived myself. The only way to evaluate a model is to look at the process, not the result. But in betting and media, people remember only the result. This is why so many poor models survive so long: they got lucky at the right moment, and that luck was recorded as talent.

The third counter-intuitive point: humility can be measured. A model that knows when to say N/A is a more mature model than one that always has an answer. In my trade, this is seen as weakness. Clients want numbers, not hesitation. But a number without an emptiness warning is a promise the model cannot keep. I would rather give a prediction with a line admitting uncertainty than give a confident prediction based on dead data.

There is a cultural dimension worth noting here. Living between Vietnamese and Chinese football, I see the two football cultures treating data very differently, yet sharing one mistake. In one place, people worship numbers as talismans. In the other, people dismiss numbers because they believe football is emotion. Both extremes ignore the central question: is this number still alive? The difference in frames of reference does not make me write in a who-is-better style. It only shows me how data migrates from one culture to another, degrades along the way, and is often worshipped in the wrong place upon arrival.

To readers following the major tournament season, I want to say one blunt thing. During World Cups or major tournaments, the pressure to produce empty data is higher than usual. Media needs content every day. Pundits need to speak every hour. Models need predictions every match. No one has time to check whether the data source is still alive. As a result, during major tournaments, the share of predictions built on empty data spikes, and the failure rate rises with it. This is when readers most need to stay alert, not the moment to let emotion run loose with flags and stories.

I am not saying this to appear objective. I am saying it because I was once the man on live television declaring that Brazil would beat Belgium, and I know the feeling of drawing a firm conclusion from a model that was in fact empty. I do not want anyone to repeat that feeling, whether they are making the prediction or believing it.

So what can an ordinary football fan do with all this? No need to read source code. Just ask one question of every prediction you encounter: is this number based on live or dead data? If the person making the prediction cannot answer where their data comes from, chances are their model is empty. Confidence is not proof of quality. Sometimes it is only proof that people have not checked their own pipeline.

Back to the central question of this piece: when data goes empty, what must a model do?

My answer, after many years and many losses, is this: the model must fall silent. Not silent as surrender, but silent as honesty. A model that knows it is empty and says so is more useful than an empty model that believes itself full. In an industry where everyone wants an answer, the person brave enough to say they do not know is the most trustworthy.

I still run models every day. I still believe in data. But I no longer believe a number means anything by itself. Meaning comes from whether that number is alive, still describes reality, and still stands after interrogation. When a data source stops updating, I do not treat it as normal. I treat it as an event, a piece of information, something that must be told.

Randomness is still there, never taking a lunch break. But before blaming everything on it, I must check my own pipeline. Because in many of the failures I have witnessed, the culprit was not the god of chance. The culprit was a line reading N/A that no one bothered to read.

This major tournament season, when you hear a pundit speak with cast-iron certainty, ask yourself one simple question: is their pipeline still alive, or are they reading you a model that died long ago without their knowing? That question needs no algorithm to answer. It needs only a little doubt in the right place — something modern football, with all its data, is increasingly short of.