When the Data Sheet Is Empty: The Analyst's Discipline in an Under-Sampled Match

At three in the morning in Shanghai, I opened an analysis file and found every cell empty.

Not the kind of empty that comes from forgetting to save. The file had a title, nine sections, and a proper table framework. But from the first cell to the last, the content repeated one sentence: insufficient information. No competition name, no team, no player, not a single number to hold on to. A deep analysis file that was structurally complete and completely empty in substance.

I sat still for about ten minutes. Then I did something the version of me from ten years ago would never have done: I closed the file and went to brew a pot of tea.

That is the first reflex I want to talk about, because it is the reflex most people in my trade will break within thirty seconds. When the data goes silent, the analyst's craft gets pushed down one of two roads: build a story to fill the gap, or stop and call the gap by its proper name. I have walked the first road wrong enough times to know how expensive it is.

The lesson begins with a four-nil win

In July 2026, I was twenty-five, working as a data editor for a young football site in Shanghai. The city club beat a bottom-table team four-nil. I wrote a piece praising the manager's pressing system, and I wrote it fast, with great confidence, because the scoreline had already told me the whole story.

The home side's PPDA in that match was 8.2. For anyone unfamiliar with the number: the lower the PPDA, the fewer passes a team allows its opponent before closing in. A figure of 8.2 sounds like proof of organised aggression. But I had not checked one condition, the condition that determines what the number means: how the opponent played. The visitors sat deep and conceded the entire pitch, which meant the home side's pressing happened in a space where nobody was resisting. The metric looked beautiful, but the tactical substance had never been tested.

Three days later, the same club lost one-two to another bottom-table side. When the opponent no longer sat deep, when they stretched the pitch and played through the first pressing line with long balls, the system collapsed exactly where I had not bothered to look.

My editor said something I still remember word for word: you looked at the scoreline, you did not look at the structure. He was not being unfair. That four-nil win was beautiful in the results table and utterly ordinary in the structure table, and I had chosen to read the easier table.

From that summer on, every tactical piece I wrote had to carry a checklist of at least three advanced metrics and a mandatory line stating the match conditions. What I learned was not the list of metrics. What I learned was this: clean data without context is worse than ugly data with context, because it persuades us to go wrong faster.

The night in Russia and the limits of a confident number

A year later, I was sent to Russia as a data reporter for a World Cup. The quarter-final between the host nation and Croatia was the night that forced me to rewrite my entire frame of reference.

Our model put xG in Croatia's favour, roughly two point four against one point one. I had almost locked in the conclusion before kickoff. The match finished two-two after one hundred and twenty minutes, went to penalties, and the hosts went out there.

What kept me awake was not the result. What kept me awake was that the gap between what the model saw and what happened on the pitch was not a small error margin. It was a whole system of variables my model did not count: the fitness of a side forced to grind through one hundred and twenty minutes, the weight of home advantage as a psychological counterforce, and the way an underrated team finds an extra source of energy when it is desperate.

After that night I stayed up watching fourteen knockout matches again. I found that nine of the fourteen produced results that diverged from xG once I added two variables: the minutes played by key players and the distance covered after the seventieth minute. The piece I wrote afterwards I titled the xG trap, because the real problem is not that xG is wrong. The real problem is that we forget every metric is measured under a specific condition, and when the condition changes, the number stays the same on paper while its meaning changes on the pitch.

My writing since then has revolved around a single question for every number: under what conditions was this metric measured, and are those conditions the same as the match I am discussing. If the answer is no, I am not allowed to use it as evidence, only as a hypothesis.

That night in Moscow, I did not watch a football match; I watched raw data laugh in the face of every probability. I keep that sentence unchanged, because it is the most accurate professional memory I own.

The theatre without spectators and the layers of data that are not numbers

By the pandemic season, I was running a data content team for a sports platform. Global football stopped, then returned inside stadiums with no people.

Over six months I rewatched more than a hundred old matches with tracking data. Early in the 2026-21 Bundesliga season, I found a paradox: teams pressed roughly twelve percent more often, but the effectiveness of that pressing fell by around eight percent. The laziest explanation is that teams simply played better. The explanation I trust more is that with no stands, a familiar source of psychological pressure was withdrawn, and a pressing system that relied on opponents being squeezed by the roar of ten thousand people became less frightening.

This was the first time I saw clearly that a football match contains two layers of data. The first layer is what can be measured: passes, recoveries, distance covered. The second layer is what can only be read from breathing, from a stride that shortens in the seventy-fifth minute, from a player standing in the wrong position because the mind tired before the legs.

Many of my colleagues treat the second layer as noise. They are not entirely wrong, because the second layer is very hard to encode and very easy to invent. But if you discard it completely, you will never explain why a higher metric attaches to a worse result.

Russia taught me that the variable does not live in the spreadsheet, it lives in the player's pulse. I wrote that line after the trip, and it only became more convincing when I watched beautiful pressing systems crumble in silence.

I did something I still defend. I was assigned to build an automated writing system for matches with empty stands, and I refused. The reason is that I believed an automated system would recreate exactly the mistake I made in 2026: it would read the beautiful number and ignore the context. I proposed a model combining tracking data with remote interviews of coaching staff, and what I got back, beyond a small promotion, was a habit my colleagues found annoying: every piece I wrote from then on had to include a small section titled data collection method.

The goal in the one hundred and nineteenth minute

Euro 2026 gave me the reverse test. In the quarter-final, hosts Spain faced Germany. In the one hundred and nineteenth minute, from a cross by Olmo, Merino headed in the decisive goal.

Our probability model placed that situation at roughly one point two percent. One point two percent means that in the machine's eyes, that moment almost should not have happened. And yet it happened, and it decided an entire quarter-final.

I wrote the piece that same night. I pulled the data on that team's successful headers over their last fifty matches and found a striking ratio: nearly one fifth of them came from positions the model considered impossible to score from. In other words, those low-probability situations are a systematic part of the team's skill set, not a statistical accident.

That was when I stopped using the word AI as an answer. Our model was not wrong to say one point two percent. It was simply honest within its own scope, and that scope did not include Merino arriving in exactly the right place with exactly the jump his collective had rehearsed hundreds of times. A probability model needs to be challenged by people, which is very different from being replaced by them.

The contrarian angle: an empty space is not a failure

There is a way of thinking I deliberately take against the majority in my trade.

When an analysis file returns all empty cells, the amateur reaction is to treat it as a technical incident, something to be patched immediately with more data, more models, more inference. The professional reaction is to treat it as a valid result. Insufficient information is a conclusion. It sounds harder to swallow than a tactical judgement, but it is more honest than any tactical judgement built on sand.

The danger lies in a very old psychological fact: the human brain hates a vacuum. Give someone a post-match story and they feel comfortable. Give them an empty table and they feel uneasy, and they tend to build a story out of thin air. In my trade, that behaviour has a name: storytelling instead of evidence. It is dangerous because it does not look like deception. It looks like work.

And this is where correlation stands far from causation. A high-pressing team wins three in a row, and people rush to conclude that high pressing is the cause. But perhaps the high pressing is only coinciding with an easy run of fixtures, and when the fixtures get harder, that same system breaks. I have been inside that trap. Three wins are not three samples. One win is not one sample. And a file full of words that are all speculation is not a sample either.

There is a detail in this trade I rarely say out loud. Emotion is not the enemy of data. In the identity of someone who writes with numbers, it is easy to develop a habit of treating players' emotions as noise to be discarded. I used to think that way, and I was wrong in many matches. The right approach is not to sweep emotion off the table, but to encode it as behavioural indicators: a stride that shortens in the seventy-fifth minute, a touch heavier than usual, a passing sequence drifting steadily toward safety. Once encoded, emotion becomes a valid data column, not a subjective remark.

The method I use when the data says nothing

I turned that three-in-the-morning night into a process.

The first step is to separate two columns clearly: hypothesis and evidence. The hypothesis is allowed to fly far. The evidence is not, and it must never be bent to fit the hypothesis. With a file of empty cells, my hypothesis is that the match does not yet have enough samples to be read. The evidence sits exactly there, no more.

The second step is to identify the smallest sample that still says something. Instead of dumping the entire season's data into one place, I look for one match whose conditions are closest to the match I need to analyse, plus one metric. A small sample with the right conditions is more useful than a large sample with the wrong ones.

The third step is to leave the deviation intact and name it. When the data deviates from expectation, the reflex of someone who likes control is to round it off, to pick an angle of reading that makes it fit the mould. I force myself to leave it intact and call it a blank point. A blank point is not proof of weakness; it is proof of honesty.

The fourth step is to state clearly which metric is missing and why. A judgement missing fitness data for extra time must say it is missing that, so the reader knows where they stand on the map.

Clean data cannot rescue a dirty hypothesis

There is a line I use so often it has become my identity. The summer of 2026 was the most expensive tuition I ever paid to realise: clean data cannot rescue a dirty hypothesis.

I repeat it because recently I have seen many younger colleagues invest heavily in the cleanliness of data: good sources, good models, beautiful charts. But data does not live at the input. Data lives along the entire path from the opening question to the final conclusion. A perfect model running on a wrong hypothesis will still produce a perfect number and a wrong conclusion, and the worst part is that the beautiful number makes the wrong conclusion harder to challenge.

That is why I keep a habit I know tires other people. I always turn every analysis into a hypothesis test. I set a specific tactical hypothesis, then go looking for data from matches with similar conditions to confirm or refute it. If the data refutes it, I rewrite the hypothesis, not the data. That two-column table is my most important defensive weapon, and it takes ten seconds to set up.

A system does not collapse in one night; it cracks from the moment I stop questioning the foundation. The first crack is usually silent. It is a number I accept without checking its conditions, a match I praise for the scoreline, a metric I treat as truth. People only hear the crash when the whole block has separated from the foundation, and by then it is too late to repair.

When the stands are empty, I hear most clearly

I began my analytical career with my eyes. I learned to read tables of numbers. Then I learned to read footage. And by the pandemic season, I learned to listen.

When the stands are empty, I hear most clearly the sound of pressing footsteps under the darkness of the pandemic. With no roar to cover it, I can hear the rhythm of ten men pushing up together, hear the moment that rhythm slackens, and hear the breathing of the system before it breaks. The silence of the stands does not remove information. It filters out noise so the remaining information rises more clearly.

I think about this every time an analysis file returns empty cells. An empty stadium is not a stadium without a match. An empty data file is not a file without value. Both are telling me something; they simply are not speaking in the language I am most used to hearing.

Numbers only tell part of the story; the rest I hear with ears that were once burned by my own arrogance.

Looking forward

I am preparing for the next tracking cycle, and this time I am carrying a list of signals I will observe rather than conclusions I have preset.

I will watch the delay between a team losing the ball and that team beginning to reorganise, because that number speaks to structure while the scoreline speaks only to results. I will look at effective distance covered after the seventieth minute, something the naked eye cannot see but which decides who is still standing when the match drags on. And I will clearly log every empty cell in my own data, because an honest analysis file must say what it does not know, rather than pretend to know everything.

If there is one thing I want readers to take from these lines, it is a way of asking questions. Next time, when you read a judgement about a match and find it flowing, full of numbers and with no hesitation anywhere, ask yourself one thing only: is the writer telling you what he does not know.

Because an empty data sheet, in the right hands, is not an ending. It is the first coordinate of a new map.