The Empty Spreadsheet and the Four Times Data Spoke Against the Crowd

The Empty Spreadsheet and the Four Times Data Spoke Against the Crowd

At five in the afternoon on a November day in Shanghai, an assistant editor slid a spreadsheet across to me. The core-metrics column was blank. The starting-lineup column was blank. The context-notes column was blank: no note on whether the stadium was full or empty, no fixture-density note, no weather note, no pitch-condition note. He said the morning bulletin needed a long match-analysis piece, roughly fifteen hundred words, and asked whether I could finish it in two hours.

I opened the file, scrolled to the bottom, closed it. Then I gave an answer the me of ten years ago would never have dared to give: without data, I do not write.

He thought I was joking. At thirty-eight, after twenty-two years inside the sports and esports industry, I have heard every reason to write: audiences need the piece, partners need the piece, the editor needs the piece, and above all the algorithm needs the piece. But there is one thing the algorithm cannot give me, and cannot take away: a sample thick enough to say something verifiable.

That spreadsheet is the subject of this article. Four times in my career, I stood against the crowd because the spreadsheet contained numbers. Once, I stood against the crowd because the spreadsheet contained only emptiness — and I wrote anyway. That was the time I was wrong.

The data context, before anything else

I work as a sports data analyst in Shanghai, reporting on football and esports for the Chinese market. My job is not to make entertaining predictions. My job is to turn a match into a structure that can be checked again: what happened, how often it happened, and under what conditions.

There is a rule I set for myself in 2026 and have never broken: before any judgement goes into print, the piece must contain at least three metrics from three different families — chance creation (expected goals, shot volume, passes into dangerous areas), off-ball control (PPDA, recoveries in the opponent's final third), and physical output (distance covered, sprints above 25 km/h). Those three families must point in the same direction. If they contradict each other, I do not conclude — I record the contradiction and leave it there.

Three metrics are still not enough. After 2026 I added a mandatory section at the top of every piece: the data context. Whether the stadium had spectators, temperature and humidity, fixture density over the previous ten days, rest days between matches, and whether the team had travelled long distance. The same number, placed in two different contexts, produces two different conclusions. Covering 118 kilometres in a domestic league with seven days of rest is normal. Covering 118 kilometres three days after a six-hour flight is something else entirely.

I mention this because the piece below is not a summary of achievements. It is an audit of method. I want to retell the four times I used data to walk against what the stands, the newsroom, and an entire country believed — along with the one time I walked against my own data because I read it without reading the context.

On the night of the Shanghai derby, I chose the numbers instead of the whole city.

In 2026 I was twenty-nine, a mid-level editor at a new football platform in Shanghai. The derby between Shanghai Shenhua and Shanghai SIPG ended 2-1 to Shenhua. The whole city talked about fighting spirit. The desk asked me to write a piece praising the winners' character.

I opened the statistics table before I opened the text editor. SIPG took twenty shots, seven on target, and generated total expected goals of 2.8. Shenhua took five shots and generated 0.9 expected goals. SIPG's PPDA was 8.1 — meaning they recovered the ball once every 8.1 opponent passes, a figure in the best pressing band of the league. Shenhua allowed opponents 19.4 passes before recovering. I also pulled distance covered: the two teams were within 1.5 kilometres of each other.

Three metric families said the same thing: the losing side had controlled the match in the opponent's half for more than seventy minutes. The winning side scored two goals from under one expected goal. I wrote the piece concluding that this result belonged in the tail of the distribution — the kind of result that occurs regularly across a season but should never be used as evidence for anything about the quality of either team.

The reaction split cleanly in two. The first half was Shenhua supporters, who called me an intruder from a newsroom that did not understand Shanghai football. The second half was analysts and a few coaches, who sent private messages asking where I got the data. The piece launched my own column, Reading the Data, and from then on I had access to deeper databases.

What I learned was not in the scoreline. What I learned was this: when a result is repeated loudly enough in the dressing room, in the news, and in the stands, people begin to treat it as evidence. The spreadsheet has no room for that repetition. The spreadsheet only has sample and deviation.

In March 2026 I wrote a prophecy. The whole of Germany laughed.

In 2026, thanks to Reading the Data, I was sent to Russia as an analytical correspondent for the World Cup finals. Before the tournament began, I spent three weeks rewatching all ten of Germany's qualifying matches, noting every pressing action, every dropped block, every midfield turnover.

It came down to PPDA. That metric measures proactive engagement: the number of passes an opponent is allowed before the team recovers the ball. The lower the number, the higher the press. The leading pressing sides in Europe at that time ranged from 8.5 to 9.5. Germany averaged 11.3. That is not a small margin of error. It is the displacement of an entire system.

I added a second data point: shots Germany conceded per match in qualifying. Then a third: the number of times their central midfielders were bypassed in transition. All three metrics pointed the same way. I wrote a prediction that Germany would be eliminated in the group stage, for one reason only: they could no longer press in midfield, and in a short tournament that is an unfixable fault.

The reaction when the piece ran is hard to forget. A colleague called me a numerology monk. A well-known commentator said I was using statistics to insult a football nation with four world titles. I collected enough mockery to remember that I was standing alone.

On 27 June 2026, Germany lost 0-2 to South Korea in Kazan. Kim Young-gwon scored in the third minute of stoppage time, Son Heung-min in the sixth, after the German goalkeeper had gone forward. Germany finished bottom of Group F with three points and were eliminated. After that night, my piece was shared more than fifty thousand times.

I will be honest about how that felt: it felt good. But it was also the moment I began making a different error — believing that because I had been right once with the PPDA scale, the scale could be applied to every match. I built a tournament-prediction model, labelled high-risk teams as slow-fuse bombs, and published the list before every major tournament. Readers started waiting for that list. They said I was causing trouble. I was only reading the ending a few months early.

What I did not write into that model, and what nobody asked: the model had never been tested at a tournament with empty stands.

Summer 2026: when the stands went silent, the whole game transformed

In 2026 the pandemic stopped the leagues. The Bundesliga was the first major league to restart, in mid-May, with no spectators. At thirty-two I was already a senior specialist with access to detailed databases across several European leagues.

I collected data from 250 Bundesliga matches played after the restart and compared it against a control sample from the same league before the shutdown. Two metrics jumped out first. Home win rate fell from 43 percent to 31 percent. Average goals per match fell by 0.4. I checked a third metric to make sure this was not sample noise: average yellow cards per match fell slightly, while duels in central midfield increased.

Those three metrics told one unified story. Home advantage in football does not live in the grass or in travelling less. It lives in the stands: in the roar that makes a referee hesitate half a second before a decision, in the invisible pressure that makes the away side choose the safe pass instead of the line-breaking one. When the stands are empty, roughly a quarter of that advantage disappears.

I wrote a long study titled A Silent Stand Is a Metric. My editor asked me to add an optimistic closing paragraph about football's recovery. I refused, and added instead that if this data was correct, then any prediction model built on data from a season with crowds would fail in a season without them.

The study was cited by several Bundesliga coaches in press conferences. I also lost my private contract with the newsroom because of my rigidity. With no crowd, football transforms. I found that out — and was rejected for it.

But this time I did not only lose a contract. I learned something more important: context is not an appendix to data. Context is a variable. Take it out of the model and the model is no longer a model, only a pretty table of numbers.

The Euro 2026 semi-final: the time I was wrong, and where I went wrong

In July 2026, the European Championship was played a year late. I walked into the semi-final between England and Denmark with the highest confidence of my career, having just finished the empty-stadium study and believing I had captured the forgotten variable.

My model produced very clear data. Denmark averaged 118.7 kilometres per match; England 112.3. Denmark averaged 18 shots per match; England 11. Denmark recorded more recoveries in the opponent's final third. All three metric families — chance creation, off-ball control, physical output — pointed the same way.

I went on a radio station and said the data said England would lose. England won 2-1 after extra time. The internet used that clip to mock me for weeks.

Looking back with data, I missed three things, and none of them lived in the basic spreadsheet.

The first was squad depth. A team's average distance covered is a metric for the first and second halves, not for thirty minutes of extra time. When a match extends, the side with more quality options from the bench holds its intensity. England could introduce elite attacking players at the seventieth minute; Denmark had no equivalent option. Jack Grealish came on and completely changed the rhythm of the attacking line.

The second was the structure of the data. My model used a whole-tournament average. But Denmark had reached the semi-final after two extended matches and an emotionally extraordinary journey. An arithmetic mean cannot see accumulated fatigue.

The third, and the biggest error, was that I ignored the psychological variable of playing at home with a crowd. My 2026 study said empty stadiums remove home advantage. But that semi-final was played at Wembley with more than sixty thousand spectators, while throughout the preceding season English players had performed in near-empty grounds. I applied the conclusion of one environment to the opposite environment.

That was when I understood that the most serious error a data analyst can make is not miscalculating. It is calculating correctly on a sample that no longer applies.

Since then, every piece I write has two separate parts. The first is the numbers. The second is the reality check — where I state plainly what the numbers cannot see: fitness, motivation, dressing-room relationships, and pressure from outside the pitch. I also added a permanent closing section: Where might the assumptions be wrong?

The counter-intuitive angle: correlation is not causation, and esports is paying for it

There is a pattern I see repeated in both football and esports: a team wins, people hunt for the metric in which that team excelled, then turn that metric into the cause of the victory. But a metric advantage in a single match is largely a consequence of taking the lead, not its cause. A team two goals up will run less, press lower, and hold less of the ball. Reading a stats table without reading the timeline of the goals is reading the arrow of causation backwards.

In esports this problem is more severe because of the speed of change. Every patch makes the previous patch's data stale. A team with a high win rate on a particular character pool can lose that entire edge after one stat adjustment. And while data needs time to accumulate a sufficient sample, the betting market has already finished pricing it days earlier.

Transfers are a fertile gamble, but I count the cards before I place a bet.

What I have observed in the esports market over many years is the gap between the speed at which money moves and the speed at which regulation updates. Tournaments have competitive-integrity rulebooks, but those rulebooks were often written for a world where no money flowed through each group-stage match. When data on betting behaviour is not published, nobody can verify whether an anomalous result is random or non-random. And when verification is impossible, both the accusation and the defence are just belief.

From the Bundesliga to esports world finals, I look for the same thing: a fact that can be repeated.

But the counter-intuitive part is this: a repeatable fact is not necessarily a meaningful one. A metric can repeat stably across ten seasons and still say nothing about the specific match being played, if the conditions of that match sit outside the distribution of those ten seasons. That is the limit of every model, and the reason I never publish a prediction without its conditions of application attached.

At a deeper level, the sports data industry has a problem of position. Analysts are edging closer to the dressing room, sometimes making tactical recommendations directly. But their conclusions are usually drawn from aggregate data, while the actual rhythm of a match — who still has legs, who is losing composure, who has just argued with a teammate — lives in no spreadsheet. I have spoken with coaches and noticed one thing: they use data to confirm intuition, rarely to replace it. That is a healthy limit, but also a warning to anyone who thinks a model can replace a person.

Where might the assumptions be wrong?

I have to state this part, because without it the piece above is just a long self-congratulation.

Assumption one: I treat expected goals as a good measure of chance quality. That holds on average across a large sample but fails at the level of a single match, especially for teams with idiosyncratic finishing or teams that deliberately concede low-quality chances.

Assumption two: I treat PPDA as reflecting tactical intent. But a high PPDA can also come from a team protecting a lead and deliberately dropping its block. The metric measures behaviour, not intent.

Assumption three, and the one that worries me most: I treat what I observe in European football as transferable to esports. That holds at the level of logic — both are competitive systems with finite samples — but fails at the level of speed. Football changes over seasons. Esports changes over weeks.

And the final assumption: I treat myself as having learned the lesson from the Euro 2026 error. Perhaps I have not. A person who writes about data tends to repeat the same mistake in a new shape, and to notice only after publication.

The data context of this very article

To be honest, I will set out the environment of this piece. The Shanghai derby figures come from internal data I collected in 2026, a single-match sample, low confidence for comparing two teams, high confidence for describing one event. The Germany PPDA figures come from ten World Cup 2026 qualifying matches, a small sample, biased by mismatched qualifying opponents. The empty-stadium figures come from 250 Bundesliga matches in 2026-2026, a large sample in conditions that cannot be repeated. The Euro 2026 figures come from Denmark's seven matches and England's seven, a very small sample.

These four datasets are not the same grade of evidence. I place them side by side to tell a story about method, not to add them into one unified conclusion.

Signals for the next cycle

After many years I have settled on a working habit I consider the most honest available: always keep a list running in parallel with the prediction list. The first list is what I believe will happen, with probabilities. The second is the conditions under which, if they appear, the entire first list must be thrown away — an empty stadium, a congested schedule, a mid-season patch, or an injury in exactly the pivot position.

When I received that empty spreadsheet on that November afternoon, I applied exactly this rule. An empty dataset means the second list swallows the first entirely. There is no condition to test, so no conclusion may be published.

People usually assume the job of a numbers reader is to deliver answers faster than everyone else. I think the opposite. My job is to deliver answers more slowly, but with a trail that anyone can check. The spreadsheet is an altar, and I give myself to every number — including the numbers that say I have nothing to say yet.

Every crowd is wrong. The only thing that is not wrong is probability. And probability, when the sample is zero, is not a number. It is an emptiness that deserves respect.

The next round will begin with a question I have not answered: if data keeps multiplying, are we moving closer to the truth, or only learning to lie more elegantly with accurate numbers?