Ink Traces on the Statsheet: Four Tournaments, One Data-Tracing Method

Kazan, June 27, 2026. The clock at Kazan Arena ticked into the 90th minute plus three. Kim Young-gwon stabbed the ball past Manuel Neuer from a corner, the assistant referee raised his flag for offside, the Korean end of the stadium stopped breathing, and then VAR intervened and overturned the call. Three minutes later Son Heung-min ran the length of the pitch while Neuer was stranded in the opposition half. The score finished 2-0, and the defending champions left the tournament at the group stage.

Twelve hours later I opened the official statsheet. Germany dominated possession, took far more shots, completed more passes and completed them more accurately. South Korea were described in the usual two words: defensive and counter-attacking. That reading was right about the result and wrong about the cause.

In my own spreadsheet I pulled out one number: South Korea's PPDA in that match was 9.8. The metric measures how many passes the opponent completes for every defensive action a team makes in the opponent's half; the lower the figure, the more aggressively a team presses high. The average across the 2026 World Cup group stage sat between 12 and 13. South Korea did not park a bus in front of Neuer. They split Germany's midfield lines, forced Mats Hummels and Toni Kroos to circulate the ball sideways more than they wanted, and waited for the single moment to break.

PPDA 9.8 is not defending — it is how a team declares war with a number.

The piece I wrote then predicted Germany would be eliminated for one reason: their expected-goal differential was too thin for the volume of chances they created. The collapse of a giant always starts with a fragile xG. The article drew around 40,000 views, along with a few dozen angry replies suggesting that a fourteen-year-old had no business talking about the reigning champions that way.

My work began a year earlier, with a much smaller discrepancy.

July 12, 2026, K League 2, Busan IPark at home to Seoul E-Land. I sat in front of a screen with a squared notebook, ruling four columns: minute, passer, receiver, outcome. Three and a half hours later I had counted 412 completed passes for Busan. The official statsheet published 389.

Twenty-three passes of difference. Not large. But it forced me to ask a question that later became a working principle: which definition produced that number?

Four hundred and twelve passes, and the official figure is a polite lie — that is the line I wrote at thirteen, posted on a small forum, and received just enough argument to understand I was missing half the truth. Years later I still use the line, but with a clause my thirteen-year-old self lacked the patience to write: data providers do not lie, they simply answer a different question from the one I asked. Does a pass deflected by a defender's leg still count for the original passer? Is a blocked cross still a pass? Is a delivery from a set piece inside the sample?

Every pass leaves an ink trace if you are willing to follow it. But a trace only means something when you know which pen wrote it.

I kept that habit for years: archiving the raw data of nearly fifty matches, counting by hand, cross-checking one match at a time. My method grew out of that and still has three fixed layers.

The first layer is raw event data: every pass, every duel, every shot, with coordinates and timestamps. This layer is almost impossible to argue with once you accept the recording convention.

The second layer is definition. This is where most arguments on social media happen without anyone realising, because the two sides are comparing two numbers born from two different rulebooks.

The third layer is context: the competition format, the state of the match at the moment the data was generated, the schedule, the weather, and whether the stands had anyone in them.

Remove the third layer and you have a correct number that means nothing. Remove the second and you have an argument with no end.

Across the nearly fifty matches I counted, the gap between my tally and the published figure ranged from 2 to 9 per cent depending on the provider and the event type. Passes were the most volatile category, because they depend on who touched the ball last. Shots were the most stable. The lesson was not that official numbers are wrong, but that every event type carries its own margin, and an analyst has to know that margin before entering a debate.

Back to Kazan. The Germany–South Korea match became the first case in my file because it exposed the limits of possession.

Germany held the ball and passed the ball, but their passing chains mostly ran sideways and backwards. When I mapped the passes, most of the traffic sat in the middle third and the wide channels, with very few line-breaking balls into the fourteen metres in front of goal. Their shot count was high, but the distribution of shooting positions drifted further from the box with each group-stage match.

Volume of chances did not match quality of chances. Germany's expected-goal differential was thin enough that a single inefficient finishing night could bring the whole structure down.

A team can win the possession column while having already lost the chance-quality column before the ball was kicked.

South Korea, by contrast, did not defend by dropping deep. They defended by compressing the distance between their lines, accepting risk in the wide channels to seal the middle, and turning every German turnover in the Korean half into a counter-attack within five seconds. PPDA 9.8 reflects exactly that structure.

This is where I have to argue against myself, and I did so inside the 2026 article itself: PPDA depends on match state. South Korea entered the final group game needing a win. The later the match went, the higher they had to push, and the pressing figure naturally fell. Part of the 9.8 came from deliberate tactics; another part came from forced circumstance. Quoting the number while dropping the second half of that sentence would turn a description into a declaration.

That is also why I never conclude from a single metric. Possession tells you who held the ball. Expected goals tells you who created danger. PPDA tells you who imposed the tempo. The pass map tells you where that tempo was deployed. All four have to agree before a conclusion stands, and when they disagree, the disagreement is the story.

Two years later the pandemic shut the stadiums and handed me a natural experiment no league could have built ethically.

In May and June 2026, the Bundesliga returned in silence. I took the 2026-20 season data, split it into the period before and after crowds disappeared, and compared each club's home expected-goal differential.

For Borussia Mönchengladbach, the differential had been +6.2. With the stands empty it fell to -1.8. The decline in home advantage I calculated for that club was roughly 28 per cent.

Home advantage is not atmosphere; it is a number that knows how to evaporate.

The crowd leaves the stands, and the home equation loses its largest variable.

I have to be explicit about the limits of that calculation. The sample covered only a handful of matchdays, against different opponents, on a schedule that was not randomly assigned. More importantly, not every club declined at the same rate. Some barely moved at all. The variation between clubs is the valuable part of the information, not the league-wide average. Reporting only “home advantage fell 28 per cent” would have been a correct number that concealed everything interesting.

The hypothesis I put forward then: most of home advantage does not live in the noise of the crowd, but in the decision tempo of referees in 50-50 duels, and in the confidence level of the home side in the opening fifteen minutes. With the stands empty, both lose part of their anchor. That hypothesis was never tested tightly enough for me to call it a conclusion, and I will not call it one.

November 2026 brought a different problem, this time about injury.

On November 2, 2026, Son Heung-min suffered a facial injury and underwent surgery. He returned for South Korea's opening group match against Uruguay on November 24, 2026, wearing a protective mask. The match finished 0-0.

I tracked his positioning data in that game against his pre-injury matches. Distance covered fell by roughly 18 per cent. Involvement in duels fell. Shot quality, measured as expected goals per shot, dropped sharply.

The mechanism is not hard to read. A player in a mask avoids unnecessary contact and tends to read the game a fraction later, especially inside the box where every decision must be made in milliseconds. To compensate, Son dropped deeper to join build-up play, where the physical pressure is lower. That choice gave him more touches but pulled him away from goal — precisely where his value is greatest.

The chain of consequences I predicted then: form would not recover the moment the mask came off, because a habit of playing safe had worked its way into his reflexes and would take time to undo. By February 2026, Son had gone nine matches without scoring.

I do not tell this story to praise myself. I tell it because it illustrates a principle: injury data does not predict that a player will play badly, it predicts that he will play differently — and that difference only becomes a problem when the tactical system is not adjusted around it.

Since 2026 my work has moved into esports, but the three layers have not changed.

In esports the public data layer is rougher than in football. A champion's or a composition's win rate is usually published as an aggregate for an entire period, pooling multiple patches, regions and formats. Split the same data by patch, by side, by match duration and by role, and a champion can move from being nerfed to being the best early-game pick that is useless if the game runs long. Two opposite conclusions, one dataset, different splits.

That is the same error as the Kazan statsheet: an aggregate read as a single fact.

A win rate that rises after a patch does not necessarily reflect new strength; it may simply reflect that more strong players started picking it at the same time. This selection effect is the most common trap in esports data analysis, and it only becomes visible when you check how players are distributed across ranks instead of staring at the mean.

Here I have to say the thing my readers usually do not want to hear.

Correlation is not causation, and my spreadsheets are not immune. Gladbach's 28 per cent decline is a correlation. It could be caused by the absence of crowds, or by a tougher run of fixtures in that window, or by players' fitness decaying after a three-month break, or by opponents preparing better because the whole league had to adapt at once. I present it as a signal, not a law.

The same applies to South Korea's PPDA of 9.8. Without separating the tactical share from the situational share, the number is a description, not an explanation. And a description presented as an explanation is the most common failure in my line of work.

There is another field where the lack of transparency matters more than it does in data: refereeing.

At Kazan in 2026, VAR corrected a wrong decision. But during the interval between the assistant's flag and the referee's final call, the stadium received almost nothing. Fans in the ground had to wait for a signal on the big screen, and that signal was usually a short line of text. A system can be right about the verdict and still fail at explaining it. To someone in the stands, an unexplained verdict is indistinguishable from an arbitrary one. Transparency stays a slogan until the crowd inside the ground hears the reason, at the moment the decision is made.

The same pattern shows up in the transfer market. Models that price young players are very good at measuring potential: minutes played, rate of improvement, age. They barely measure what decides a dressing room — acceptance by the existing group, the ability to handle pressure in a new city, the quality of the relationship with the coaching staff. A signing can score very high on the model and still fail, and when it does, blame usually falls on the player rather than the model that omitted a variable. Dressing-room chemistry does not sit in any of my spreadsheets, which is why I state that limit before making any prediction about a transfer.

Four tournaments, four different problems, and one conclusion about how to work.

What matters in the next matchweek is not the number on the scoreboard. It is in three signals.

Pressing metrics will keep being normalised against match state, which means PPDA will only carry value when it travels with the scoreline at the moment of measurement. Anyone reading PPDA without the score will repeat the mistake I made at fourteen.

Home advantage will keep being repriced. Leagues with thinner crowds, or tournaments staged at neutral venues, will supply more data to answer the old question: how much of the edge belongs to the crowd, how much to the referee, and how much is simply habit repeating itself.

And in esports, the signal will not come from aggregate win rates, but from datasets split by patch and by match duration. The finer the split, the faster a dataset touches the truth.

I started this work with a squared notebook and twenty-three passes of discrepancy. Years later I am doing exactly the same thing: following ink traces, checking definitions, and stating clearly which part of an answer is evidence and which part is a guess.

When you read a statsheet in the next matchweek, ask one question before you believe it: who generated this number, under which definition, and in what conditions?