The Empty Cells: When Football Data Is Forced to Stay Silent

At 10 PM on 15 July 2026, I sat on the second floor of a library on William Brown Street in central Liverpool, staring at a spreadsheet with 64 rows. Each row was a match from the World Cup in Russia. The final column, the one I had labelled "set-piece xG", was completely empty. Not because I was too lazy to fill it in. My model had no cell to fill it with.

Four weeks earlier, I had published the longest analytical piece of my writing career, predicting France would win the tournament from the group stage, based on an average of 2.4 xG created per match. The comment section was unkind. Croatia quietly walked all the way to the final. By the time Luka Modrić lifted the Golden Ball, I had locked myself in that library for two weeks, reopening every file and rechecking every formula.

What I found was not a miscalculation. It was a variable I had never entered. Corners, direct free kicks, dead-ball situations my model, built to describe open play, had assigned a probability close to zero.

That empty column taught me something it took me another six years to write down properly. In football analytics, the most dangerous thing is rarely a wrong number. It is usually an empty cell that somebody fills with a guess dressed as data.

Twenty-eight years after I walked into the sports desk of Belgrade Television with a staff pass still smelling of printer ink, I still do exactly one job: read spreadsheets and try not to lie. But the spreadsheets are far bigger now.

An industry that has learned to speak in numbers

In 2026, not yet twenty, I joined the sports department of Belgrade Television. My job was to record scores, stack tapes, and learn not to comment before I had enough evidence. That discipline came from an old editor who always asked one question before letting anything go on air: "How many times did you count it?"

Thirty years later, that remains the hardest question in the trade. It is also the question most of modern football analytics has quietly stopped asking.

Now everything has a metric. A Premier League match generates over a thousand geolocated events. Every pass, every duel, every sprint is logged with timestamps accurate to a hundredth of a second. Leading English clubs employ dozens of data analysts, recruit physics and applied mathematics graduates straight out of university, and pay some of them more than youth-team coaches. xG, expected goals, has moved from a niche term among spreadsheet obsessives to something shown on national television, right beneath the team name.

That popularity carries a trap. When data becomes a shared language, the pressure to speak it rises for everyone. And when forced to speak, people tend to fill every blank, including the blanks the data is not yet thick enough to answer.

The most recent major tournament showed me this more clearly than ever. The 2026 World Cup in the United States, Canada and Mexico is the first edition with 48 teams and 104 matches, opening on 11 June 2026 and ending on 19 July 2026. Going from 64 to 104 matches means roughly sixty percent more recorded events. Mathematically, that is good news: bigger sample, less noise.

But there is another side few mention. With 48 teams split into 12 groups, the quality gap between group-stage pairings widens. A striker can score three goals against an opponent ranked more than a hundred places below his own side in the FIFA ranking, and those three goals sit on the same row as three goals scored against a quarter-finalist. The spreadsheet does not distinguish. The person reading the spreadsheet has to.

In twenty-eight years in this trade, I have never seen a period in which people were this enthusiastic about numbers. Post-match studio shows now assume at least one chart. Transfer bulletins attach a target player's progression metrics. Youth academies teach fifteen-year-olds how to read a personal report. That shift is real, and mostly good.

But precisely because data has become a shared language, it has become the easiest thing to abuse. A wrong number can be caught. An empty cell filled with a guess is far harder to catch, because it looks exactly like every other cell in the same table.

That is why I want to recount four times my model was wrong, and one time it was right.

Anfield, January 2026: when the number was more convincing than the eye

On 14 January 2026, Liverpool hosted Manchester City at Anfield. City arrived unbeaten in the Premier League that season. In the ninth minute, Alex Oxlade-Chamberlain opened the scoring with a long-range strike. Leroy Sané equalised before the break. In the second half, Roberto Firmino put Liverpool ahead, Sadio Mané doubled the lead two minutes later, and Mohamed Salah made it 4-1 on 68 minutes. Bernardo Silva and İlkay Gündoğan pulled two back in the final ten minutes, but Liverpool won 4-3. It was City's first Premier League defeat of the 2026-18 season.

Throughout that campaign, I tracked Jürgen Klopp's Liverpool with a single metric: PPDA, the number of passes an opponent is allowed before a Liverpool player intervenes. Liverpool's average was 8.2, the lowest in the league. Manchester United under José Mourinho sat at 15.7. A gap of nearly double, describing two almost opposing philosophies.

As a transfer market administrator, I was able to follow matches through event data, not only with my eyes. My job then was to assess players and put proposals to the recruitment department, and I began putting pressing metrics into every report. Some weeks I would watch the same match three times: once with my eyes, once through a heat map, once through a chronological event list. Each time I found a detail I had missed. And each time I saw more clearly the limits of what I was holding.

I wrote a long piece on gegenpressing, published it on my personal blog, and received a stream of furious replies. "Robotic." "Football isn't played on a spreadsheet." "You've never stood on a pitch." I read them all, and honestly, part of me thought they were right.

But there was one detail I kept in my head and left out of the piece. In that 4-3 win, three of Liverpool's four goals came from situations where the ball was recovered within seven seconds of losing it. That is a very small number, a very thin sample, not enough to assert anything.

I learned at Anfield that belief is also a variable. Klopp's belief in his system is a variable PPDA cannot measure, and it is what allowed a figure of 8.2 to persist across three seasons without collapsing. My model measured consequences, not causes.

Russia, summer 2026: the empty column and the price of solitude

The 2026 World Cup had 64 matches and 169 goals. I built a simple open-play xG model: shot location, body part, pressure from the nearest defender, type of preceding pass. It gave France the highest chance-creation figure of the tournament, an average of 2.4 xG per match. I concluded France would win and that Croatia's run was luck.

France won. But Croatia reached the final, and the manner of that run turned my conclusion into a punchline. Three consecutive Croatia knockout ties went to extra time or penalties. They won through moments my model did not contain.

The person who is right before his time always pays in solitude. But that solitude is only worth something when it is used to re-examine yourself, not to feel sorry for yourself. Two weeks in a library taught me that.

My error was specific. My model scored every shot with the same function, regardless of where the situation began. A header from a corner was graded on the same scale as a shot from a four-on-three counterattack. Technically, that is a false assumption. Set pieces have a completely different probability distribution from open play: more players in the box, different ball speed, shorter goalkeeper reaction time, and most importantly, the creator is the free-kick taker rather than a teammate in motion.

I recounted by hand. Nearly a third of knockout-stage goals at the 2026 World Cup came from set pieces. My model assigned that group of goals a total xG far below reality. Croatia were not luckier than France. Croatia were better at exactly the thing my model could not see.

That empty column was not a technical error. It was an abandoned confession.

Spring 2026: when context vanished from the spreadsheet

In March 2026, football stopped. Liverpool were 25 points clear of Manchester City and all but certain to win the Premier League after a thirty-year wait. Then the season was suspended. I wrote three drafts and deleted all three.

The problem was not that I had nothing to write. The problem was that my data had no variable for a pandemic. Every model I built assumed the world outside the pitch was a constant. That spring proved otherwise.

When football returned in June behind closed doors, I began logging every match. After a few rounds a trend appeared: the home win rate in the Premier League fell from around 46 percent in previous seasons to around 39 percent. That is a bigger shift than any tactical change over the same period.

Empty stadiums did not distort the data, but they made the truth feel hollow. The home advantage I had long treated as a constant turned out to be largely the singing of twelve thousand people behind a goal, the invisible pressure on a referee in the 89th minute, the heartbeat of a player hearing the groans after a misplaced pass.

From that season on, I added a section to every model called "data context". Before trusting a metric, I ask myself four questions: under what conditions was it measured, how large was the sample, who collected it, and what has changed in the surrounding environment since the data was generated.

Those four questions made my writing longer and less definitive. I accept that.

Euro 2026: when data became a conversation

In July 2026, through an analytics forum, I connected with an Italian analyst working in the youth coaching system. He sent me internal data on Italy under Roberto Mancini. Average distance covered of 112 kilometres per match, not the highest in the tournament. But ball-circulation speed in midfield was outstanding: very short time on the ball per player, and the number of passes before the ball beat the opponent's first line was well below the tournament average.

I wrote a piece arguing that Italy were not a defensive team but a movement machine. It was widely shared, and for the first time in my career I felt something I had missed for twenty years: the pleasure of analysing with someone else checking the work.

xG is a revolution, but every revolution needs time before people accept it. And the fastest way to acceptance is not a lecture. It is putting the data out and inviting others to argue with it.

Italy won Euro 2026 at Wembley, ending a 34-match unbeaten run. Neither of us predicted the correct final score. But both of us were right about how that team operated.

The transfer market: where empty cells cost the most

If there is one place where the habit of filling blanks causes damage in real money, it is the transfer market. Here, every valuation model is forced to return a number, because money is waiting for that number.

In July 2026, Atlético Madrid paid 126 million euros for João Félix, who at that point had fewer than 50 senior appearances in his career. In January 2026, Chelsea paid 121 million euros for Enzo Fernández, who had moved to Europe only half a year earlier. These are real, published, verifiable deals. The point is not that clubs bought the wrong player. The point is that the valuation model was forced to assign a figure accurate to the nearest million to a player with half a season of top-level football.

Every number in the transfer table is a fate waiting to be written. But that table rarely records its own level of uncertainty. No cell says the sample is only 26 matches. No cell says the player has never performed under the pressure of a must-win game. No cell says the league he is leaving is at a completely different level of competitiveness from the one he is joining.

Based on my experience watching the market across many seasons, there is a fairly stable rule: once a deal passes 100 million euros, its uncertainty does not fall, it rises. The fee turns the player into the centre of a sporting project, which brings expectation, which brings pressure, and pressure has never been in the valuation model.

There is also a technical detail rarely mentioned. Modern big deals are usually structured as an upfront payment plus several performance-linked instalments. The published figure is the maximum possible, not the certain one. In a club's internal spreadsheet those are two different rows. On a newspaper front page they are one.

The bubble in young-player prices is a naked gamble. Not because young players are not worth money. But because the price is calculated by a formula that returns a definite answer to a question that has no definite answer.

VAR and the vaguest clause in the laws of football

There is another place where uncertainty is hidden, and it has nothing to do with money.

The VAR protocol revolves around one phrase: "clear and obvious error". The phrase sounds like a technical threshold. It is in fact a human threshold written in technical language. There is no mathematical definition of "clear". There is no unit of measurement for "obvious".

Over years of watching, I realised that most VAR controversies are not about technology. They are about assuming a single objective answer exists, when the laws hand a group of people the decision of whether the deviation has crossed the threshold. That threshold shifts by league, by season, by the light of a December afternoon.

Semi-automated offside technology, used at the 2026 World Cup, removed some ambiguity from drawing the line. It did not remove the larger question: when is an error too small to intervene on? What about a player ten centimetres offside? Five centimetres? One?

The fact that leagues adjusted the thickness of the offside line across consecutive seasons is evidence that the threshold was chosen, not discovered. Nobody found the correct threshold. People simply chose one that provoked less reaction than the last.

Every time I ask myself this, I think of the empty column in the 2026 spreadsheet. Both are the same kind of problem: a grey zone that the system is forced to answer with a binary value.

The counterintuitive angle

There is a widespread belief in the industry that more data leads to more fairness. I am not sure.

More data leads to more concentrated power. A club with twenty analysts will extract more from the same public dataset than a club with two. That gap does not narrow over time; it widens, because analytics is an activity with increasing returns to investment scale. Wealthy clubs can buy both the data and the people who read it.

The second belief, and the more dangerous one, is that correlation is itself an explanation. A team with a low PPDA usually wins more matches. That does not mean lowering PPDA will help a team win more. Both may be consequences of something else: squad quality, stability of the coaching staff, or simply that the team often leads and does not need to press.

This is what I realised when I revisited Liverpool's 2026-18 data. Their PPDA of 8.2 was beautiful, but it was only sustainable because the squad contained players willing to run more than normal, and because Klopp had enough authority to persuade them to do it for three straight years. In reverse, if a club simply asks its players to press more without those two conditions, the metric will look better on the screen and the results will get worse on the pitch.

In a world of long seasons, the awakened can only rely on their own spreadsheet. But an honest spreadsheet must have room for rows that cannot be answered. If every cell is filled, there is a good chance the person filling it is selling you a certainty they do not have.

What to watch in the next round

The 2026 World Cup will be the first tournament where data exceeds the verification capacity of most viewers. With 104 matches in 39 days, the volume of events will be so large that nobody can watch it all, including people in my trade. That is the ideal condition for empty cells to be filled as loudly as possible.

Three signals I will be tracking. First, how public xG models handle set pieces: whether they separate them into their own variable or keep them in a single function. Second, the top-level minutes played by players valued above 80 million euros, and whether that figure appears in any valuation table. Third, how leagues publish their VAR intervention threshold, and whether it is written as a concrete number or left as a vague phrase.

I will log every match, as I have for twenty-eight years, and I will leave empty the cells I have no data to fill.

Perhaps this time, someone will read that spreadsheet and understand that a blank is not something to hide.