When the Data Is Empty, Football Invents Its Own Truth

In February 2026, in a meeting room in Boston, a screen displayed a transfer dataset of 1,240 rows. Every column was full: player name, parent club, transfer fee, contract length, medical date, agent name. Only one column was empty — the source column. The technical team ran a valuation model over that table. The system raised no error. It returned results: complete, correctly formatted, plausible enough that nobody in the room thought to double-check. And it was entirely wrong.

I saw, with my own eyes, how an analytical machine invents its own truth. It does not lie in the human sense. It fills the gap with whatever sounds most plausible. Football has a politer name for the phenomenon: transfer rumour.

Every transfer window, thousands of rows like that are pushed onto the market. A deal is described with a fee, a contract length, a release clause, a weekly wage, a medical date. Fans read it and believe it. But trace the source column — the only empty one — and most of those figures have no root. They are generated from a void, by a process that was never designed to stop when data is missing. It was designed to continue.

In systems engineering this is called a fail-open fault: the input is broken, yet the system keeps running and returns output instead of halting. The opposing principle is fail-closed — if data is missing, return null, raise an error, and produce nothing at all. Modern football, at the information layer, runs in fail-open mode. That is why I am writing this.

Three tiers of truth

In sports data analysis, I impose one rule on myself: every claim must be labelled by tier. Tier one is what the source document states explicitly. Tier two is reasonable inference from what already exists. Tier three is high-probability speculation with no anchor. The problem in football is not that all three tiers exist. The problem is that all three are presented in the same tone of voice.

A typical transfer report blends all three tiers across four sentences. The first names a player. The second names a figure. The third names a date. The fourth names a source close to the deal. All four read identically. None declares its own tier. The reader has no instrument with which to separate a fact verified from two directions from a guess delivered in a confident register.

I once received a transfer dataset from an international platform in which 62% of rows carried a specific transfer fee but only 9% carried independently verified sourcing. Six in ten figures were presented to the precision of a single euro, yet fewer than one in ten could be traced. The precision of a metric says nothing about the reliability of that metric. A wrong entry can still be written to six decimal places.

Pressure, momentum, and what the scoreline hides

My founding principle has not changed in eighteen years: the scoreline is a lie that time has memorised; xG is the testimony. A match can end 0-1 while the event data tells a completely different story. Most spectators only ever receive the scoreline, and most reports are written from the scoreline alone.

In June 2026, at Foxborough, I watched New England Revolution host Toronto FC. Toronto held 72% of possession, fired 21 shots, posted an aggregate xG of 2.3 — and lost 0-1 to a single Diego Fagundez goal. That day I was an intern writing match reports. My editor asked me to celebrate the miracle. I reopened the StatsBomb event data, recalculated every shot, and wrote a piece arguing Toronto deserved to win 3-0, that the result had lied. It reached 50,000 reads in 24 hours and forced the newsroom to publish a correction.

The lesson went far beyond the phrase “the data is always right.” When numbers and narrative conflict, both must be re-examined before either is chosen. The scoreline is a lie that time has memorised. But a strong feeling that one side deserved to win can be a different, subtler kind of fraud.

The Croatia case, 2026

A year later, on the back of that piece, I was invited to build a metrics table for a new sports platform during the 2026 World Cup. Ahead of the quarter-finals I compiled a PPDA table for all 32 teams. Croatia sat at 8.9 — meaning they allowed opponents an average of just 8.9 passes before a defensive action, the lowest of the remaining eight sides.

A reading of 8.9 is usually taken as an indicator of pressing intensity. That reading is mechanically correct and essentially wrong. I wrote about Marcelo Brozović: 13.8 km covered against Argentina, nine ball recoveries before being substituted. What made Croatia different was not the distance. It was that the whole team agreed to suffer together, in one rhythm, against one opponent.

The Croatia PPDA board of 2026 does not measure pressure; it measures pride. That is the line I still use when explaining the metric to coaches. Pressure, at the data layer, is the number of opponent passes completed before an intervention. Pressure, at the human layer, is a collective decision about how much pain a group is willing to absorb. The spreadsheet captures only the first layer. The second has to be read with the eyes.

When Croatia reached the final, my name began circulating in professional conversations. A Championship club hired me as a part-time data consultant. I moved from writing by feel to writing by system: every claim must carry a metric threshold, and a statement of how much that threshold can be trusted.

The empty stadium of 2026: a natural experiment

In early 2026, the pandemic emptied stadiums worldwide. The Boston consultancy where I worked cut 40% of its staff. I did not ask to be spared. I wrote a report titled The Stand Effect: Evidence from 372 Bundesliga Matches Before and During COVID.

The result: home win rate fell from 45% to 31%. Penalties awarded fell 28%. This is the class of data I call a natural experiment — a single variable removed at global scale, isolating the crowd effect from everything else. The empty stadium of 2026 is a natural test: football does not need a crowd to reveal its nature.

After that report, Huddersfield Town hired me for the final eight rounds of the Championship. I proposed a rotation model built on sprint distance above a 6m/s threshold: any player who fell below 80% of his individual sprint baseline in two consecutive matches would be benched. Huddersfield took 14 of 24 points and survived with exactly one point to spare.

The real value of the model lay in the coaching staff's response. They did not ask whether the metric was correct. They asked what happens if the metric is wrong. We built a fail-closed procedure: if a training session's GPS data fell below the minimum reliability threshold for sample points, that session's metrics were flagged as null and dropped from the model rather than interpolated. No data, no decision. That is a habit most football information platforms do not have.

Morocco 2026 and the cost of an amplified metric

Ahead of the 2026 World Cup I published a series arguing that Morocco do not defend — they operate on data. I pointed out that goalkeeper Yassine Bounou posted a goals-prevented above expectation of +4.3, and that Achraf Hakimi averaged 6.8 progressive passes per match. I predicted Morocco would reach the semi-finals. When they beat Portugal 1-0, international platforms came calling.

In the summer 2026 transfer window, a Saudi investment fund asked me to assess Cristiano Ronaldo for a contract extension. I wrote a 40-page report. The conclusion: Ronaldo's actual xG creation sat at 0.55 per 90 minutes, but was inflated to 0.82 by set-piece situations and by the way composite metrics are constructed. I recommended against paying more. The fund disagreed. Three months later, Ronaldo's market valuation had fallen 15%.

That story is usually told as a data victory. I do not see it that way. If my report was right, it was right because I isolated one specific effect — the set-piece advantage — out of a composite metric that blends many things. It was not right because I am cleverer than anyone else. Transfer data is like a tide: you cannot read it from the surface; you have to measure the seabed. And the seabed, here, is the structure of set-piece situations.

Correlation is not causation, and a gap is not a fact

Here I have to argue against myself.

My trade has one great temptation: concluding early when the data is still a fragment. A five-match sample looks like a trend. A metric rising over three rounds looks like a tactical turning point. A player covering two more kilometres than last season looks like a physical transformation. All three can be noise.

The same happens in how we read transfer news. When a club pays 70 million euros for a striker, we assume the price reflects ability. It may reflect ability. It may also reflect a bidding war between two clubs, a release clause triggered at the right moment, or money that needed to leave the books before a deadline. The price is a market event. Ability is a technical event. The two correlate, but correlation is not causation.

xG does not judge anyone; it merely exposes the truth the scoreline conceals. Precisely for that reason, xG cannot license conclusions about a person's essence either. A player with low xG across ten rounds is not necessarily a bad player. He may be playing in a system that generates no chances for his position. If I forget that, I turn a metric into a verdict and myself into a poor judge.

The principle of the gap

This is where the lesson from the Boston meeting room matters most.

When a data column is empty, the reflex of a modern system is to fill it. When information is unverified, the reflex of a platform is to write it in a confident voice. When a player has not been contacted by a club, the reflex of the market is to sell him a scenario. All of it is fail-open. All of it manufactures something with the shape of a fact and none of its content.

The only counter is to accept the gap. In the reports I send to clubs, some cells are left blank and marked: insufficient information to assess. That is a complete professional statement, not a concession. A properly flagged gap is more valuable information than an invented metric.

In an internal experiment we compared two ways of handling the same transfer dataset. The first filled empty cells with the column mean. The second left them empty and flagged. The first produced a model that looked complete, with an average error of 18%. The second produced a model with visible holes, yet every prediction on the populated portion carried an error below 6%. A model that knows it does not know is an honest model. A model that appears to know everything is a dangerous one.

Signals for the next cycle

The current transfer window is at its loudest phase. This is when the source column is emptiest, and also when figures are written with the greatest apparent precision. That paradox survives only because readers lack a filter.

The filter I propose is simple. Before believing a transfer story, look for three things: the structure of the clauses, the effect on the wage bill, and the name of the party responsible for payment. The transfer fee is a metric. The clause structure is a fact. And in most deals, the clause structure is the real story.

I have never kicked my data habit; I have only changed suppliers. Better supply is not more supply. Better supply is supply that can be traced, that carries dates, that carries names, and that leaves some cells deliberately empty.