Four Times My Data Collapsed — and Why the 2026 Transfer Window Forced Me to Relearn the Words 'Insufficient Information'

3:40 A.M., and a Column of Zeroes

My second monitor displayed a tracking sheet of 4,812 posts, 219 source accounts, and a single column containing the number 0. The column was labelled “verified contract documents.” Four thousand eight hundred and twelve pieces of noise, not one piece of paper.

I sat looking at that zero in Chicago, in July 2026, and what I felt was not the curiosity of someone holding a story. It was the emptiness of someone being asked to write an analysis of something that does not yet exist. Fourteen years in sports data analysis had taught me how to read PPDA, how to calibrate an expected-goals model, how to build a heat map for every pressing action. No classroom had ever taught me how to sit still when the data table returns zero.

That week I received four article requests from three different editorial desks. All four shared the same shape: a player rumoured to be moving to a club, a rumoured fee, and a request for “tactical impact analysis.” I read all four briefs again, and then did something I would never have done ten years ago: I replied that I did not have enough data to write, and that if I wrote, I would write about that very gap.

Northampton, the Spreadsheet, and the First Lesson

In March 2026, I was a sociology master’s student volunteering as a data analyst for Northampton Town in League One. The club had no analytics department. It had a head coach named Justin Edinburgh, a video room, and a desktop computer onto which I had to install spreadsheet software myself.

At Northampton, we did not have technology. We had patience and one spreadsheet.

I entered 34 matches of data by hand, counted every opponent pass from recorded video, and then calculated PPDA — the number of passes an opponent is allowed before Northampton makes a defensive action. It came out at 8.7, the lowest in the division. That meant Northampton gave opponents very little time on the ball before closing them down. Alongside that, the team’s chance-conversion rate sat at 14.2%, unusually high against the rest of the division.

Those two numbers sat side by side in my spreadsheet for three weeks. And I nearly misread them.

The misreading was easy: high pressing plus high conversion equals a good attacking side. I wrote a forty-page report concluding that Northampton’s high press was in fact “active defending,” that the pressure was generated not to win the ball in the opponent’s half but to force opponents to pass into zones where the back line was already waiting. Edinburgh pushed the report aside. He said the team was short of centre-backs, not short of a model.

Then came five straight defeats. He called me back, and the only change applied was to drop the pressing line eight metres deeper. No personnel change, no formation change. Northampton stayed up. But what I carried out of that season was not the result. It was the question: if a 14.2% conversion rate is just a small sample, what was my entire forty-page report?

Every number is a story waiting to be verified.

World Cup 2026: When the Definition Is Wrong, No Model Can Save You

In June 2026, I began writing analytics pieces for a football data outlet during the World Cup in Russia. On 17 June, Germany lost 0-1 to Mexico in their opening group match. I ran my own expected-goals model and published that Germany had created 2.1 expected goals, that they “should have won.” The next morning, a veteran analyst sent me a four-paragraph email. He pointed out that I had not subtracted shot angle coefficients and defender pressure from the model, and that the error inflated the output by 34%. The 2.1 was in reality closer to 1.4.

I spent the next six weeks, nearly the rest of the tournament, rewatching all 64 matches and recalibrating the model with tracking data from every phase of play. When Germany were eliminated in the group stage, I wrote a piece arguing against myself, calling my first analysis “a rushed conclusion from raw data.” The lesson was not that my model was wrong. The lesson was that my model was wrong in a very specific direction, and that direction came not from the data but from how I had defined the variable.

Data never lies, but the person defining it can.

From then on, every piece I write includes a short section listing the variables I could not control. That section does not add much length, but it forces me to admit a limit before the reader discovers it on their own.

92 Matches Without Crowds: The Variable That Cannot Be Entered Into a Spreadsheet

In June 2026, the Premier League returned after the pandemic with 92 matches behind closed doors. I was then a junior analyst at a sports consultancy in Chicago. My client was a Championship club wanting to assess the financial impact of losing home crowds.

I took six years of historical home and away performance data, built a model, and produced a forecast: home advantage would fall by roughly 15%.

The reality: home win rate dropped 28%. Average goals per match rose from 2.6 to 2.9. My client lost millions of dollars betting on the number I had given them.

What I had overlooked has a very simple name: the crowd effect. Not fandom in an emotional sense, but a mechanical variable — noise affecting defenders’ ability to hear instructions, affecting whether referees award decisions, affecting the speed of players’ decision-making under pressure. No column in my spreadsheet could record that variable.

The crowd leaves, but the numbers stay — and for the first time I saw them as empty.

After that episode, I built a mandatory assumption-testing process before running any model: interviewing at least five coaches and three players about match psychology, recording it in words, then cross-referencing against the quantitative variables. That process costs far more than simply running the model. It costs far less than paying for a wrong forecast.

Euro 2026: When the Metric I Had Never Modelled Was the Answer

In July 2026, I was assigned to write an analytical piece on Italy under Roberto Mancini for a major newspaper. My model, based on expected goals and PPDA, predicted Italy would be eliminated in the quarter-finals. The basis: Italy created only 1.2 expected goals per match on average, 25% below Belgium. Their total expected goals ranked only seventh in the tournament.

Italy won the tournament.

I went back through the footage, and what I found was not in any column I had ever built. The average distance between Italy’s two centre-backs — Giorgio Chiellini and Leonardo Bonucci — was just 21.4 metres, the smallest in the tournament. That number created no goals, no shots, appeared in no attacking ranking. It created something else: control of tempo. The narrow gap smothered opponents’ counter-attacks before they could become shots.

I wrote the piece “My Mistake: Italy Did Not Need Expected Goals, They Needed Positioning.” It drew 12,000 reads in 24 hours. But its real value lay elsewhere: from that point, I began building spatial metrics into my analysis — distance between lines, team width when losing the ball, speed of ball circulation from flank to central corridor.

A wrong measure is more dangerous than measuring nothing at all.

The Transfer Window: Where Noise Is Packaged as Data

The four stories above are not four anecdotes. They are four repetitions of the same error: I took a real dataset, defined it in whatever way suited the story I wanted to tell, and then presented the result as a conclusion.

The transfer window is the perfect environment for that error, because it supplies a form of data that looks very much like data: engagement counts.

A transfer rumour can be measured by number of posts, number of views, number of feed appearances. Those numbers are real. But they measure the spread of a claim, not the probability that the claim is true. When an account posts that a player will join a club, and that post receives 40,000 engagements, that 40,000 enters my model as a signal. It is not a signal. It is an echo.

In my tracking sheet in Chicago, I classify sources into four tiers.

Tier one is verified documentation — contracts, official club announcements, player registration records. Tier two is reporting from journalists with a track record of accurate reporting on that specific club. Tier three is information from agents or intermediaries, which may be true but often serves a negotiating purpose. Tier four is everything else.

In that week of 4,812 posts, the distribution sat almost entirely in tier four. Tier one was zero. Tier two had three items. Tier three had one.

And here is where I have to say the thing I know will displease some editors: when tier one and tier two are both zero, then a tactical analysis of that player, of his impact on the new club, of how he fits a formation, is a piece without foundation. Not difficult to write. Without foundation.

The structure of the release clause and the wage bill is the real story, but they only become a story once there is a contract. Before that, they are numbers leaked deliberately by parties to shift the negotiating position. A leaked fee is not a fee. It is a negotiating position brought into the light.

I understand why that kind of piece exists. A reader in the middle of a transfer window is not seeking truth; they are seeking something to follow. A piece saying there is nothing yet to analyse gives them nothing to follow. But a piece that constructs certainty out of nothing gives them something worse: a belief that will have to be dismantled, and a reason not to trust the next piece.

The Counterintuitive Point: “Insufficient Information” Is a Professional Output

In analytics, returning an empty result is usually treated as a sign of weakness. A model that returns “undetermined” is considered a broken model. An analyst who returns “not enough data” is considered someone who has not finished the job.

I think that order has it backwards.

In medicine, a test without sufficient specificity will not be used to reach a conclusion. In statistics, a sample of too small a size will not be used to generalise. Only in sports does the industry accept taking whatever dataset is available, regardless of what it measures, and calling its output a conclusion.

Every match is a data sample, but belief is the one variable that cannot be entered.

What worries me most is the cumulative effect. Each time a piece is built on tier-four data and presented in the same confident register as a piece built on a signed contract, readers learn that the two kinds of information carry equal value. When that happens, the entire source-evaluation system collapses. And when the entire source-evaluation system collapses, publishing any number at all becomes meaningless — including the correct ones.

There is a reasonable counterargument I have to face. One could say: if you wait until there is a contract, the piece no longer has news value. True. And that is the price of holding a standard. But it is also worth looking at the structure: the news value of a rumour-based transfer piece lasts about 48 hours. The news value of a piece analysing contract structure, release clauses, and wage-bill impact lasts until that contract expires — sometimes three years. We are trading three years for 48 hours.

I have tested this with my own data. Over the last four transfer windows, I tracked 212 high-spread transfer rumours. The hit rate — in the sense that the player actually signed for the rumoured club — was 31%. Of those 212, only 19 had tier-one or tier-two documentation at the time of publication. The hit rate for that group of 19 was 84%.

The gap between 31% and 84% does not lie in the quality of the reporter. It lies in whether or not there was a document to cross-check against.

Signals for the Next Round

I do not believe in intuition. I believe in data — and it was data that taught me to trust no one.

But data taught me something else too, and it took me fourteen years to write it plainly: when there is no data, the only correct move is to say there is no data.

For the rest of this transfer window, I will track three specific signals. First, the ratio between number of posts and number of verified documents per rumour — when that ratio exceeds 1,000 to 1, I read it as a sign that information is being pushed, not discovered. Second, order of appearance: documentation first or rumour first. When a rumour appears more than seven days ahead of documentation, the probability it is accurate drops sharply in my dataset. Third, which side leaks — the selling club, the buying club, or the agent. Each side leaks for a different reason, and that reason determines how I read the number.

Some will say this approach makes me miss the beat. Possibly. I have missed a few pieces. But in fourteen years, I have never had to retract a number because I built it out of nothing — and in this profession, that is the kind of record that never appears on any table.

And that column of zeroes on my second monitor? It is still there. I have not deleted it. It is the most honest data row in the entire spreadsheet.