A Blank Cell Is Not a Zero: The Silent Trap in Sports Data Analysis

On a Friday morning at my desk in Los Angeles, I opened a nine-section report. It had clear headings, subdivided tables, a risk assessment ranked by severity, and a recommendations section. I read it top to bottom in twenty minutes. Then I read it again, slower.

All nine sections said the same thing: insufficient data. No tournament name. No team name. No player name. No patch number. No date. No source. An analysis of an event that had never been established as anything at all, with a single surviving label at the top of the file: esports.

What stopped me was not the emptiness. In twenty years of this work I have seen plenty of empty files. What stopped me was the packaging. An empty dataset, formatted to standard, will be read as a conclusion. The tables did the persuading that the content never did, and a skimming reader cannot tell a report that concludes "no problem found" from one that concludes "nobody looked".

I work in two stages. Stage one extracts: what happened, who was involved, when, from which source, and how good that source is. Stage two analyses what stage one has dug out. The habit comes from the job: every figure that lands on my desk gets traced to its origin before it is used as evidence.

When stage one returns an empty list, stage two has no raw material. The only correct behaviour at that point is to stop and raise an error. A stage two that keeps running and produces a polished nine-section document anyway is a failure of process, not of analysis.

My work depends on data from wildly different sources. In football, professional providers log every touch by coordinate. In esports, publisher APIs supply match logs, draft lists and pick-ban records. The exposure gap between competitions is enormous. A major domestic league produces minute-by-minute data; a semi-professional regional event may leave behind nothing but a few phone photos of a scoreboard, archived by nobody.

Same rulebook, same scoring, two completely different levels of visibility. And here is the part I want to sit with: the system does not crash. It returns a blank cell. A blank cell makes no noise.

Every sports dataset contains three different things that software tends to render identically. A zero is a measurement taken that came back empty: no shots on target, no cards, no recorded wage arrears. A blank is a measurement never taken. "Unknown" is the state of knowing that no measurement was taken.

Collapsing those three into one cell is the foundational error of this trade. Data never returns "innocent"; it only returns "not found". The distance between those two sentences is the whole of my job.

On 27 August 2026 I watched from a desk in Los Angeles as Liverpool beat Arsenal 4-0 at Anfield, with goals from Roberto Firmino, Sadio Mané, Mohamed Salah and Daniel Sturridge. The traditional shot count did not suggest that margin: 18 to 9. Read that column alone and a familiar story assembles itself — a reasonably even game settled by a few moments.

The first time I ran expected goals on that match, the output was 3.6 against 0.3. A gap the eye could not see.

I did not believe it immediately. My temperament does not allow it. I logged the full match data, then tested the model across the next ten rounds. It held on roughly eighty per cent of them. I changed how I wrote.

The lesson was not that the metric is gospel. Before trusting a figure, ask where it came from, who recorded it, under which definition, and what was dropped along the way. I apply that question to the metrics I build myself, and it is why I never sign an analysis without a stated error margin.

Three years later, June 2026, the World Cup group stage. Germany held 74 per cent of the ball, took 26 shots, and generated 1.8 expected goals. South Korea took four shots and generated 0.8. I trusted the model. Germany lost 2-0, both goals in stoppage time — Kim Young-gwon's after a video review, Son Heung-min's into an empty net.

The numbers were not wrong. They measured the chances each side created, accurately. They did not measure a team forced into must-win territory, a defence sitting so deep that every shot was blocked at close range, or the weight on a player's legs at the ninetieth minute. My model read the first half well and could not read the second. The model was not wrong; the world changed while I was not looking.

Since then, two variables go into every projection I write: the opponent's PPDA, and the game state — which side is obliged to win. That sounds simple. It only became simple after the model failed once in a way I could not forget.

In May 2026 football returned to empty stadiums. I logged 157 Bundesliga matches and found the home win rate had fallen from 43 per cent to 36 per cent. The drop is small enough to wave away as noise. I split the data by month, then by league position, and the trend held. Home advantage in my model had never mostly come from the pitch or the travel. It came from the stands.

Small data is what big data keeps exposing — but only if you are willing to split it open. A tidy aggregate table is usually where a trend goes to die.

Then I went back to that nine-section report. It contained a compliance checklist. Every cell was empty. A reader skimming column headings, seeing no warning flags raised, could walk away with something very specific: this organisation is clean. No match-fixing allegation. No unpaid wages. No contract dispute.

The only fact in evidence was that nobody had looked.

I call this the free certificate of innocence, and it is everywhere in this trade. A player absent from an injury list is not necessarily fit; the league may simply not publish one. A competition with no disciplinary rulings is not necessarily clean; the organiser may not disclose them. A transfer window with no big deals is not necessarily a quiet market; the reporting may only cover the top four clubs.

Draw a four-cell grid: data present and a clean conclusion; data absent and a clean conclusion; data present and a problem; data absent and undetermined. The most dangerous cell is the second. It returns the same answer as the first by an entirely different route. On a spreadsheet the two look identical, and no software will separate them for you.

Expected goals sits inside the same logic. It is not truth, it is a mirror — and a mirror does not lie. What matters is remembering that a mirror only reflects what stands in front of it. It shows me the team. It does not show me the stands, the dressing room, or the clauses in a contract. It is honest within its range. The problem is how rarely we ask how wide that range is.

The whole industry is saying one thing: we need more data. I am not convinced.

More dirty data just makes longer tables. The danger in this trade is not ignorance. It is confidence with good formatting. An empty file with nine headings does more damage than an empty file with none, because the shell has already done the persuading. Readers feel no obligation to audit the structure of a document that looks this careful.

I do not trust correlation either. Two metrics moving together does not mean one drags the other. Fewer home wins without crowds is a real correlation, but the cause is not the advertising boards or the grass. The cause is the noise people make. Assign the cause wrongly and every later projection fails in a way that is very hard to detect, because it still fits the old data.

And here is the part I suspect the industry will not enjoy hearing. A system that returns an error is a healthy system. A system that returns a blank cell and stays quiet is a system covering for its own defect. Between the two failure modes, only the loud one leaves a trail. The quiet one leaves a file that looks immaculate, and someone files it under "handled".

The irony is that the rewards flow toward outputs that look certain. Nobody is reprimanded for filing an empty report. People are reprimanded for filing a broken one.

The signal I will track over the next cycle is not in the score column. It is in the fill rate of the "undetermined" column in every report I read. People publish what they measured and bury what they did not. Anyone who states plainly what they do not know is lending me their credibility.

If you open a stats table tonight and meet a blank cell, the useful question is not what the result was. It is who left that cell empty, and why nobody went back to fill it in.