The Empty Report and Football's Broken Data Ingestion Layer

On a Tuesday morning, a nine-dimension analysis file landed in my inbox. I opened the first line: insufficient information. Second line: insufficient information. By the ninth line, I was checking whether the file itself was corrupted. It was technically intact. It simply contained no football event at all.

It felt like sitting in a video room, rewinding a match, and discovering the footage has lost its picture but kept the crowd noise. You know something happened out there. You do not know what.

The only surviving token in that file was a label: football.

A professional football analysis today passes through four layers before it reaches a reader. The ingestion layer pulls the raw text from the source. The parsing layer slices it into a title, a source, an article type, and a list of information points. The entity-recognition layer extracts club names, coach names, player names, competitions. Only then comes the analysis. Those four layers stack like the lines of a defensive block. When the first line breaks, the other three can only stand and watch.

For a football article, the minimum input needed to open the tactical dimension is four things: a named subject, a time window for the match or period, at least one tactical descriptor, and one process metric. That metric might be xG, xGA, PPDA or pass completion. If the piece covers a single match, you also need the opponent, the scoreline and the key substitutions. The financial dimension needs a fee, a contract length and a selling club. The league-landscape dimension needs at least two named clubs to build a comparison axis.

One more field was missing: the publication timestamp. Without it, nobody can place the article in a phase of the season, and every judgement about competitive pressure collapses. The same 0-0 means two entirely different things on matchday five and on matchday thirty.

The file I received that Tuesday was empty in every one of those cells. No title, no source, no article type, no information points, no entities. Only one field carried a real value, and that field was the domain label.

My first task was to locate which layer had failed. Three possibilities existed.

First, ingestion failed. The source page sat behind a paywall, or its content was rendered by JavaScript, so the fetch layer received an empty body. In that case, no downstream layer has any work to do.

Second, parsing failed. The body arrived intact, but the DOM selector pointed at a block that does not exist, so the text was trimmed to an empty string before any model ever read it.

Third, entity recognition ran and returned nothing. That possibility is weak, because a normal summariser still retains at least one proper noun, even from a thin article.

The traces in the file lean toward the first or second explanation. The reason is that the domain label survived while every content field was empty. A label like that is usually assigned from metadata or from a URL path, not inferred from body text. The system knew the article was about football. It had never read a single sentence of it.

The incident is small. How the industry reads it is not.

Through the regular season I track matches in fifteen-minute blocks. That slicing carries a technical consequence few people notice: each block is an independent data unit, and an empty block is not the same as a quiet block. If nothing happens between minutes 60 and 75, a model may record "no pressure". But if the footage lost exactly those fifteen minutes, the model records the same thing. Two completely different situations, one identical output.

An empty data cell is not a zero. It is an unknown value, and every sum performed on it was already wrong before it began.

I learned this early. In 2026, as a first-year student writing a tactics blog, I analysed RB Leipzig's 4-2-2-2 under coach Ralph Hasenhüttl, focusing on how Timo Werner moved into the space behind the defensive line. A commenter asked what a girl knows about pressing. I did not answer. I rewatched fourteen matches, counted 212 pressing actions, built heat maps and published them alongside the piece. Had I accepted an empty dataset that day and called it "nothing worth noting", the number 212 would not exist.

Three years later, when the Premier League returned during the pandemic with five substitutions, I tracked twenty Liverpool matches and found they intensified pressing between minutes 60 and 75, exactly the window in which opponents usually threw on three players at once. Their expected goals rose by 0.23 after those substitutions. A student in my class once asked why I did not simply skip the matches with missing minute-by-minute data. The answer sits inside that 0.23: fill the missing minutes with zero and the increase disappears, and the conclusion flips.

Every number is a witness statement. My job is to make sure it cannot lie. When a statement is absent, the first task is to record the absence, not to record that the witness had nothing to say.

The counter-intuitive angle sits here: the biggest risk in a football data system is not wrong data, but missing data read as clean data. An empty file passes the correct process, meets the correct format, and returns the finding "no issues detected". On a risk register it gets marked low. In reality, "not ratable" is the correct answer, because with zero identified exposures there is nothing to score by probability multiplied by impact.

Football is already used to this error at a smaller scale. A centre-back never appears among the tackle leaders because his team holds seventy per cent of the ball. A midfielder is underrated because he plays in a league without metric coverage. A transfer story cites no source, then gets quoted three times until it becomes true. Bias is just noise data the market has not learned to process.

The transfer market is a chessboard where spectators only see pawns move. But even pawns need a board with coordinates. Tuesday's file had no coordinates at all.

If I had pushed it into a transfer meeting unchanged, readers would have concluded that the club involved had no fitness problems, no financial risk, no dressing-room pressure. Not one line said so. The layout said so.

I do not predict. I just read data one beat faster than everyone else. And the first beat of any process is checking whether there is data to read at all.

What I want to see in football data workflows this season is not a more accurate prediction model. I want a hard gate: if the input carries fewer than one information point and one named entity, the system must emit "insufficient input" instead of filling a blank template. Engineering calls it null handling. Analysis calls it honesty.

A mature analysis system is not measured by how many dimensions it covers, but by whether it dares to say "I do not know" in the right place.