Model Error Before Measurement Error: The Silent Flaw in Modern Football Analytics

On the night of November 12, 2026, at the FC Seoul training base, I placed a 47-page report on the coaching staff's table. They flipped to the summary page, nodded, and folded it shut. The other forty-six pages — the coordinates of every central-channel reception, the average distance between the two lines, the heat map of zones nobody touched — stayed on the table.

I sat there until nearly dawn. Across 38 K League Classic matches that season, the side under Hwang Sun-hong produced an average of just 1.7 shots per match from the central channel, the lowest in the league. On its own, that figure means nothing. It only began to speak when I drew it as five geometric boxes on a single A4 sheet: which zones were left empty, which were filled with harmless sideways passes.

Six years later, I opened another file on my machine. It contained no calculation error. It was simply empty. And inside that emptiness I read my own work back to myself: a fluent, reasonable, wholly confident analysis — of something that had never been measured.

That was the moment I understood why what I fear most is not measurement error, but model error. A measurement error is a small deviation, fixable by re-running. A model error is a building raised on an empty foundation, and it collapses exactly when you need it most.

When the pitch becomes a coordinate grid

The 2026 World Cup in Qatar was the first time FIFA deployed semi-automated offside technology at the scale of a full World Cup. The system used 12 dedicated cameras mounted under the stadium roof, tracking 29 data points on each player at 50 times per second, plus a sensor inside the ball transmitting 500 times per second. FIFA published this configuration in July 2026.

Which means that at some given instant, we know to the millimetre where a defender's toe was. And at that same instant, we have no idea why the forward facing him chose to run toward the far post instead of the near one.

This is the central paradox of the decade. Data resolution has grown exponentially; the resolution of meaning has barely moved. We measure the foot, not the intention. We measure the distance covered, not the decision made.

A professional match now generates thousands of event data points, plus positional data for all 22 players across 90 minutes. No human can read that volume with their eyes. So we do the only sensible thing: we build a model to read it for us.

And every model begins with an assumption. That is where the risk is born.

The gap gets filled with story

In June 2026, in Nizhny Novgorod, I sat in the stands watching South Korea face Sweden. The match ended 0-1; Andreas Granqvist converted the penalty in the 65th minute after a VAR review.

After the final whistle I did not go back to the hotel. I sat and rewatched all six Asian qualifying matches, rewinding and replaying the passages without the ball.

I was not looking for an individual mistake. I was looking for distance.

In the first half, the average distance between South Korea's midfield line and forward line when the team pushed up to press was roughly 48 metres. Son Heung-min, at the top of the shape, received nine passes across the full 90 minutes. Nine. A forward of continental class, isolated in a space half the width of the pitch.

I wrote 200 pages of notes on that campaign and published exactly one short piece. Those two hundred pages are still on my hard drive. I criticised myself for not daring to say plainly what I had seen.

But the larger lesson lay elsewhere. South Korea 2026: we did not lose on the pitch, we lost from the moment we believed we had won. That belief does not appear in any statistical table. No data provider sells you a metric for the degree of complacency. It has no column, no unit. And because it has no column, in every model it defaults to zero.

That is the mechanism I want to describe.

When an information point is missing, the model does not stop. It interpolates. It assigns a default value. Or worse: the analyst standing in front of the gap fills it with story — with intuition, with reputation, with the feeling of the crowd. It sounds harmless. But the story stuffed into that gap flows downstream: into the news bulletin, into the scouting dossier, into the selection decision, into the odds.

Three tiers of risk, ordered by ascending danger.

The first tier is definition drift. The same shot returns two different expected-goals values from two different providers, because they define a clear chance differently, weight defender positions differently, handle blocked shots differently. Nobody is lying. They are simply two different rulers carrying the same name.

The second tier is lost context. Event data has no column for a player in pain, for a defender whose concentration just slipped, for a team that flew four hours and slept three. In the empty-stadium season of 2026, with the stands bare, I could hear the breathing of defenders and the cracking of tactics. No sensor records that cracking sound.

The third tier, and the one that worries me most, is the commercial motive of the data itself. Live data supplied to betting companies is the darkest side effect of the digitisation of sport. When money flows toward transmission speed, the pressure is no longer to analyse correctly but to analyse quickly. And in every system, fast always beats right.

In the 2026 K League season, I built a database of the gaps between lines. I divided the pitch into cells, counted how often the ball entered each cell, then cross-referenced that with the shots generated from it. The result emerged as an inverted triangle: the team controlled the ball most in the wide corridors, created the fewest chances centrally, and connected the two zones with sideways passes that never cleared the opponent's second defensive line.

Not one data point there is wrong. The error lay elsewhere: if I presented only the possession figure, everything looked fine. The error only became visible when I was willing to draw the space as a shape.

A tactical system only survives until it meets a larger system. Analytical models are the same. They survive until they meet a question their data was never designed to answer.

This industry pays for confidence, not for accuracy

There is something I have to say, even though it is not easy listening for my own profession.

If I submitted a 47-page report today in which 40 pages read insufficient information to assess, I would be replaced within a week. Nobody pays for an empty table. But if I submitted a 12-page report with five handsome charts, three decisive conclusions and one clear recommendation — even if those three conclusions rest on assumptions never tested — I would be invited back the following week.

That is the incentive structure of the industry. It rewards the appearance of certainty and punishes caution. Which is why most professional football analysis does not fail at the calculation stage. It fails at the admission stage.

There is a counter-argument I hear constantly: more data means more certainty. I think that is true in physics and false in football. Every additional variable does not merely add information; it opens one more way for you to be wrong. With ten variables you can be wrong in ten directions. With a thousand variables, the number of ways your model can drift from reality grows faster than the rate at which you collect them.

The day I realised that data does not judge, it only exposes. It exposes what we measured, and equally what we chose not to measure.

So where is the blind spot in execution?

It sits in the distance between the person doing the data and the person making the decision. Coaching staff do not read 47 pages. They read one page, or none, and instead they trust their eyes on the training pitch. That is reasonable. But when a decision is made on the basis of an eye test plus a handsome chart that was misread, the outcome is not a better decision. It is a decision camouflaged in numbers.

I learned this the expensive way: every report I have written since must fit on a single A4 page, and must contain at least one spatial model. If I cannot draw it, I have not understood it.

What to verify next match

If you are about to watch a match and you have a data table in front of you, try one thing.

Before reading any metric, ask yourself: which information point is missing, and what has the model filled it with?

When you see a team with 63 percent possession losing 0-2, look for the gap between the lines before you look for an explanation. When you see a forward who has not scored in five matches, count the passes he received in dangerous zones before you blame his form. When you see a coach criticised for getting his tactics wrong, check how much real data he had for his midfield.

And when you cannot find the answer — leave the cell empty.

Across three decades, I have come to realise: football changes its shirt, but the core remains a contest of minds. That core is not in the spreadsheet. It is in the moment someone dares to say I do not know yet, before saying I am certain. A mature analytical culture is not measured by how many models it owns, but by how many times it dares to leave a cell blank.

The question I carry into the next match is very simple: if I deleted every metric from this analysis, what would I still have left to see on the pitch?