When the Data Goes Silent: Lessons from a Failed VAR Model

Minute 67, Round 29 of the 2026 K League Classic, Seoul World Cup Stadium. Lee Dong-gook received the ball inside the box, turned, shot, and the net rippled. In the editing room some 30 kilometres away, I — then a 23-year-old VAR assistant — had already seen his run put him roughly 0.3 metres beyond FC Seoul's last defender. But I was still checking the rear camera angle to confirm the point of contact. My alert left my hands 14 seconds late, double the 7-second standard FIFA recommends for clear offside situations. The referee awarded the goal. No mechanism existed to reverse it.

For three nights afterwards, I rewound that footage again and again. Not to hunt for the referee's mistake. I rewound it to find the mistake inside the process I had just operated.

Every VAR error is a crack in the mirror that reflects the laws of the game. It took another five years, and one badly wrong forecasting model, before I understood that the deepest crack is not in response speed. It sits in the fact that we almost never say the sentence: “There is not enough data to conclude.”

A system designed never to fall silent

VAR runs on an unstated assumption: every incident has one correct answer, provided there are enough camera angles and enough time. That assumption is physically true. A player is either offside or not. The ball either touches a hand or it does not.

But what we call a “decision” is less physical than we imagine. It is a chain of cognitive latency, camera vantage point, the order in which signals appear, and how fast a human processes under pressure. When I started keeping an automated “VAR Decision Analysis” log for every incident from 2026, the purpose was simply self-review. By incident number 300, a pattern had emerged: errors rarely live in the person. They live in the limits of the observing tool.

The 7-second standard FIFA recommends is not an arbitrary figure. It is derived from the average time a VAR assistant needs to recognise an incident, select an angle, confirm the point of contact and transmit the alert. But it rests on another unstated assumption: that the incident is one of those visible clearly from at least one angle. For my 14 seconds in 2026, the incident was clear. For many others, it is not.

In an annual league season, the pressure grows heavier. Every matchday, every report must end with a verdict. Nobody pays for an analysis that closes with “insufficient data”. Newsrooms need a headline, an index, a name for the front page. Referees on the pitch need a whistle so the match can continue.

The noise of the stadium is not written into the laws, yet it carries legal weight. The 62,000 people in the Seoul stands that day did not need to know I was 14 seconds late. They only needed to know whether the goal stood.

Four lessons from the data

The first lesson came from the 2026 World Cup. I was sent to Russia as a VAR analysis assistant for a Korean broadcaster. My task was to track every handball incident across the tournament. When I tallied them up, I had 27 incidents and a result that cost me sleep: only 31 per cent were handled consistently under IFAB's new rule. Nearly seven in ten similar incidents received different verdicts.

I wrote a 40-page report and sent it to the newsroom. They published one small chart, uncaptioned. I started a personal blog and posted the entire raw dataset. It drew 50,000 reads, mostly from referees, sports lawyers and supporters. The 2026 trap was never in the hand; it was in the belief in a definition that does not exist. There is no definition of a “natural hand position” in the laws. There is only a gap, and 27 arms falling into it.

The second lesson came from the 2026 shutdown. That March, global football stopped, and my contract with the broadcaster was cut for budget reasons. I retreated into research as an escape. Six months later I held 1,247 VAR decisions from five European leagues played in empty stadiums. Referee review time at the monitor fell 22 per cent. The rate at which the on-field decision was upheld rose 15 per cent.

What does that mean? With no crowd, referees decided faster. They were also less willing to overturn themselves. Greater speed did not come with greater accuracy. I wrote a 60-page report and posted it to an academic forum. A director at the Asian Football Confederation made contact and invited me to serve as a data analyst for a referees' committee. That was the first time I understood that a valuable finding does not need to declare who was right and who was wrong.

The third lesson is a failure named Kim Min-jae. In 2026, working as a mid-level staffer for a consultancy, I built a player evaluation model on VAR data. The model showed the defender committed 0.73 fouls per match in Serie A, a level I classified as “high card risk”. I advised the firm not to recommend him to major clubs.

Napoli signed him anyway. Kim Min-jae became a cornerstone of the side that won Serie A in 2026-23. My model was not wrong in its arithmetic. It was wrong because it ignored two unmeasurable variables: the covering ability of his team-mates, and the difference between how Italian referees read the laws and how Korean referees do. The same defensive action was legal in Naples and a booking in Seoul. At the end of that year I wrote a ten-page self-critique and pulled the model from the system.

The fourth lesson is unfolding right now. The transfer market has unwritten laws, where price is set by expectation rather than by sample size. A player with fewer than 50 top-flight appearances can be valued at 100 million euros. When I ran a risk model on players under 23, forecast error was more than three times larger than for players over 27. The younger group is not worse. We simply have too little data on them for the model to say anything trustworthy.

The problem is not the data, it is the rewards

I am not arguing that data is useless. I am arguing that we have built a reward system that pays only for decisive verdicts.

In esports, which I have followed for 16 years, a player's career is far shorter than a footballer's. The academy pipeline is thin, and there is almost no safety net after retirement. Yet whenever a 19-year-old shines for three matches, the whole industry slaps the label “next generation” on him. Three matches. Not enough for a statistical model to separate talent from luck. Plenty enough to generate a headline.

The same mechanism runs inside the VAR room. Nobody is reprimanded for issuing a wrong conclusion from thin data. People are reprimanded for issuing no conclusion at all.

A wrong decision does not ruin a match; the silence after it is what corrodes trust. In my case, the legally correct silence — the sentence “insufficient data” — was the hardest thing to say, because it produces no product to sell.

What stays with me

We search the pitch not for justice, but for an excuse to stop arguing. Yet there are matches where the argument should not stop, because the data is still missing, and that absence is itself the real information.

What I have carried from 2026 is not a complete method. It is a habit: before concluding, ask whether the data actually permits that conclusion. For Vietnamese football, where public VAR data remains thin and the number of matches with adequate camera coverage is small, that habit matters more than any model.

The natural position of an analyst is not that of the person who delivers the verdict. It is that of the person who points precisely to where the mirror stops reflecting — and has the nerve to say that this time, the mirror was not bright enough.