When Football Data Learns to Lie Like Theranos

Elizabeth Holmes blinked six times in one minute. A body-language expert sat in front of a screen, started a stopwatch, and delivered a verdict: this woman is hiding something. An average person blinks 15 to 20 times a minute. The clip replays the detail of six blinks, and the room nods as though it has just heard a sentence handed down.

The footage sits in the trailer for the documentary “You Can See Everything”, premiered at the Telluride Film Festival, scheduled for cinemas from October 16. Body-language expert Patti Wood, clinical psychologist Denise Dudley and behaviour analyst Traci Brown take turns dissecting Holmes's expressions. Directors Lance Oppenheim and Nathan Fielder stand behind the project.

Nobody needs further evidence. One metric will do.

I watched that trailer four times. Not for Elizabeth Holmes. I watched it because I had seen the same move somewhere far more familiar: on the analysis board before every matchday.

The machine never ran, but the report was flawless

Theranos was founded by Holmes in 2026, was once valued at roughly 9 billion dollars, and dissolved in 2026. In January 2026 she was convicted on four counts of fraud. In November of the same year the court sentenced her to 11 years and three months in prison. The company's blood-testing device never worked as advertised. The investor decks carried not a single scratch.

Theranos did not collapse because of data. It collapsed because of a belief: that a clean metric equals a verified fact.

Football analytics grew up in exactly that window. Between 2026 and 2026, information moved from scouts' notebooks onto paid data platforms, from elite analysis rooms down to amateur commentary channels. The promise was beautiful: measure everything. The risk went unmentioned: measure, very beautifully, things that should never be measured.

Last week, inside a system I monitor, a record was tagged “football”. Inside it was a story about Holmes, about body-language experts, about a film screening. No club. No player. No scoreline. The tag sat there, clean and confident.

One wrong tag does not ruin a match. A thousand wrong tags ruin an entire model. What worries me more: the tag was born at the ingestion layer, before anyone had a chance to read the content. That is the most frightening execution blind spot of all — wrong before it could ever be right.

A glove priced by its passing

In July 2026, André Onana left Inter Milan to join Manchester United, with reported fees between 47 and 51 million euros. His profile at the time ranked among Europe's best in one very specific category: progressive passes under pressure and involvement in build-up from his own half.

That metric is real. It measures something real. The problem is that it does not measure what a goalkeeper is paid to do.

In his first season at Old Trafford, Onana's save numbers fell and his errors leading to goals rose. But the fee had already been paid, based on a dataset that described him as a deep-lying midfielder rather than a man guarding a goal. A pass is just a pass until you can read the intent of the entire block of space around it.

In the opposite direction, Emiliano Martínez was undervalued for years, even though he owns the hardest thing to teach a goalkeeper: positional instinct inside the box, command of aerial balls, and sheer stubbornness in a penalty shootout. Those qualities do not produce pretty charts. They only produce points.

Goalkeeping distribution has been sanctified to the point where a keeper with declining reflexes keeps his transfer value, while an excellent shot-stopper with average passing is written off as obsolete. Space does not lie — only people lie to themselves with numbers.

The laboratory with no crowd

In 2026, when the Bundesliga restarted in empty stadiums, I sat down with 88 matches. Home win rates fell from around 42 percent to around 30 percent. I built a separate xG model for teams that defend deep, and found that what disappeared was not skill. It was kindling.

A match without a crowd is a pure laboratory — but I used to fear it.

On August 18, 2026, in Lisbon, I predicted Leipzig would not overturn Paris Saint-Germain in the Champions League semi-final. The basis was not squad quality. It was that Leipzig's pressing system needs a roar from the stands to trigger it. Once the trigger vanished, the structure lost about 12 percent of its intensity in the second half. The match ended 0-3.

That year's numbers collapsed, and so did I — then I learned to rebuild from fragments of doubt. Since then I no longer write “Team A lost focus”. I write “Team A lost its organisational capacity when pressing intensity dropped 12 percent”. The difference is that a colleague can now verify me, or refute me.

Applause does not appear in the table

The same logic applies to referees. I do not believe in a secret plan between football officials and big clubs. I believe in 60,000 people in the stands, in dozens of close-up cameras, in a media cycle that runs three days after every contentious decision.

Home advantage in penalty and corner counts has been documented across many studies of European football over two decades. But that advantage is not evenly distributed. It pools toward the club with more fans, more television contracts, more headlines. Referees treat giants and small clubs differently. Pressure, not conspiracy, is quantified through the very metrics we use to praise big clubs.

The execution blind spot

Here is the counter-intuitive angle: the biggest mistake in football analytics lies in hunting for a metric clean enough to end an argument, not in using data at all.

The six-blinks-per-minute figure was never produced to understand Holmes. It was produced to close a conversation. xG, progressive passes, aerial duel win rate — all of them risk being used the same way: as a full stop instead of a question mark.

And the “football” label stuck onto an article about a convicted fraudster is not a reasoning error. It is an input error. The model is not wrong. What the model received was already wrong, and nobody checked. Many source fields in that system read “Source: None”. The report still looked beautiful. So nobody asked.

In Qatar in 2026, I once held a Croatia analysis back for three days, waiting for a perfect model of pressure on Joško Gvardiol. Another analyst published a similar idea the next day and took all the attention. I arrived late because I wanted a perfect map; it turned out the match had already redrawn itself.

Takeaway

Next matchday, I will do one small, verifiable thing. For every goalkeeper praised for his passing, I will time how often he actually intervenes inside his own box across the full ninety minutes. If the gap between those two datasets is wide, I will write it down — not to prove anyone wrong, but to check whether the machine is really running.