Football Is Poisoning Its Own Data

In July, Amazon confirmed an eight-episode order for Stillwater, a horror-thriller adaptation of the graphic novel of the same name owned by Skybound Entertainment. The lead role of Daniel West went to Ben Hardy, a face previously seen in Bohemian Rhapsody, 6 Underground and X-Men: Apocalypse. The script was written by Greg Berlanti and Carly Wray, both of whom also serve as executive producers alongside Leigh London Redman, Sarah Schechter, Robbie Rogers and a Skybound producing bench comprising Robert Kirkman, David Alpert, Rick Jacobs and Glenn Geller. Original author Chip Zdarsky and artist Ramón K. Pérez are also credited.

The item runs under three hundred words. The only attributed source is Amazon, and only for the episode count. Twenty-one information points. Not one club. Not one player. Not one match. Not one league table. Not one release clause, not one wage bill, not one expected-goals figure.

The label attached to it: football.

I read that extract three times. The question that stopped me is the kind I ask myself whenever a data set looks too clean: if the classification system is wrong on one record nobody checks, how many times has it been wrong on the records nobody sees?

And in a transfer window, when thousands of items are pumped into the same pipeline every day, that question stops being a technical matter.

When the whole world believes in the bracket, I believe in the data. But belief in data is only worth anything when you are willing to audit the pipeline that produces it.

---

Context: a stray record inside an industry worth tens of billions

To understand how a casting announcement can sit in the same bucket as Manchester City, you have to understand how the football information pipeline runs.

Every day, sports data systems sweep through hundreds of thousands of articles in dozens of languages. Each article is decomposed into information points: names of people, names of organisations, events, numbers, timestamps. The system then assigns a domain label. That label decides where the record flows: into a tactical analysis store, into a transfer-market sentiment index, into a broadcast-rights valuation model, or into the bin.

In theory, this is a beautiful architecture. In operation, it is a chain of decisions governed by three very human things: keywords, name collisions and time pressure.

I have tasted that pressure. In 2026, while studying for a master's in movement science, I sat up at night building fourteen heat maps from nine Manchester City matches in the 2026/18 Premier League season, purely to prove something everyone at the time dismissed as nonsense: Kyle Walker was not a right-back in the traditional sense. He averaged ninety-eight touches per match, more than David Silva across three consecutive games. The fan community called me a madman destroying sacred tactics. Two weeks later, Guardiola used the word quarterback to describe Walker.

That piece taught me a formula: contrarian argument, visual data, insider quotes. It also taught me something less discussed: public opinion can only verify what makes it into print. The submerged part of the iceberg — records mislabelled, miscounted, summed into totals that nobody reconciles — will never be called out by anyone.

In 2026 I tried another experiment. In May of that year I published four reasons Germany would not survive the group stage at the World Cup in Russia. The data was simple: Germany's possession share against weak opponents in qualifying was only fifty-eight per cent, ten points below the previous cycle; the average age of the midfield was twenty-eight point six, the third oldest of the thirty-two teams. Six hundred mocking comments before kick-off. When Germany lost nil-two to South Korea, the piece was dug up and shared fifty thousand times.

The lesson of 2026 was not that I am good at forecasting. The lesson was that a correct conclusion is only worth something if the input data is correct. If the possession figures I used that day had been mis-sourced, mis-leagued or double-counted, my conclusion would have been right by luck and wrong by system.

That is exactly what is happening with the Stillwater record. It is not a joke played by an algorithm. It is evidence that the pipeline has no domain gate at the entry layer, and therefore no mechanism to stop a bad record reaching the deep-analysis layer.

---

Anatomy of the error: how an actor lands in a football data table

Let us decompose the twenty-one information points exactly the way a system would.

The people group comprises Ben Hardy, Greg Berlanti, Carly Wray, Sarah Schechter, Leigh London Redman, Robbie Rogers, Robert Kirkman, David Alpert, Rick Jacobs, Glenn Geller, Chip Zdarsky, Ramón K. Pérez and Jonathan Gabay. Not one of them is a registered player, coach, sporting director, referee or licensed agent. The organisations group comprises Amazon Prime Video, Amazon MGM Studios, Warner Bros. Television, Berlanti Productions and Skybound Entertainment. No club, no federation, no competition. The works group comprises Stillwater, Bohemian Rhapsody, 6 Underground, Only the Brave, X-Men: Apocalypse, The Conjuring: Last Rites, The Girl Before, The Woman in White and EastEnders. None is football-related.

The only quantified datum in the entire record is the episode count: eight.

So where does the error come from?

The first possibility is a name collision. Proper names are not unique identifiers. A system that matches character strings cannot distinguish an actor from a lower-league footballer of the same name, an assistant coach, or a scout. In the global football database, the number of duplicated personal names runs into tens of thousands. Without an occupation field in the entity schema, every proper name carries a probability of mislabelling.

The second possibility is keyword collision. Words such as series, executive, clause, contract, squad, transfer and signing appear densely in both football copy and television copy. A classifier trained on a football corpus, encountering an item with a high density of administrative vocabulary, will lean toward the football label if the decision threshold is set low to optimise recall.

The third possibility is a cascade effect. Once a bad record receives a label, it can drag related records in the same topic cluster into the same wrong label. In a news environment where hundreds of sites republish the same content, one root error can multiply into hundreds of child errors.

These three possibilities are not mutually exclusive. They resonate.

What matters is the structure of the error. This is not a missing-data error. It is a surplus-label error on top of complete data. The record contains twenty-one information points, meaning the system read and decomposed it very carefully. It understood the content. It simply misunderstood the industry.

And this is the point I want to press hardest in this entire article: a false positive in an information pipeline is more dangerous than a false negative, because it does not create a gap — it creates noise that looks like signal. A missed item only leaves the model short of data, and everyone knows it is short. A mislabelled item gives the model extra data, and nobody knows the extra is rubbish.

In my trade, that is the worst class of mistake. Nobody audits a data table that already looks full.

---

What the record actually reveals: Amazon's capital envelope

Now strip away the wrong label and read the record as an economic fact. At that layer, Stillwater is not meaningless to anyone working in football.

The record establishes three real events. First, Amazon MGM Studios and Warner Bros. Television are co-producing an eight-episode series. Second, Skybound Entertainment holds the underlying rights and retains a deep executive-producer position, meaning the rights holder has meaningful consultation power. Third, the product is in pre-production, with no release date, no disclosed budget and no audience data.

Read alone, this is an entertainment item. Read at the parent-group layer, it is a spending line inside the same capital envelope Amazon uses to pay for football rights.

That is the real blind spot.

Recall the structure of Amazon's assets over the past decade. In 2026 the group completed its acquisition of MGM for roughly eight point five billion dollars, later consolidated as Amazon MGM Studios. On the sports side, Amazon Prime Video held a Premier League package in the UK for several seasons, streaming two rounds per season, including a midweek December round and the Boxing Day fixtures. From the 2026/25 season, Prime Video holds the UK Champions League package for Tuesday matches. In the United States, Amazon holds the NFL's Thursday Night Football under a long-term deal that financial press has valued at about one billion dollars per season.

Every one of those lines is a multi-year, pre-paid commitment that cannot be cancelled mid-term without penalty. And every one of them competes directly with the scripted-content production budget in which Stillwater sits.

This is the structure I call intra-envelope competition: the modern fight for football rights no longer happens between Sky and DAZN, between ESPN and Fox. It happens between the sports division and the scripted division inside the same balance sheet.

There is a telling fact here. Amazon once held Premier League rights in the UK and did not retain that package in the following renewal cycle, while continuing to spend heavily on the Champions League and the NFL. That choice was not random. It was the output of an internal calculation: which package delivers better subscriber retention per dollar spent.

And when a group decides that way, the consequences do not stop in England.

A title is never a surprise to anyone who can read data. What is true of a league table is true of a balance sheet. A rights package is not lost because a rival bid higher. It is lost because someone chose to spend the money on something else that returns more inside the same envelope.

In this specific case, Stillwater is not the cause of Amazon withdrawing from any football package. But it is part of the same allocation decision. Every approved eight-episode order is a line that has been weighed against a football round that could have been bought instead.

This is why I read entertainment news. Not to learn who plays the lead. To learn where the money is going. Do not ask what a star is worth; ask what the team looks like without him. In this case, ask what the league looks like when the paying partner chooses something else.

---

The Vietnamese mirror: when rights are priced by feel

Vietnamese fans consume both ends of that envelope without knowing it.

At one end, they pay subscriptions to platforms holding international league rights. For years, the Vietnamese market has watched broadcasters pay very high prices for Premier League, Champions League and other major rights, then struggle to recover the outlay through subscription price rises, package sales or loss-cutting. At the other end, those same fans are consumers of scripted content made by technology groups, often through the same app, the same subscription, the same credit card.

Which means: an internal allocation decision taken by a group in Seattle can directly change what a viewer in Hanoi or Ho Chi Minh City switches on at nine o'clock on a Saturday night.

Domestically, the story is harder still.

Broadcasting rights for Vietnam's professional leagues have long been an unsolved equation. Domestic television rights values sit far below market potential, while production and operating costs keep rising. The cause is not purely sporting quality. It is the pricing layer.

And that pricing layer is fed by precisely the data we are discussing.

When a broadcaster negotiates rights, it brings a bundle of indicators: viewership, engagement, discussion volume, media reach, social frequency. Most of these are not measured from actual viewing behaviour. They are measured from the volume of content published. And published-content volume is exactly what gets contaminated when classification pipelines go wrong.

Picture the mechanism concretely. A discussion-index model sweeps a football corpus and counts occurrences of each entity. If the Stillwater record sits in that corpus under a football label, it contributes to the occurrence counts of the entities inside it. On a single record, the error is too small for anyone to notice. Across a batch of tens of thousands of records, it accumulates and shifts the entire baseline.

And the dangerous part is this: a skewed index does not set off its own alarm. It still looks reasonable. The chart still looks good. Only the meaning of the chart is wrong.

I have seen a version of this error. In 2026, when the pandemic forced leagues to play in empty stadiums, I obtained detailed Bundesliga data from the first nine rounds after the restart and found a figure never seen before: the home-win rate fell from forty-two per cent the previous season to twenty-six per cent. Eleven of eighteen matches went the away side's way. I wrote an eighteen-hundred-word piece in two hours under a deliberately provocative headline: home advantage is dead.

Conservative journalists attacked me on television. But data analysts invited me onto tactics podcasts, and for the first time I was paid to voice a controversial opinion of my own.

What I learned from that was not whether the twenty-six per cent figure was right or wrong. What I learned was that a nine-round sample is enough to open a hypothesis and not enough to close a conclusion. With domain classification errors, the problem is more serious still, because the sample is not merely small. It is dirty from the labelling stage, and there is no way to know how dirty without actively sampling it.

In the Vietnamese market, where most international data arrives through republication by aggregator sites, the probability of such contamination is higher, not lower.

---

Index contamination: how an actor raises the temperature of a league

Let us go into the technical mechanism with a checkable example.

Suppose a collection batch contains ten thousand records labelled football. Suppose three per cent are mislabelled, meaning three hundred records. Three per cent sits inside the typical error band of keyword-based classifiers operating on multilingual data.

With three hundred junk records, three consequences occur at once.

First, the baseline of every entity-occurrence index shifts upward. Unfamiliar names appear at low but persistent frequency. Nobody audits every low-frequency name.

Second, trend-detection models become noisy. An out-of-industry name can cross a display threshold if it appears within a large cluster of republished records, because such models typically count traffic rather than independent origins.

Third, and most important, the model learns from the bad data itself. If the system uses its own output to update itself, the error doubles over time.

Now place two citable numbers side by side. Across the entire Stillwater record, the number of information points with a clear attribution is two: one for Amazon confirming the episode order, one for the remarks by Berlanti and Wray. The remaining nineteen are unattributed interpretation. The verifiable-source ratio: roughly ten per cent.

Ten per cent. In an entertainment item, that ratio is acceptable, because the consequence of being wrong is a corrected article. In a record feeding a multi-billion-dollar valuation model, that ratio is a catastrophe.

This is where I have to say something plainly that many in the industry do not want to hear: we are building sports-asset valuation models on a data foundation whose source-verification rate is lower than that of a personal blog.

And I am not saying this to shock. I am saying it because I checked.

While tracking transfer news across several markets, I built myself a simple reliability ladder in four tiers.

Tier one is an official announcement from a club, federation or league authority. It is the only source that needs no cross-check.

Tier two is specialist press with reporters who cover clubs directly and whose hit rate can be audited over time. This tier carries high value but is not absolute, because specialist press has its own interests and is sometimes used by agents as a negotiating instrument.

Tier three is aggregating press, republishing tier-two content without adding independent verification. Its informational value approaches zero; its propagation value is very high.

Tier four is content deliberately planted by an agent, club or third party, typically appearing across many channels simultaneously within a short window to manufacture a sense of consensus.

The Stillwater record sits in tier three. It comes from an aggregating outlet, cites no specialist source, and attributes a single datum to Amazon. That is why it was both easy to mislabel and easy to spread without anyone reconciling it.

The transfer-window lesson is direct: if you are filtering rumours, filter by tier, not by the fame of the outlet carrying them. A transfer rumour appearing on thirty sites at once does not mean it has thirty independent sources. It may have one source, multiplied by thirty.

---

Contrarian angle: where I could be wrong

A critique is only credible if it states its own weaknesses. Here are three scenarios in which I could be wrong.

Scenario one: broad labelling is a design choice, not a defect. In many data-collection systems, operators accept false positives to guarantee recall, because the cost of missing an important record is higher than the cost of admitting a junk one. If this pipeline was designed to maximise recall for a coarse screening task, then Stillwater slipping through is a pre-calculated price, and the real check sits downstream, where humans filter.

My response: that is sound in principle, but only if the downstream filter actually exists and actually runs. In this case the record passed two layers, was fully decomposed into twenty-one information points, and entered analysis without any domain gate. If a downstream filter exists, it did not operate. A safety mechanism that does not operate is not a safety mechanism.

Scenario two: the capital transmission channel between scripted content and football rights is too slow to matter. Group capital envelopes are set annually; seasons are decided on three-year cycles. The two rhythms are out of phase, so month-to-month variation may say nothing at all about rights decisions.

My response: correct on timing. But my point is not that a single record causes change. My point is the direction of money. A rights renewal cycle is prepared twelve to twenty-four months in advance. Anyone wanting to forecast the outcome of that cycle should read the group's capital allocation decisions, not rumours about clubs.

Scenario three: I am extrapolating from one case. A single mislabelled record proves nothing about the quality of the whole system.

My response: also correct, and this is the biggest weakness of this article. One case is an anecdote. But I am not presenting it as statistical proof. I am presenting it as a sample test. In quality control, when a defect is found in a batch, the correct procedure is not to fix that defect and carry on. The correct procedure is to inspect the whole batch, because the probability of an isolated defect arising in a systematic process is low.

And this is where I place my bet. When the whole world believes in the bracket, I believe in the data. But when the data itself is contaminated, I believe in the only thing left that can be verified: the process that produced it.

---

An honest conclusion: transmission equals zero

I have to state this clearly, even though it makes the article less attractive.

The Stillwater record has no impact on football. No club, no federation, no league, no player, no betting market, no academy is affected by Ben Hardy playing Daniel West. The correct finding on this record's football impact is: zero.

Saying zero out loud matters more than it looks. In an environment where every item must be converted into an argument, the ability to say that something is irrelevant is a professional skill, not ignorance.

The only football-relevant value in this record sits at another layer: it proves the information pipeline is running without a domain gate, and therefore may be quietly skewing the very indices used to price sports assets. That is not football news. It is news about how football understands itself.

---

Takeaway

In this transfer window there will be thousands of rumours, a few hundred real deals, and a small number of decisions that genuinely shape the future of leagues. The shaping decisions are not in the rumour column. They are in the capital allocation tables of corporations whose financial reports none of us read.

History does not care whether you dare to speak; it only waits for you to be right. And being right here means admitting something uncomfortable: the data system the entire football industry leans on may be counting wrong, and nobody is checking.

If a casting announcement can slip into a football data store unchallenged, what else is already in there?