From Ali Dia to the 2026 spreadsheet: how labeling errors misprice footballers

In November 2026, Graeme Souness picked up the phone in his office at Southampton. The voice on the other end claimed to be George Weah — the 2026 European Footballer of the Year, then at AC Milan — recommending Ali Dia, a striker said to be his cousin, who needed a club. Souness was short of forwards. He offered Dia a trial, then a one-month contract.

On 23 November 2026, Dia came off the bench against Leeds United, touched the ball a few times, and was substituted after 53 minutes. Two weeks later the contract was terminated. The caller was not Weah. It was an acquaintance of Dia, and the story of the Ballon d'Or winner's cousin went straight into a Premier League club's decision system without a single verification step.

Thirty years later, that system is software, APIs and data contracts worth hundreds of thousands of euros a season. The error is unchanged. Only the speed of transmission is different.

This week, in the inbox of a content-processing pipeline I work with, there was a package labelled "football". Inside there was no club, no player, no competition, no transfer. It was a general news item. The classifier had mislabelled it at the very first layer, and without a cross-check it would have travelled onward into the database, been counted as football signal, and contaminated every model downstream.

I raise it because it is a miniature of what happens daily in scouting, only at a different scale.

A modern transfer passes through four layers: collection, labelling, verification, decision. Collection is the most expensive and best funded — wide-angle cameras, positional data, contracts with metric providers. Labelling is almost free, often done by a junior analyst on a Friday afternoon. Verification is the first thing cut when budgets tighten. Decision belongs to whoever holds authority, and that person rarely reads the three layers beneath.

The result is a paradox: the industry spends hundreds of millions of euros to learn that a striker covers 11.2 km per match, yet not a cent to confirm he is actually a striker rather than a winger pushed inside because the squad was short.

The cheapest bad record is the one nobody checks.

Ali Dia is not the only case. Carlos Henrique Raposo, known as Carlos Kaiser, signed for Botafogo, Flamengo, Fluminense and Vasco da Gama — four of Brazil's biggest clubs — across nearly two decades, while barely playing a professional match. His real skill was manufacturing false data: he faked injuries to avoid training, used his own phone to call club press offices posing as a European club's representative to force contract renewals, and posed for photographs with Romário and other stars, using them as evidence of a career that did not exist.

The book "Kaiser: The Greatest Footballer Never to Play Football" (2026) documents the whole chain. Reading it, I do not see a brilliant conman. I see a weak verification system. Kaiser never needed to breach a server. He needed a photograph and a phone call.

What troubles me is how small the distance is between those stories and how clubs operate today. Modern recruitment has a thousand times more data than Souness had, but the decision structure still rests on a chain of trust: a credible recommender, a credible confirmer, a credible signatory. When all three are credible and the underlying record is wrong, nobody catches it.

Consider the tool nearly every recruitment department uses: Transfermarkt. Market values there are community-contributed and moderated, but they are estimates. They differ from transfer fees in one fundamental respect: a fee is what a club actually paid, while a market value is what a group of people believe a club should pay.

The two numbers frequently sit in the same spreadsheet. When they do, the model stops measuring ability. It measures crowd expectation, and crowd expectation is shaped by journalism.

Numbers do not lie, but the people who produce them always have motives.

CIES Football Observatory does this better: it publishes its own valuation model with an open methodology and explicit assumptions. But even CIES must assume a league, an age, a remaining contract length. Every assumption is a place where a label can go wrong, and every wrong label propagates through the entire calculation chain.

One concrete case I rebuilt in a spreadsheet: a striker with 18 goals in a European second tier. The record labels him "top scorer". The record does not say that six of the 18 were penalties, that the league has one of the worst defensive standards on the continent, and that 11 of the 18 came in four matches against the bottom three. The club buys the number 18. Six months later it discovers it paid for a label, not a player.

The financial layer makes labelling errors far more expensive than they look. When Juventus reported a loss of 89.7 million euros for 2026-20, with Cristiano Ronaldo's net salary reported at around 31 million euros a year, most analysis focused on revenue lost to empty stadiums. The harder problem sat in transfer amortisation: a 50 million euro signing on a five-year deal places 10 million euros on the balance sheet every year, whether or not the player features. If he does not feature because he was mislabelled from the start — wrong position, wrong league, wrong tactical fit — that 10 million still sits there, and still counts against financial fair play.

A contract has three truths: the seller's, the buyer's, and the writer's.

Those three truths rarely align, and my job lives in the gap between them. The selling club leaks to inflate the price. The buying club leaks to reassure fans that work is being done. The agent leaks to invent a rival bidder who does not exist. Three sources, three motives, one headline.

The lesson on source quality I took from that mislabelled data package is simple: information must be classified by level of verification, and the label "unconfirmed" must be preserved rather than deleted for a cleaner headline. A story no authority has concluded must be recorded as inconclusive. A transfer not yet registered with a federation must be recorded as incomplete. Stripping out that qualifier is the fastest route from a three-minute rumour to a three-month fact.

I also took from that report a concept scouting still lacks: quarantining contaminated data. When a record shows anomalies — age, position, league, provenance — it should be pulled out of the training set and out of the valuation model until re-verified. Almost no club does this systematically. They delete only once it is too late.

And this is where I want to linger, because it is the industry's largest blind spot.

The most mislabelled metric in football is not market value. It is possession.

Based on my experience watching matches in Serie A and across European youth competitions, I have repeatedly seen a side hold 62 per cent of the ball and lose 2-0 while nobody understands why. The passing map explains it: most of those passes are sideways between two centre-backs and a deep midfielder, in areas with no pressure and no value. The 62 per cent does not measure control of the game. It measures that the team passed a lot. Two entirely different things, sharing one label.

Once the label "controlling the game" is attached to the number 62, it enters the punditry, the scouting report, and the decision to buy a central midfielder who is good at sideways passing. The error is not in the data. It is in the labelling.

The same mechanism explains why Aleksandr Golovin joined Monaco in 2026 for around 30 million euros after Russia reached the World Cup quarter-finals — several times his pre-tournament valuation. Nobody bought Golovin because of one tournament. They bought the label "star of the surprise national team", applied within four weeks.

Conversely, Luka Modrić's Ballon d'Or-winning run at that same World Cup generated no significant transfer, simply because Real Madrid held absolute negotiating power and had no need to sell. Same stage, same media exposure, opposite outcomes. A label is only worth what someone will pay for it.

A player's value exists only until someone dares to pay it.

In 2026, when Paul Pogba returned to Manchester United from Juventus for a recorded 105 million euros, the statistical model I had built valued him at around 72 million. The 33 million gap was not error. It was brand, expectation, the story of the returning son. No model can price a story, and no model should try. But a model can state clearly: this portion is narrative, not ability.

That is the entire difference between a good scouting report and a handsome spreadsheet.

The industry does have tools to do this at system level. FIFA Clearing House, launched in 2026, centralises training data and solidarity payments in a single database, ending the loss of records when players cross borders. Before it, many small clubs lost money simply because no record proved they had trained a player. Labelling failure, at macro scale, is real money lost.

But tools only fix the data layer. They do not fix the culture.

That culture shows up here: when a bad record appears, the default response is to adjust it to fit the existing conclusion rather than revisit the conclusion. When a 40 million euro signing performs badly, the default is to blame the player or the coach, not the record that made them believe he fitted.

I have seen this repeat often enough to believe it is structural, not incidental. Clubs do not buy players. Clubs buy a set of labels about players, assembled by people with their own motives, and pay for that set of labels.

And here I have to say the thing most recruitment departments do not want to hear.

The whole industry spends money answering "how good is this player", when the more expensive question is "are we naming correctly the thing we are measuring".

A false negative — missing a good player — costs only an opportunity. A false positive — buying a player because of a bad record — costs five years of amortisation, a wage slot, a place in the squad, and sometimes a manager. Yet the transfer market is designed to fear false negatives more. Nobody is sacked for missing Golovin. Plenty are sacked for buying wrong.

That is the paradox: the fear of missing out is precisely what pushes clubs into their worst contracts.

A three-minute call can kill a three-month negotiation.

And a three-minute call can also create a contract that, three months later, nobody admits to having approved.

The one thing I can state with high confidence concerns my own trade. We are the amplification layer. When an anonymous source says a deal is done, we publish it as an event. When the club denies it, we publish that as a new development. Both reports are technically accurate and informationally wrong, because both describe a state that does not yet exist.

The only way I know to resist that is to keep the verification level inside every sentence, even when the headline becomes less attractive. A headline containing "according to two independent sources" is always less fun than one containing "official". But the price of a fun headline is a contaminated data system, and that system will keep mispricing the next players.

What I want to leave behind is not a call for transparency. Transparency is something everyone endorses and nobody defines. What I want to leave behind is a small change in the order of work: verify first, label second.

If recruitment departments spent one tenth of their data budget re-verifying old records — age, position, league, provenance — they would discover that a meaningful share of their models runs on a false foundation. That would not immediately help them buy better players. It would merely stop them buying players they never meant to buy.

In a market where every club has the same data, the only remaining edge is knowing which of your data is wrong.

I do not write about contracts. I write about partings.

And most partings in football begin with one wrongly filled cell, on a Friday afternoon, entered by someone whose name nobody remembers.

The question I leave for any recruitment department reading this: if your system were wiped tomorrow and only records verified by two independent sources survived, how many players would you have left — and of those, how many do you genuinely believe you have named correctly?