Football Data Integrity: When a Complete Report Contains No Evidence

A late evening at the end of the month: a 4,200-word file arrives in a sports desk inbox. A match analysis from a national league — nine sections, thirty-two tables, every table carrying numbers, every number carrying a unit. Pressing intensity, chance-creation passing rates, revenue structure, pressure on the coaching staff, probability of breaching financial fair play, injury risk, media cycle, industry transmission chain. Not a single section missing.

When I traced every column back to its origin, the answer was identical across all thirty-two tables: the system had never read the source article. It did not invent numbers. It filled the gaps with something harder to detect than invention — structure.

The editor had no way of noticing. Every cell was populated. Every heading was grammatical. Every table was aligned. An empty product had learned to wear the clothes of an authoritative one.

The machine that produces reports

Sports content has become a formatting industry. Fifteen years ago, a match analysis in Vietnam needed three things: the narrative, the opinion, the prediction. Today a piece that qualifies as complete must carry an xG table, a heat map, a PPDA figure, transfer fees, the wage bill, head-to-head history, and usually an injury-risk section as well.

That demand does not come from readers first. It comes from algorithms. Since 2026, search tools have graded sports content on “information gain” — how much new information a piece adds beyond what already exists online. A recap of a match everyone watched carries no value. To survive, a newsroom has to add data, comparison, forecast.

The gap between that requirement and real capability is bridged in two ways. The first is hiring people: one editor covering twelve matches per round, entering data by hand, cross-checking everything. The second is hiring systems: a model reads the source article, extracts entities, queries a database, then outputs a nine-section analysis.

The second route is roughly two orders of magnitude cheaper. That is why it is everywhere.

The problem sits in a technical detail almost nobody explains to the desk. In data-pipeline design there is a rule called null handling. When a system cannot find information, it has two choices: fill in a guessed value, or return an explicit marker that the information does not exist.

The second school is more honest and is treated as more useless. A nine-section report where every section reads “insufficient information” is methodologically correct and impossible to sell to anyone.

The failure signature is distinctive. The domain label is populated correctly — football. The title is blank. The source is blank. The article type is filed as “unclassified”. The entity list does not exist. The most plausible explanation: the classification step ran, the extraction step did not. Half the pipeline alive, half dead — and the dead half is the decisive one.

Four kinds of silence

The most forgivable silence is a genuinely empty source: a wire stub with a headline and no body. The system returns “insufficient information” and it is right. The fault belongs to whoever commissioned the work, not to the machine.

Another kind is a source with content where extraction fails. The original runs two thousand words; the system retrieves only the domain label. The fault is in the pipeline, and it is usually invisible because the output still looks tidy.

Then there is the source that extracts cleanly but is lost at the formatting stage. Entities exist, numbers exist, but they are pushed into the wrong cells or truncated when the report exceeds its length limit.

The final kind is the most dangerous: a complete source and a system that speculates. No technical error at all. Only a model trained to always have an answer.

A report can be truthful in every cell and still lie in its overall shape. The lie is not in the numbers; it is in the layout. Readers do not inspect each cell and form a judgement. They see a structured page, and the brain defaults to assigning that structure a level of credibility.

Data tables know how to speak; few people have the patience to listen. But one more thing needs saying: an empty table says nothing at all, and the table is the part everyone sees.

The craft of grading sources

The transfer market is where this skill is paid for most visibly. Consider three deals that generate the same headline format.

Enzo Fernández left Benfica for Chelsea on 31 January 2026 on an eight-and-a-half-year contract for a fee of around £106.8m — at the time a British record. Moisés Caicedo left Brighton for Chelsea in August 2026 for a fee of around £115m. Kim Min-jae left Napoli for Bayern Munich in July 2026 through a release clause worth around €50m.

Three entirely different mechanisms. The first was the product of a negotiation that ran to the final minutes of the window. The second was an auction with multiple bidders. The third was a line in a contract that triggered itself, requiring no negotiation at all.

A source-grading system that treats all three as the same category is not grading. It is decorating. The right question is not “where did this come from”, but “what mechanism produced this, and who has an incentive to leak it”.

Every figure I dig up buries a myth that the media created.

The metric layer

xG — expected goals — measures the quality of a chance based on the probability it becomes a goal. xA is its counterpart for the final pass. PPDA, passes allowed per defensive action, measures pressing intensity: the lower the figure, the higher the press.

These three metrics have become a shared language. Which is exactly why they are the most abused.

A single match's xG says nothing without shot volume attached. A low PPDA does not prove a good press if the team is chasing the game and forced to push up. More importantly, different data providers use different models for the same shot, and two tables from two sources can diverge enough to reverse a conclusion about a player.

My experience at the 2026 World Cup is an example of the metric layer's power. On 27 June 2026 in Kazan, South Korea beat Germany 2-0. Before kick-off I published an analysis based on Germany's group-stage pressing data, concluding that their back line was being pulled open. Based on my experience of watching matches, it was the first time I saw a defensive metric tell the story the eye had missed.

I was mocked for ninety minutes. The piece was then shared several hundred times over.

But precisely because I believe in that power, I have to state the reverse: a correct metric in football is one that comes from a published model. A metric with no model behind it is just a coloured drawing.

The governance layer

Financial fair play and profitability-and-sustainability rules operate on a different logic entirely. Here, entities are always named.

Everton were docked ten points in November 2026, reduced to six on appeal in February 2026. Nottingham Forest were docked four points in March 2026. Manchester City face 115 charges dating from February 2026. Juventus have previously been sanctioned over misstated financial accounts.

The striking feature of this news category: silence has never meant compliance. A piece about club finances that names no club, no governing body and no applicable rule framework is not under-informed — it is the wrong genre.

That is why a nine-section analysis with an empty “compliance risk” section can still do damage. Readers see the blank, and part of them reads the blank as “nothing is wrong”. The actual cause is simpler: the system had no data.

The reverse transmission chain

During the behind-closed-doors season of 2026, I collected data from more than 130 matches in the K League and the Bundesliga to test the home-advantage hypothesis. The result showed the gap between home and away performance narrowing sharply with no crowd present. I published the raw dataset and invited verification within forty-eight hours.

The first response was not an analysis. It was an accusation: someone said I had fabricated the numbers.

They called me a data cheat because they could not call me wrong.

That story repeats at larger scale when machine-generated analysis enters an editorial system. An extraction error becomes a sentence in an article. A sentence becomes a supporter's belief. That belief becomes a betting line. And the betting line becomes a reason for a club to change coach.

None of those links is football. All of them are information.

The industry is worrying about exactly one half of the problem. The half it worries about is the model inventing players, inventing scorelines, inventing transfer fees. That is the easiest error to catch, because it breaks a fact verifiable in thirty seconds.

The half it does not worry about is counterfeit honesty. A report can be correct in every individual cell — because every cell reads “insufficient information” — and still cause identical damage, because its shape is read as a conclusion.

I may be wrong here. Perhaps nine-section frameworks exist for good reason: they force an analyst through a checklist, and for a newcomer a checklist beats a blank page. Perhaps in many cases the source genuinely contained nothing — a headline-only stub — and the system behaved correctly by returning an empty marker. In that case the fault lies with whoever paid for an analysis built on a headline.

And perhaps I am asking for too much. Football has never been an industry that demanded traceability. Supporters go to the ground for emotion, not for data provenance.

I do not need anyone to agree with me. I need someone good enough to argue back.

What will change

Source labels will be attached directly to individual figures — not a footnote at the bottom, but a convention: every metric accompanied by model name, provider and sample size.

At the same time, publishing an empty marker will be treated as a valid result. A section stating “insufficient data to conclude” will stop being read as a bad article. It will be read as a checked article.

What remains depends on readers. When every outlet has xG and PPDA, those metrics stop being an advantage. What is left is the ability to explain how a number was produced.

I once wrote that the crowd is always safe, and that is precisely why the crowd is always mediocre. In data analysis, the mediocre version of safety is a table packed with figures that nobody dares ask about.

If a data table cannot be argued with, is it doing the job of a poster, or of evidence?