When Football Analysis Outruns the Evidence: A Lesson in Data Integrity

On an empty Anfield in February 2026, there were no chants and no whistles. Liverpool lost a fifth consecutive home match, a run that had never happened under Jurgen Klopp. That night I stayed up until three in the morning with a spreadsheet, and what kept me awake was not the scoreline but a professional question: where is the evidence, and how much of the conclusion I am about to write does it actually hold up?

Liverpool's PPDA, the number of passes allowed per defensive action, rose from 8.2 to 12.5 during the behind-closed-doors period. On the surface, that is evidence of collapse. But with only that number in hand, I would have written a wrong piece. A rising PPDA does not automatically mean pressing weakened. It can mean the back line pushed higher without crowd noise to synchronise it, or simply that opponents chose longer passes to avoid the pressure zone. The empty stadium taught me that noise is data.

I entered football analysis with one very specific fear: writing a conclusion that sounds certain with nothing underneath it.

In 2026, when I was eighteen and a first-year sociology student in Guangzhou, I watched the Russia World Cup with a notebook. The quarter-final between France and Uruguay changed how I saw the game. France controlled 39 percent of possession but generated 2.1 xG; Uruguay generated 0.4. The winner was not the team with more of the ball. I spent the next three weeks rewatching matches and building my own xG tables, in direct opposition to the emotional commentary in mainstream media.

xG, expected goals, is a model estimating the probability that a shot becomes a goal, based on location, angle, shot type and context. It does not measure goals; it measures chance quality. xGA is the xG a team concedes, used as a defensive-process metric. PPDA measures pressing intensity, where lower values indicate more aggressive pressing.

By 2026, these metrics are no longer obscure. Every transfer report, every post-match column, every television breakdown cites them. FBref, Understat and StatsBomb are household names for supporters. Football has acquired a new layer of language, and that layer sounds highly professional.

At the same time, a new problem has appeared that few people name. The more people learn to use data, the more conclusions are generated from numbers with no provenance. A metric gets quoted without anyone rechecking it. A model gets invoked without anyone knowing how it was built. A number travels faster than the speed at which it was verified. And behind all of it, a question gets skipped: if you strip every statistic out of the piece, how much real evidence remains?

That is the paradox of football's data era. We have more numbers than ever, and more conclusions lacking evidence than ever.

Process metrics are not verdicts

When I wrote about Euro 2026, the temptation was enormous. Federico Chiesa scored two goals and provided one assist on Italy's run to the title. The media called him the breakout star of the tournament. But when I opened the data, the picture looked different: Chiesa's xG was only 1.8 across five matches, while he scored two. His shot-on-target rate stood at 41 percent, below the average for top European wingers.

Two goals above xG is not an achievement; it is a variance signal. A player who outscores xG across a short tournament is usually running ahead of the model, and the model tends to catch up. I wrote a 2,000-word analysis for my personal blog concluding that Chiesa's performance was unlikely to be sustainable. The following season, he suffered an injury and his form dropped.

I tell this story not to congratulate myself. I tell it because it illustrates a principle: data does not confirm the story the media wants to tell. It only offers a different reading. Every number tells a story. The story is not in the number.

But here is the reverse side of that same principle. With the same dataset, if I treated Chiesa's xG as absolute proof, I would also be wrong. xG depends on the model, and different providers produce different numbers for the same shot. That is why I cross-check at least two sources, usually FBref and Understat, sometimes adding StatsBomb, before writing anything. A number from a single source is a hypothesis, not a fact.

Process metrics such as xG, xGA and PPDA are extremely useful. But they are tools, not verdicts. They describe possibility; they do not judge outcomes. The gap between those two things is where errors are born.

The transfer market, where impatience is priced

Now to the financial side. This is where numbers without provenance do the most damage, because they shape the decisions of real people handling real money.

A transfer fee is accounted for through amortisation, meaning the fee is spread evenly across the contract's years. A player worth 80 million euros on a five-year deal goes into the books at 16 million euros a year. That number appears in no transfer report, yet it is what determines a club's financial health, not the headline figure.

Here I hold a fairly firm position, and it comes from a simple observation: signing fees for free agents are more corrosive than transfer fees, because they sidestep the core scrutiny of FFP. When a player's contract expires, the money that would have gone to the selling club is instead paid directly to the player and the agent as a signing-on fee. That payment is not a transfer fee, so it is not amortised across the contract, and it frequently fails to appear in standard aggregate tables. It disappears from view.

FFP is UEFA's financial rulebook; PSR is the Premier League's equivalent, capping the losses a club may record. Both systems rest on reported numbers. If a large payment can be structured so that it falls outside the definition of a transfer fee, the rule remains correct on paper while reality has shifted. This is not a conspiracy theory. It is a definitional problem.

My rule when reading a deal is to separate three numbers: fee, wages, and signing-on payment. Reports usually contain only the first. The other two hold most of the story.

Contract years and the new-manager bounce

Two recurring patterns in the data deserve close attention, and both are examples of correlation misread as causation.

The first is the contract year. Players entering the final year of a deal often show form volatility. The phenomenon is widely acknowledged, but the popular explanation is suspect. The popular explanation is that players try harder to earn a renewal. The better-evidenced explanation is small sample size, heavy media pressure, and selective attention magnifying a handful of good performances. Before concluding, I always ask what the sample size is.

The second is the new-manager bounce: short-term improvement in results after a coaching change. The effect is real and relatively easy to measure. The problem is durability: it typically lasts a few matches and fades, while media presents it as a tactical turning point. What I track is not the first three results but the structure of process data after ten matches.

People see form. I see sample size.

Precedents any analysis must know

No discussion of data integrity in football can ignore the cases that shaped the current rules.

Everton and Nottingham Forest were docked points in the Premier League for breaching PSR. Manchester City face 115 charges relating to the league's financial rules. Juventus were sanctioned in a case involving player valuation and capital gains. These three cases differ in nature, but they share one important feature: in every instance, what was contested was not only the number itself but how the number was created and reported.

At the transfer level, FIFA bans TPO, the model in which a third party owns a player's economic rights. FIFA's Article 19 restricts the international transfer of minors. And multi-club ownership groups, such as City Football Group or the Red Bull network, create eligibility conflicts in European competition when two clubs under the same ownership could meet.

What all these structures share is that they make money harder to trace. When money is harder to trace, public data becomes less reliable, not because the data is wrong, but because it no longer fully describes reality.

There is also a mechanism few people notice. FIFA's solidarity mechanism distributes a share of transfer fees to clubs that trained a player within certain age brackets. The existence of the mechanism is the system's admission that value is created before the first professional contract is signed. Yet in most transfer analysis, that portion of value is entirely absent.

The empty-input principle

This is the part I want to weight most heavily.

In analysis, there is a professional temptation I call laundering an empty input. You are handed an incomplete dataset, a record with unclear provenance, a case in which every metric is missing. The temptation is to write a plausible-sounding conclusion to fill the gap. The product of that temptation is the most dangerous thing in the trade: confident prose anchored to nothing.

My rule is simple and testable. If no entity has been identified, with no club, no player, no competition, no transaction, then any conclusion about tactics, finance or governance is fabricated attribution. The correct behaviour is not to write something to fill space, but to stop and state clearly that the input is insufficient.

I know this sounds unglamorous. The profession carries enormous speed pressure. The Chinese market, where I work, especially favours controversy and pace. A shocking hot take will travel; a piece saying there is not yet enough data to conclude gets ignored. But the long-term value of a writer lies in being trusted. The moment you publish a conclusion without foundation, that trust is gone, and it is very hard to win back.

Data does not make revolutions. It only strips the paint off legends. But the paint only comes off when the number genuinely has provenance, and when the writer is willing to admit what they do not yet know.

At this point I have to pose a counter-argument to myself, because I belong to the cautious school and I know its weakness: cautious people tend to cling to old models and resist anything new. That is a real trap, and I have fallen into it.

When neural-network models began replacing linear models for estimating xG, my first reaction was scepticism, mainly because I could not read inside the black box. But opacity and error are different categories. I forced myself to test the new models on old data before rejecting them. The results showed that many new models were better at separating chance quality from finishing skill. The lesson is that data analysis is not about doubting everything. It is about testing before believing, and testing before rejecting.

But there is a reverse counter-argument, and it matters more. There is a style of analysis that sounds very data-driven but is really just statistics placed side by side with no hypothesis behind them. You can pile ten metrics into a chart, and the chart answers no question at all. Data is not analysis. Analysis is the process of choosing the right question, choosing the right metric, and, most importantly, choosing what you admit you do not know.

Data does not erase emotion. It explains why emotion exists. A silent Anfield is not a number; it is an explainable variable behind a strange run of results. My job is to turn that emotion into a question, not into a ready-made conclusion.

From one season to the next, the analyst's problem never changes. Data will keep growing, models will keep getting more complex, and the gap between evidence and conclusion will keep getting easier to blur. The decent writer is the one who holds that line.

If there is one signal I want readers to carry into the next matchday, it is this: when a number is quoted and nobody can say where it came from, hold off on sharing it. It may be right. But being right without provenance makes it unusable.