Verifying Football Data: From 1,204 Shots in Ligue 1 to the Empty Tables Nobody Wants to Read

In Marseille, summer is measured not by calendars but by the hum of ceiling fans. For me, however, the summer of 2026 began with a spreadsheet containing 1,204 rows. Not lines of text. Lines of numbers. Each row was a shot in Ligue 1 during the first half of the 2026-18 season, recorded by my own hand, cross-checked against the actual goals scored on the pitch.

I was 57 then. People said my reaction was slow. Opta had already released xG tables for Ligue 1, forums were arguing about them, English journalists were using xG like a mantra. I was still sitting there, rewinding every passage of play, counting every shot, noting every position, every angle, every defender's pressure. Verification comes before use. That is the first principle, and also the last principle of this profession.

The correlation coefficient I calculated was 0.84. Enough to start building my own striker valuation dataset. Not enough to call it a truth. The distance between those two things is the entire subject of this article.

In the summer of 2026, I learned to trust something nobody had yet named: xG. And the accompanying lesson was harder: trust it, but never worship it.

Context: an industry learning to speak before it learned to count

To understand why I spent half a year verifying a single metric, you have to understand the landscape of European football in the 2010s. Football entered the data era not with an explosion but with a slow bleed. Statistic companies like Opta, StatsBomb and Wyscout expanded event collection from every match to every half to every touch. Data became merchandise. And like all new merchandise, it was sold before it was certified.

As a transfer market administrator, I saw this daily. A striker scores 12 goals in half a season, his price triples within two weeks. A midfielder has a 92% pass completion rate, the papers call him an "architect," while 60% of his passes are sideways deliveries under 15 metres with zero pressure. Those numbers are not wrong. They were simply asked the wrong questions.

I am 66 years old, and I have sat at this table long enough to watch three waves pass. The first wave was goals and assists — numbers from the previous century, simple, easy to grasp, and easiest of all to mislead with. The second wave was action metrics — passes, tackles, pass completion — measuring what a player does, but not whether what he does has value. The third wave is model metrics — xG, xA, PPDA, progressive passes — measuring value, but demanding that the reader understand sample size, confidence intervals and context.

The problem in this industry is not the third wave. It is that most consumers of data are using third-wave tools with a first-wave mentality: one number, one conclusion, one headline.

Based on my experience following matches in Ligue 1, Ligue 2 and European competitions across nearly four decades, I have drawn a fairly simple conclusion: new metrics are not automatically more correct than old ones. They are merely capable of being wrong in subtler ways, if the user does not verify. A 92% pass completion rate can be dismissed with the naked eye. An xG of 0.08 for a shot is much harder to dismiss, because it has been labelled "scientific."

That is why I spent half a year on 1,204 shots.

The core: a chain of evidence, not a chain of conclusions

Step one: verifying xG against the match itself

When Opta released xG tables for Ligue 1 in the 2026-18 season, I set three questions before believing anything.

First: what is the sample size? An xG figure computed from 50 shots has no statistical meaning at club level. I needed at least several hundred shots per team for the coefficient to stabilise. In the half-season I surveyed, each team averaged roughly 60 recorded shots in my sample — not enough for a club-level conclusion, but sufficient at league level when pooled.

Second: is match context included? A 25-metre shot while trailing 0-2, with the opponent already sitting deep, carries an entirely different probability to the same shot at 0-0. If the xG table does not state its contextual variables, I have to verify by rewinding the tape myself.

Third: what is the confidence interval? This is the question almost nobody asks, and it is the most important one. A team with 1.4 xG against a team with 1.1 xG in a single match — that gap sits inside the noise. Publishing an article claiming "Team A attacked better" on a 0.3 xG gap is unprofessional conduct, not analysis.

After recording and cross-checking, the correlation between xG and actual goals in my sample reached 0.84. That is a strong figure. It allows me to use xG as a valuation tool, not as a prophecy. That is the entire difference.

I built my striker valuation dataset on three layers: xG per 90 minutes (chance quality), xG per shot (shot selection), and the gap between actual goals and xG (finishing above or below expectation). The third layer is the most misunderstood. A striker with 15 goals and 10 xG in a season is not an "elite striker." He is a striker at the peak of a variance cycle, and that cycle usually does not repeat the next season. Transfer market history is full of large contracts signed at the peak of such cycles.

Here I want to state something about method clearly, because it relates directly to how I have read every table of numbers since. When you have an empty data series — no information, no club, no player, no dates — the correct thing to do is explicitly write "insufficient information to assess," not fill the gap with speculation. In my profession, an honest empty table is worth more than a full table of invented figures. Readers usually dislike this. But readers are usually not the ones who bear the consequences.

Step two: PPDA and the 2026 World Cup test

Thanks to the dataset accumulated in Marseille, a sports newspaper invited me to contribute for the 2026 World Cup. I was 58. I followed all 64 matches and counted PPDA — the passes an opponent is allowed before each defensive action.

PPDA is a beautiful concept. It measures pressing intensity by counting backwards: the lower the PPDA, the more aggressively a team presses. But it has a blind spot few notice.

The first blind spot: low PPDA cannot distinguish organised high pressing from a team simply chasing the ball in vain. Both produce low PPDA.

The second blind spot: PPDA is heavily distorted by game state. A team leading 1-0 and deliberately sitting deep will have a higher PPDA than a team trailing and forced to push up, even when their pressing capacities are identical.

The third blind spot: PPDA does not measure the quality of the defensive action. A tackle that wins the ball in midfield and an air-swing at the edge of the box both count the same if they fall within the definition.

In the Croatia-England semi-final, my recorded figure was Croatia allowing England 8.2 passes per defensive action, while England allowed Croatia 12.5. I wrote a piece predicting Croatia would win through sustained pressing into extra time. Croatia won 2-1, with the decisive goal coming in the second period of extra time.

I did not shout in celebration. I reopened the spreadsheet to look for outliers. This is a professional reflex I built after years of being fooled by myself: a correct prediction does not prove a model correct. It only proves the model was not falsified in that particular instance. The distinction sounds small. In practice it is the boundary between an analyst and a fortune-teller with a spreadsheet.

Croatia won a tournament of low PPDA? Then PPDA is merely one letter. At the 2026 World Cup, the winning team France did not lead any pressing metric. France won through defensive structure, transition speed and timing. No single metric tells that story. To tell it, you must layer multiple datasets and read the intersection.

After the tournament I changed how I wrote. I moved from describing matches emotionally to presenting PPDA tables, distance covered and successful pressing counts. The phrase "the team pressed better" appears only when the numbers genuinely support it, and always with game-state context attached. Some colleagues found that style dry. I accept that. Football does not lack good writers. Football lacks verifiers.

Step three: empty stadiums as a natural laboratory

In 2026, the pandemic closed the stands across Europe. An editor assigned me to follow the Bundesliga when football restarted, because of my data experience from the 2026 World Cup. I was 60, sitting in Marseille, analysing 81 matches played in empty stadiums in the 2026-20 season.

The result: home teams won only 26% of matches, against roughly 43% before the pandemic.

An empty stadium is the finest laboratory for anyone who loves data. Home advantage is one of the most studied variables in football statistics, and for decades people argued about where it came from. There are three main hypotheses. The first: the crowd influences referees, producing decisions more favourable to the home team. The second: the crowd affects the psychology of away players, raising tension and lowering performance. The third: the crowd affects home players, raising motivation and intensity.

When the stands emptied, all three hypotheses were eliminated at once. And home advantage collapsed by nearly half. This is one of the cleanest pieces of evidence the football data industry has ever had.

I wrote a report whose title said empty stadiums kill home advantage. It was widely read. A Ligue 2 club, Le Havre, used it in a negotiation to buy a young striker. The player had strong scoring figures at home, but most of his goals came before the lockdown period. The club used my numbers to lower the price. The deal still happened, just at a smaller fee.

That was the first time I saw clearly that data analysis is not only for understanding matches. It is for pricing. And once it is used for pricing, it is no longer neutral.

After that report, every statistical table in my articles separated home and away figures. I reminded readers not to trust pre-lockdown records when evaluating a player. And I reminded myself that every dataset has a point in time after which it stops being true.

Step four: the inverted full-back and the 34% gap

The empty-stadium report reached Canal+, so they sent me to Qatar for the 2026 World Cup when I was 62.

When the pundits praised Achraf Hakimi — 142 sprints and roughly 2.3 chances created per match, in a hybrid full-back and winger role — I dug into positional data for the player across the tournament and found something troubling.

The corridor behind Hakimi was unoccupied for roughly 34% of playing time. In other words, for more than a third of his time on the pitch, the space behind his position was not protected by him.

This is not a flaw in Hakimi. It is the nature of the role he was given. A full-back who pushes high creates space behind. That is the conservation law of space. No tactic creates space out of nothing.

What kept Morocco safe early on was that their centre-backs had exceptionally high covering speed — in my tracking data, defenders were recorded above 31 km/h in recovery runs. This is the compensating variable. Without it, the system fails.

I wrote a note warning that this system only holds if the back line is fast enough to compensate. In the match against France, the opponent attacked the Moroccan right flank relentlessly, and that space became an exploitable factor.

The lesson I took for my own method: do not praise a new tactic without listing the compensating variables. In every analysis since, I have tried to state the necessary and sufficient conditions for a system to function. A system that only works with the right players is not a system. It is a dependency.

I remember that night in Doha, sitting in the stand, opening my phone and writing a line in my personal notes: "Everything I praise today must be re-checked tomorrow against the list of conditions."

The counter-intuitive angle: correlation is not causation, and an empty table is not a failure

Here I must address the hardest part of the profession, the part nobody wants to hear.

The entire chain of evidence above — the 0.84 xG correlation, the 8.2 versus 12.5 PPDA, the 26% versus 43%, the 34% gap — consists of strong numbers. But none of them is causal. They are all correlations. And the gap between correlation and causation is where most modern football analysis collapses.

Take the empty stadiums. The collapse of home advantage in the 2026-20 season could be due to the absence of crowds. But three other hypotheses need consideration before concluding.

The first: the schedule was compressed after lockdown, and high fixture density affected all teams differently depending on squad depth. Smaller clubs, with thin squads, suffered more. Since smaller clubs are often not the home side in major fixtures, this effect may have been misattributed to the crowd factor.

The second: the change in substitution rules, from three to five, altered second-half dynamics in ways unfavourable to home teams that traditionally relied on fitness.

The third: the crowd may affect not only players and referees but the goal frame in a psychological sense — meaning the effect may exist but be far smaller than the raw 26% versus 43%.

I presented all three hypotheses in the original report. Nobody quoted them. People only quoted the number.

This is a law of the industry I have observed for nearly fifty years: the numbers get cited, the conditions get dropped. A metric with conditions attached does not spread. A metric stripped of conditions spreads very fast, and at some point becomes a belief.

Players are variables, the market is a function, but most of my life has been a constant. I sit in one place, re-read the data, and try not to join the loud arguments. My job is to check whether the function still holds with the new variables.

And here is the truly counter-intuitive part, the part I want readers to carry away.

In football data analysis, the greatest failure is not drawing a wrong conclusion. The greatest failure is drawing a conclusion from an empty dataset without noticing it was empty.

I have seen this many times. A report presented in perfect formatting, with section headings, tables, checkboxes and a conclusion. On the surface, professional. But every content field is empty, every data row absent, every judgement resting on an unverified source. Perfect format. Hollow content.

In this profession we have a rule: if there is insufficient information, state that information is insufficient; never fill it with speculation. The reason is simple. An empty report labelled "fully analysed" is more dangerous than a report that admits missing data, because it carries false authority.

I have seen this in transfer deals. A club receives a report on a player, complete with sections on technique, fitness, mentality and adaptability. But the author watched three matches on video and told nobody about the three-match sample size. The club signs the player. The player fails. And nobody can trace it back, because the report looked so complete.

That is why I spend most of my time on what nobody sees: checking the empty cells.

A little about my method, written down so others can use it

I am 66. I do not have much time left for long articles, so I want to package the method here so it does not disappear with me.

Principle one: every metric must come with its sample size. No sample size, no judgement.

Principle two: every metric must come with context. The same shot in two different game states is two different events.

Principle three: every gap must be tested for whether it lies outside the noise band. A 0.3 xG gap in a single match is not enough to say which team attacked better.

Principle four: every dataset must state its time window and collection conditions. Data collected before lockdown cannot be used to evaluate a player in a season with crowds, and vice versa.

Principle five: before concluding, write out at least three alternative explanations. If you cannot write three, you have not understood the problem.

Principle six: when a prediction proves correct, go back and look for outliers before celebrating.

Principle seven: when the data is empty, say it is empty.

Principle eight: do not use emotional language to describe data. Emotion belongs to the reader. The writer must keep the cold part.

Principle nine: a single metric is never a conclusion. It is one letter. To read a sentence, you need several letters in the right order.

Principle ten: data analysis is not neutral when it is used for pricing. The analyst must know whom he serves.

These ten principles are not science. They are discipline. Science is the method; discipline is doing it on the three-hundredth day when nobody is checking anymore.

Where modern football is heading with data

I want to spend a section on what I have observed in recent years, and it is somewhat concerning.

Football data is gradually shifting from describing matches to describing people. Current models do not only compute the probability of a shot succeeding — they compute the probability of a player improving, of a young talent peaking, of a transfer turning a profit.

From one angle, this is professionalisation. From another, it is turning human beings into variables in a financial function.

I have seen valuation models for 17-year-old players based on 400 minutes of professional football. Four hundred minutes. That is fewer than five matches. Applying the ten principles above, such a model must be explicitly marked as having insufficient sample size. But it survives, because the market needs a number to start negotiating.

This is the intersection of data analysis and sports business, and it is why I always tell younger colleagues: know whether your data is serving a sporting decision or a financial one. Those two decisions require two different standards. Financial reporting pressure usually bears down on sporting decisions, and when that happens, the data model gets bent to serve a conclusion that already existed.

I see the same in esports, which I followed for a time out of curiosity about how data is used in an environment that was born digital. A mouse click on an esports screen carries the shape of a pass. It has a position, a timing, a pressure, an expected value. And there I saw the same problem: professionalisation turns players into an assembly line product, and individual play gets worn smooth in digitalised training.

That is a subject for another article. Here I only note it to remember that the problem is not data. The problem is what data is used for.

What I am tracking for the next cycle

The next major tournament cycle is approaching, and I have started taking notes.

First, I am tracking how national teams handle fixture density. Expanded tournaments mean more matches in the same window, which means squad depth becomes the decisive variable. In my historical samples, at high-density tournaments the champion is not the team with the strongest eleven, but the team with the smallest gap between starters and substitutes.

Second, I am tracking how national teams handle pressure from metrics. When a player knows his xG is low, his shooting behaviour changes. He starts choosing higher-xG shots, meaning fewer long-range efforts, meaning fewer spectacular goals. This is an effect I call "visible optimisation," and I believe it will appear more clearly in this cycle.

Third, I am tracking national teams from high-density domestic leagues. Their players enter the tournament in a state of accumulated fatigue. Minutes played in the three months before a tournament is one of the most predictive variables I have, and it is almost never mentioned in prediction pieces.

Fourth, I am tracking teams that change systems mid-tournament. Every mid-tournament system change requires bedding-in time, and at a tournament with few matches, that time usually does not exist.

Fifth, and perhaps most importantly, I am tracking refereeing decisions and their impact in environments with no crowds in some matches, or crowds of varying intensity in others. I learned from the 2026-20 season that the crowd factor does not vanish. It only shifts from one form to another.

I record all of it, and I leave room for what I do not yet know. That is how I work. I am 66, old enough to know a number never tells a story unless we ask. And the most important question in my profession, after nearly fifty years, is still the first one: where did this data come from, from how many samples, under what circumstances.

An empty spreadsheet is not a failure. It is a reminder. It reminds us that football still has a great deal that has not been measured, and what has not been measured cannot be summed up in a headline.

I will keep taking notes. And if a new metric appears in the next round of matches, I will spend half a year learning to trust it — as I spent half a year on 1,204 shots in Ligue 1, the summer when I was 57 and first learned that trusting a number and understanding a number are two entirely different jobs.