Marseille Data Diary: xG, PPDA and What the Spreadsheet Never Tells You

Marseille, November 2026. The clock on the office wall read 23:40. On screen sat a spreadsheet of 1,204 rows, each one a shot taken by the twenty clubs of Ligue 1 in the first half of the 2026-18 season. The left-hand columns held location, angle, preferred foot, number of defenders in close proximity. The right-hand column held the real outcome: goal, or no goal. At row 847, I finished typing the final formula and pressed Enter. The correlation coefficient appeared: 0.84.

I sat still for a long while. Not out of joy. Because I had just lost an old belief — the belief that a scout's eye, after ten matches in a row, could substitute for a dataset. The number 0.84 did not shout anything. It only whispered that if you close your sample tightly enough, football leaves measurable traces. The hardest part came afterwards: knowing when those traces tell the truth, and when they are merely repeating themselves.

In the summer of 2026, I learned to trust a thing nobody had yet named: xG.

When Opta first published xG tables for Ligue 1, colleagues in the transfer room called it "one more column for decoration". Fair enough. For a century football had been read through three columns — goals, assists, cards — and those three had entered the bloodstream of an entire industry. A new column does not automatically become truth. I did not rush to believe either. But I did something most people in the trade never do: I counted it myself.

That is why I call myself a record-keeper rather than a commentator. A commentator is allowed to say "I saw". A record-keeper must say "I counted, here is the sample, here is the confidence interval". With 1,204 shots from the first half of 2026-18, I split them into two groups: shots taken under direct defensive pressure and shots taken without it. I weighted shot angle by deviation from the goal axis. I discarded efforts from beyond 35 metres because they dilute variance without adding information. After standardisation, the correlation between accumulated xG and actual goals per club reached 0.84 — high enough for me to build my own striker valuation dataset, but not high enough for me to abandon the human eye.

I am 66 years old. I have lived long enough to know a number never tells a story unless you know what question to ask it. And long enough to watch data fashions arrive and depart: eras when people worshipped pass counts, eras when they worshipped distance covered, eras when they worshipped possession share. Each fashion is right in part, and abused in the remainder.

In 2026, when the World Cup was held in Russia, a sports newspaper invited me to contribute after hearing about the Marseille dataset. I accepted on one condition: I would not write about emotion, I would write about tables. Across 64 matches I counted PPDA — the number of passes the opponent is allowed before each defensive action. This metric has a property I admire: it forces you to look at both directions of a match at once. To know how strongly a team presses, you do not measure that team; you measure its opponent.

The semi-final between Croatia and England was where my spreadsheet collided with received wisdom. England were favoured thanks to an easy path and an in-form striker. Croatia were seen as old, slow, having already played two extra times. I opened my table and read the opposite. Croatia allowed England only 8.2 passes per defensive action, while England allowed Croatia 12.5. Croatia were forcing England to pass hurriedly; England were letting Croatia hold the ball far more comfortably than was healthy.

I wrote a short note predicting Croatia would overturn the game in extra time through their ability to sustain pressing as the opponent faded physically. Croatia won 2-1. I did not shout in celebration. I reopened the spreadsheet and went hunting for outliers, because a correct prediction is never evidence that a method is correct. That habit has stayed with me all my life: after every correct call, I search for the crack.

Croatia winning a tournament of low PPDA? Then PPDA is only one letter.

That line is not meant to diminish the metric. It is a reminder that a metric only means something when placed beside other metrics and beside context. Croatia in 2026 did not win the tournament — they reached the final and lost to France — but they went far beyond nearly every prediction based on age and distance covered. Read one column and you conclude wrongly. Read the PPDA column beside extra-time running data, beside the three-man midfield structure, and you begin to see a system rather than a number.

In 2026, European football restarted after the pandemic in empty stadiums. My editor assigned me Bundesliga coverage, a league I was not a specialist in, but he wanted someone with World Cup 2026 experience to handle the analysis. I took the job and chose an unusual angle: treating the empty stand as a natural laboratory. This is the kind of setting a data man dreams of — a variable suddenly removed across an entire continent, designed by no one, lasting long enough to build a sample.

I analysed 81 matches played in empty stadiums in the 2026-20 season and compared them with the pre-lockdown period. The home win rate fell from 43% to 26%. Away goals rose noticeably in the second half. Yellow cards for home teams fell — which is more interesting than it looks, because it suggests part of home advantage comes from invisible pressure on referees, not only from crowds urging players on.

An empty stand is the finest laboratory a data lover could ask for.

I wrote a report titled "Empty Stands Kill Home Advantage". The report did not stay in journalism. A Ligue 2 club — Le Havre — made contact and used my data as one piece in a negotiation, driving down the price for a young striker whose scoring record at home had been markedly better than away before the pandemic. I do not know the final gap, and I do not need to. What I know is that from then on, every statistical table I built separated home and away columns, and every player assessment in my writing carried a warning line: do not trust pre-lockdown data until you understand the conditions in which its sample was collected.

In 2026, the empty-stadium report reached Canal+, and they sent me to Qatar for the World Cup at the age of 62. There I collided with what I consider the most important problem in modern football: the inverted full-back. Achraf Hakimi was praised everywhere — 142 sprint bursts and 2.3 chances created per match. Reports competed to call him the model of the modern full-back. I went into the data from a different direction: I measured the space behind him.

Result: the corridor behind Hakimi was vacant for 34% of match time. That is a large figure. Morocco kept clean sheets for much of the tournament, but not because that structure was safe — rather because their centre-backs covered it. I measured the top speed of Morocco's centre-back line and found many readings above 31 km/h. In other words, Morocco's system survived on a very specific compensating variable: the speed of its centre-backs.

The necessary and sufficient conditions for a system to work matter more than its glamorous surface.

Read only the 142 sprints and you conclude Hakimi is a perfect full-back. Read the 34% vacancy as well and you understand he is an attacking weapon placed on top of a defence assigned to cover him. Not every club has centre-backs running at 31 km/h. When Morocco met France, the opponent funnelled the ball repeatedly into Morocco's right flank, and that corridor — pre-marked in my spreadsheet — became the place where the match happened.

I say this not to praise myself. I say it to point out something analysis usually skips: when a tactical fashion wins, we tend to assume it won because it is good. It won because the surrounding conditions permitted it. Change the conditions and it collapses.

Here I must say plainly something few in the industry want to hear: correlation is not causation, and modern football is drowning in conclusions built from correlation without testing causation.

Take my own 0.84. It says accumulated xG and actual goals move together at club level over half a season. It does not say a shot with 0.3 xG will go in 30% of the time in any specific passage. It does not say the team with higher xG always deserved to win. And it absolutely does not say a striker with a large xG-goal gap is an undervalued talent — far more likely he is in a system generating few but high-quality chances, or facing superb goalkeeping.

I see transfer reports quoting xG as a seal of confirmation. I see reports concluding a defender is "bad" because his tackle count is low, without asking whether the system assigns him tackles at all. I see predictions based on a single season, when the sample size needed for a player-level metric to stabilise is usually three seasons or more.

In the transfer-market trade, I watch the same error repeat every window: valuation models overprice young potential and underweight dressing-room chemistry. A 21-year-old midfielder can have superior progression metrics and still fail at a new club, because no data column measures whom he eats with, whom he sits beside on the bus, and whom he threatens in the dressing room. This is not sentimentality. It is the fact that we are optimising against an incomplete objective function and calling it science.

Meanwhile, financial pressure distorts sporting decisions from another direction. When a club lists on a stock exchange, shareholders need a growth story told every quarter. Young players become assets to be recognised and resold. Academies become production lines. I do not oppose money in sport — I have worked in this trade for nearly half a century and I understand cash flow. But one thing is clear to me: when the financial statement is the primary yardstick, sporting decisions bend toward what serves the statement, not the scoreline.

Esports gives me one more observation supporting this argument. I follow esports tournaments the way I follow a youth league compressed in time. Professionalisation there happened ten times faster than in traditional football: within a few years players were turned into products of a digitised training pipeline, with every action logged and optimised. Individual flair — the very thing that makes the discipline attractive — was gradually sanded down. I see the same happening more slowly in football: young players trained to optimise metrics before being taught to understand space. A mouse click on an esports screen carries the shape of a pass — and both can be optimised to the point of losing surprise.

So what comes next for me? I am not looking for new metrics. I am looking for new contexts in which to test old ones. World Cup 2026, with its expanded format and denser schedule, will be another laboratory. I have already begun building a PPDA tracker by half, separating the final fifteen minutes, because I believe the new format will push the value of squad rotation to an unprecedented level. I am also separating scoring data from expanded tournaments from that of traditional ones, because more teams means more variance.

There are matches won on the pitch but lost on the spreadsheet, and I choose the spreadsheet — not because the spreadsheet is always right, but because it is the only place where we can argue with evidence instead of feeling.

Based on my experience tracking matches across four World Cup cycles and more than twenty Ligue 1 seasons, I draw one conclusion I want to leave with anyone reading this far: most wrong conclusions in modern football come not from bad data, but from asking the wrong question of good data. People ask "which team is better" when they should ask "which team generated more quality chances under these specific conditions". People ask "is this player good" when they should ask "good in which system, against which opponent, at which stage of his career".

A cancelled match is not a loss of points; it is a lost page of the diary. And a lost page means a sample that will never exist — which a data man like me considers a graver loss than three points.

I am 66. My hair has gone white, and most of my life is a constant: wake in Marseille, read the table, count, verify again. Players are variables. The market is a function. But the reader of the table must know what question he is asking. If you are preparing for a transfer window or a major tournament, try one small thing: before trusting any number, ask how large the sample was, where the confidence interval lies, and whether the context of collection resembles the context you are about to apply it to.

If the answer is no, pick up a pen and count it yourself. That is all I have done for nearly fifty years, and all I can suggest.