When Data Goes Silent: The 'Null' Trap in Modern Football Analytics

MINUTE 34 AT VIEUX-PORT

On 12 March 2026 I sat in front of two screens in an apartment overlooking the Vieux-Port in Marseille. The left screen showed Marseille hosting Rennes in Ligue 1. The right screen carried the live Opta data feed. At minute 34, the right screen froze. No red warning, no error message. It simply stopped updating. The score stayed 1-0, the stands kept singing, and on my desk a slice of a football match had just vanished from anything retrievable.

It took me 41 minutes to hand-record the rest: the starting point of every counter-attack, the number of times Marseille crossed the halfway line, the passes completed before losing the ball. When the feed returned at minute 75, I compared the two records. My handwritten sheet was missing 14 events the system had. The system was missing the 41 minutes I had. Both were correct, and both were incomplete.

That night I wrote a line in my notebook that more than twenty years as a transfer-market administrator had taught me but I had never said out loud: in modern football, missing data is more dangerous than wrong data. Wrong numbers still get doubted. Missing numbers get defaulted to zero.

CONTEXT: A NET THAT DOES NOT COVER EVENLY

A single Ligue 1 match today generates roughly 3.2 million positional data points and close to 2,000 ball events. With optical tracking systems such as Second Spectrum or Hawk-Eye, every player is sampled 25 times per second. The Premier League introduced optical tracking in the 2026-20 season. In France, Opta event data has been present in Ligue 1 for far longer, but positional data arrived later and is unevenly distributed across divisions.

That structure matters because it determines who gets seen. Every Ligue 1 club has an analysis department, usually three to seven people. In Ligue 2 the number is usually one, and that person also handles youth recruitment. Down in the National, France's third tier, most clubs have nobody at all. They still play, still get promoted, still sell players. They simply leave no trace.

So the football data net does not cover evenly. It has fine mesh at the top and coarse mesh at the bottom. Every player-valuation model I have ever built has to answer one question before it runs: when the net has holes, where exactly are the holes?

Statistics recognises three kinds of missing data, and I use them like three drawers in a desk.

The first drawer is MCAR, missing completely at random. A camera failing at minute 34. Noisy but harmless, because it does not correlate with anything happening on the pitch. You lose data, but you do not lose it in any particular direction.

The second drawer is MAR, missing systematically but explainable by another observed variable. Ligue 2 has no tracking data. You know why: no money. You can compensate by comparing within the division, and the error stays within a manageable band.

The third drawer is MNAR, missing because of the very thing you are trying to measure. This is the dangerous drawer, and most transfer-market mistakes are born inside it.

In the summer of 2026 I learned to trust something nobody had named yet: xG. But I trusted it in a very specific way. I hand-recorded 1,204 shots from 20 teams across the first half of the 2026-18 season and checked them against actual goals. The correlation coefficient came out at 0.84. That number only means something inside the range of data I had. Outside that range, it is an assumption wearing a jersey with a number on it.

THE MNAR DRAWER: WHEN THE GAP IS THE ANSWER

Take the example closest to my own trade. In France, clubs do not publish wage bills. The DNCG, the national financial control body, audits them but does not release the detail. The result is that the clubs in the deepest financial trouble are precisely the clubs whose numbers you cannot see. The gap here is not random. It correlates directly with the variable you want to measure.

A second example, at a more granular level. Injury data in professional football is systematically under-reported. Players and clubs have an incentive not to log a recurring hamstring problem as an injury. It gets logged as load management. Medical records therefore look suspiciously clean, and the buyer reads that cleanliness as a positive signal.

A third example sits inside the way a system defines an event. A tracking system registers a press when a player closes within a certain distance of the ball carrier. A poor presser, arriving late or from the wrong angle, generates fewer logged events than a player who does not press at all but happens to be standing in the right spot. The metric rewards the player who stands in the right place, and the gap in the poor presser's data looks like absence. That is the first mechanism by which the market misprices: the system only records what it was programmed to see.

EMPTY STANDS: THE DATA STAYED, THE CONTEXT LEFT

In 2026, when European football restarted after the pandemic, my editor assigned me to the Bundesliga. I sat in Marseille and analysed 81 matches played behind closed doors in the 2026-20 season. Home teams won only 26 per cent of them, against 43 per cent before the pandemic. I wrote a report titled Empty Stands Kill Home Advantage.

Empty stands are the finest laboratory a data obsessive could ask for. Same cameras, same event definitions, same pitches, same players. Only one variable disappeared: the sound of people. In science you rarely get to remove a variable that large while holding everything else constant.

But my report was right and incomplete at the same time. A study by Vincenzo Scoppa published in 2026 in the Journal of Economic Psychology showed that home advantage vanished mainly not because players lost motivation, but because referees lost social pressure. With no crowd, decisions favouring the home side dropped sharply. The mechanism sat with the referee, not the player.

Which means that when I wrote that empty stands kill home advantage, I described the phenomenon correctly and located the cause incorrectly. And one Ligue 2 club, Le Havre, used that report to negotiate down the fee for a young striker whose home numbers had shone during the crowdless period. The deal was sound. But it was sound for a different reason than the one I had supplied, and it took me two more weeks to notice.

A CANCELLED SEASON: A PERMANENT GAP

On 28 April 2026, the 2026-20 Ligue 1 season was cancelled after 28 matchdays. The final ten rounds were never played. Paris Saint-Germain were awarded the title on points per game. Administratively the season had an ending. In data terms it had a ten-round hole that no model can fill.

In Vietnam, the 2026 V.League 1 kicked off in early 2026, was suspended mid-year because of the pandemic, then cancelled with no champion crowned. For a league whose data infrastructure is already thin, a cancelled season is not just a missing trophy. It means an entire cohort of players enters the next transfer window with no full-season record of any kind. A cancelled match does not take away three points. It takes away a page of the diary, and that page has no backup.

The knock-on effect runs in ways few people calculate. A 24-year-old defender has 22 matches in a normal season. In a cancelled season he has 12. A 12-match sample carries roughly twice the confidence interval of a 22-match sample. But his transfer file does not display confidence intervals. It displays 12 matches. And the reader, in most cases, will read those 12 matches as a faithful miniature of 22. That is the most basic error any reader of numbers makes: confusing sample size with sample quality.

xG ACROSS BORDERS: THE MODEL DOES NOT KNOW WHAT IT IS MISSING

There is a technical problem I run into so often it has become professional reflex. The best public xG models are trained mostly on data from Europe's top five leagues. Apply them to a different league and the model still outputs a number to three decimal places, with no warning bar attached.

A shot from 14 metres through the middle of the box in Ligue 1 might be assigned an xG of around 0.11 in many public models. The same shot, in a league where defenders throw themselves in front of more shots, where pitch quality is lower, where goalkeepers come off their line differently, might be worth 0.07. The model does not know that. It was not built to know what it is missing.

The consequence is a scout sitting in Hanoi or Ho Chi Minh City reading the xG table of a V.League 1 striker and comparing it directly with the xG table of a Ligue 2 striker. The two numbers look like the same unit. They are not the same frame of reference. The biggest mistake in transfer analysis is not using a bad metric. It is using a good metric outside the range in which it was generated.

HAKIMI 2026: THE METRIC PRESENT HIDES THE METRIC ABSENT

At the 2026 World Cup in Qatar I was 62, sent there by Canal+ after my empty-stands report reached them. When pundits praised Achraf Hakimi, the numbers most often cited were 142 sprints and 2.3 chances created per match. Those are what the system recorded, and they were all accurate.

But I went digging through positional data and found something else. The corridor behind Hakimi was empty for 34 per cent of match time. That is not a metric published anywhere as a headline. You have to compute it, and to compute it you have to already suspect it is worth computing.

Morocco stayed safe for most of the tournament because their centre-back pairing ran at speeds above 31 km/h, enough to cover that space. I wrote a note warning that the fashionable attacking full-back only holds up as long as the back line retains its pace. Against France, the opponent attacked that right flank relentlessly, and Morocco lost 0-2.

What I took from it had nothing to do with the result. It was that 142 sprints is a loud number, while 34 per cent empty corridor is a quiet one. In any dataset, loud metrics crowd out quiet metrics, not because they matter more, but because they are easier to read.

CROATIA 2026: A SINGLE LETTER

At the 2026 World Cup I tracked all 64 matches and counted PPDA for every team. In the semi-final between Croatia and England, Croatia allowed England 8.2 passes per defensive action, while England allowed Croatia 12.5. I filed a preview predicting Croatia would win through pressing in extra time. They won 2-1, with goals from Ivan Perišić on 68 minutes and Mario Mandžukić on 109.

I did not shout in celebration. I reopened the spreadsheet to hunt for outliers. Croatia won a tournament of low PPDA? Then PPDA is only a letter, not the alphabet. It is a correct letter, but it is not the alphabet. Luka Modrić ran less in extra time and held the ball more, which means another mechanism was operating in parallel that PPDA could not capture. Had I published only the 8.2 and dropped the rest, I would have sold the reader half a truth in the format of a whole one.

THE CONTRARIAN ANGLE: A GAP IS NOT NEUTRAL

Here is the point most public analysis skips: missing data is not a neutral white space. It is a bias with a direction.

The transfer market makes this error in a very regular pattern. Young players have thin records. Thin records get filled with optimism. And optimism has no price ceiling. That is the structural reason why promising young transfers are consistently priced above their expected value, while dressing-room chemistry, which has no dataset at all, is priced at zero.

But I do not want to conclude on a single axis. Before any conclusion about data, I force myself to write at least three explanatory hypotheses. For the empty-stands report the three were: players lost motivation, referees lost pressure, and away teams adapted faster. I could only verify the second through outside research, and I have still not fully eliminated the third. Carrying two uneliminated hypotheses is the price of honesty.

There is one more layer my transfer-market work taught me. Financial reporting pressure tends to sit on top of sporting decisions. When a club needs to present a complete dataset to its owners or to investors, that club has an incentive to make its data look complete. A flawless dashboard does not prove a club is well run. It may only prove a club is well documented. Those are two different things, and the market routinely pays for the second as though it were the first.

I am 66, old enough to know a number never tells a story unless you ask it to. Players are variables, the market is a function, but most of my life has been a constant: the man who stays behind after the final whistle, checking what was recorded and what disappeared.

SIGNALS FOR THE NEXT CYCLE

A new role is taking shape inside European clubs' analytics departments, and it has no official name yet. I call it the null register: the person responsible for logging what was not logged. The first clubs to build that process will hold a pricing advantage for the next three to five years, because they will be the only ones who know where the net is torn.

In Southeast Asia, the highest-yield investment is not a better model. It is more complete coverage of the lower divisions. A sophisticated model running on incomplete data still produces incomplete output, only with more confidence.

There are matches won on the pitch and lost on the spreadsheet, and I choose the spreadsheet. But I choose it with one condition: I have to know what it is missing. Anyone reading a statistical table without asking where the empty cells are is reading a book with several pages torn out and mistaking it for the whole story. My job, at 66, is to count the pages again.