The Data Gap: When a Model's Silence Is Read as Safety

Data Gaps: When a Model's Silence Is Read as Safety

On the night of December 30, 2026, at Anfield, Wout Faes put the ball into his own net twice in a single match. Leicester City lost 2-1 to Liverpool. In the analysis room, Leicester's expected-goals model did not ring a single alarm. The volume and quality of chances Liverpool created sat inside the normal range for an away match in the Premier League that season. The model stayed silent. And in a great many meeting rooms, that silence was translated into a human sentence: there is no serious problem at the back.

This is the error I encounter most often after more than a decade in sports data analysis. An empty dataset, or a dataset packed with rows but missing the one variable that matters, gets read as a safe conclusion. A model's silence is interpreted as a club's calm. And almost nobody challenges that silence, because the report is presented so beautifully: a cover page, a table of contents, comparison tables, a bolded conclusion at the end.

A report that finds no faults does not mean no faults exist. It only means the report was never given enough data to look.

I once received a forty-page document of exactly that kind. Every chapter had a serious heading. Every table had columns and rows. But most of the content cells carried the same phrase: insufficient information. Nobody in the meeting read that phrase as a warning. They read it as paperwork. The forty pages had done their job: they looked professional enough that nobody asked another question.

Context: aggregate metrics versus the level of causation

From the 2026-23 season onward, I tracked Leicester City through the entire stretch in which they sat in the relegation zone of the English Premier League. After fourteen rounds, the data I collected showed a gap of 7.8 goals between expected goals against (xGA) and actual goals conceded. For most models, a divergence of that size is assigned to luck and is expected to wash out over time. But when I split the data down to the individual player level, the cause emerged in a different order: most of the divergence came from individual errors in defence, not from difficult finishes.

This is the point readers of a statistics table routinely skip past. An aggregate metric at club level can be right on average while being useless at the level of cause. Expected goals against measures the quality of the chances an opponent creates. It does not measure whether your centre-back reads the pass, turns at the right moment, or holds the correct distance from his partner. Two different questions, two different datasets, and only one of those two datasets exists inside the internal reporting of most professional clubs.

Brendan Rodgers was sacked on April 2, 2026. Dean Smith took over and moved Leicester to a back three — precisely the direction the data had indicated months earlier — but the club was still relegated on the final day of the season. The delay did not lie in a shortage of data. It lay in the fact that the correct dataset was never placed beside the incorrect one for comparison, and nobody was made accountable for putting the two side by side.

The chain of evidence: one mechanism, three shapes

The mistake of that year taught me that data never lies; only the reading is wrong. The error was not in the 7.8 figure. It was in the assumption that every divergence shares the same probabilistic nature. In sport, divergence has two entirely different origins: pure variance, and systematic error. The first washes out. The second does not, and it will keep recurring until the system is fixed. The analyst's job is not to compute the divergence. The analyst's job is to classify it.

The second shape of the same mechanism shows up in scouting. In January 2026, I scanned data from 49 European domestic leagues to find centre-back prospects for Korean clubs. I stopped at Isak Hien, a 24-year-old Swedish centre-back of Ethiopian descent, then playing for Hellas Verona in Serie A. Hien recorded 2.9 successful tackles per match. But the metric that caught my attention more than any other was the frequency of successful progressive passing: he hit that marker in more than two-thirds of his matches. For a centre-back at a club fighting in the lower half of the table, that rate signals an ability to launch attacks, not merely to clear the ball.

I wrote a deep profile of Hien, placing his record alongside Virgil van Dijk's at the same age. The piece drew attention in Korea. But when I proposed that national-team scouts look at him, the answer came back: no direct source. Four months later, Atalanta signed Hien. On May 22, 2026, in Dublin, Atalanta beat Bayer Leverkusen 3-0 to win the Europa League, ending Leverkusen's 51-match unbeaten run across all competitions. Hien was a link in that back line.

Between the transfer numbers lies a story nobody writes into the report.

Hien's story is not a story about weak data. It is a story about a process that equated "no direct source" with "insufficient grounds." Those are two entirely different statements. A missing direct source is a problem of collection channels. Insufficient grounds is a problem of evidence. In Hien's file, the evidence was not missing. What was missing was a person who had sat in the stands at Verona that winter.

The third shape appears at the market layer. I once bet on a wrong dataset and received a correct lesson. In betting analysis, people like to say the market is always right. I disagree with that phrasing, because it assigns the market a quality the market does not possess. The betting market is not wrong; it merely reflects a truth you have not yet managed to see. It is not a prophet. It is an information-aggregation mechanism: the price reflects what people know and what people fear. When the odds move before the news breaks, the movement itself is the information.

But there is an opposite case that newcomers routinely misread: when the odds do not move. Static odds can mean the market has reached consensus. They can equally mean the market has no liquidity, no takers, nobody actually watching that fixture. A silent market has never been an efficient market. The stillness of the price, in that case, is not a safety signal. It is the mark of an observation gap.

This mechanism repeats at every level of the industry, esports included. A patch that lists no changes does not mean the meta has frozen. A team that announces no roster change does not mean the roster is stable. Esports does not need luck; it needs people who read the meta faster than the server does. And the best meta readers are those who can tell two kinds of silence apart: the silence of a system running steadily, and the silence of a system that has never been measured.

The cancelled Seoul derby of 2026 was a test for every prediction algorithm. When the K-League was postponed indefinitely because of COVID-19, the Seoul World Cup Stadium sat empty for the first week. I worked remotely, analysing FC Seoul's first ten matches that season to predict which clubs would survive. The team's average running distance was 98.7 km per match, third lowest in the league. The rate of tactical fouls in their own half climbed. None of those signals appeared in any of the prediction models in common use, because those models were built on the assumption of spectators, home advantage, and a normal competitive rhythm. When the underlying assumption collapsed, the models were not wrong. They became meaningless.

In the V.League the situation is even starker. The league's data infrastructure remains thin compared with top European competitions. A club without a fitness-tracking system does not mean its players are not overloaded. A team without an injury-analysis department does not mean there are no injuries. Personally, when I follow the V.League from Seoul via streaming feeds, I usually have to rebuild the data by hand — logging duels, progressive passes, and team shape by half. That work is slow and error-prone. But it gives me something an empty data table cannot: a basis for saying I do not yet know, rather than an excuse for saying everything is fine.

Missing data and safety are two different states, but they produce the same output when they pass through a weak filter: blank space.

The counter-intuitive angle: format grants credit, not evidence

The worry is not bad data. Bad data is always easy to spot: small samples, unclear sources, large error bars, and people instinctively know to distrust it. The worry is an empty report formatted professionally enough to pass every review gate.

I call it the formal-authority effect. When a judgement sits inside a document with a table of contents, comparison tables and a risk-matrix section, readers absorb it with a higher level of confidence than its actual evidence warrants. Format does not lie. It merely extends credit. And that credit can perfectly well be extended against an empty loan, as long as the lender never checks the collateral.

In football, this confusion appears as a dataset with correlation but no causation. A team wins many matches using a back three, so the coaching staff concludes the back three is the cause. But if the fixture list was lighter in the same stretch, the opponents weaker, and the goalkeeper in form, then the back three is only one of four variables occurring together. Correlation is not causation — repeated so often it has become a cliché, yet it remains the most common reason a tactically correct decision in one period turns out wrong in the next.

Data analysis, at its deepest layer, is not a profession of giving answers. It is a profession of asking the right question of a dataset, then determining whether that dataset is even qualified to answer. I do not believe in intuition; I believe in numbers that speak once they have been asked correctly. But an empty dataset, however correctly questioned, cannot speak. It can only stay silent. And the analyst's task is to say out loud that this is the silence of the data, not the silence of the problem.

There is one small detail from the Leicester story I keep as a professional note. When I published the analysis, I stated a confidence level for each judgement and made the sampling error of the fourteen-round window explicit. Some readers read the conclusion and skipped the error bars. Others read only the error bars and concluded the analysis proved nothing. Both readings are wrong in the same direction: they took a fragment of the document instead of the whole chain of reasoning. That is also why I began splitting my writing into two layers — a foundation layer for newcomers, and a deep layer for people who do this for a living.

If a model is not given enough variables to describe a phenomenon, then the fact that it raises no alarm carries no information about that phenomenon. It carries information only about the model itself. This is the point that public discussions of sports data almost always skip, because it is less exciting than a bold prediction.

What to watch in the next round

The question for the period ahead is not which team will win the title, nor which player will shine. The question is what kind of validation gate each club operates before a report lands on the coaching staff's desk. If that gate accepts a data table with blank cells without returning an error, then every conclusion drawn from it is built on air, no matter how handsome the cover page.

I am watching to see whether, in the coming months, any club publishes a protocol that forces its analytics department to return a failure signal rather than a clean report. That would be a signal worth tracking far more than any transfer deal in the same window. Every season is a ritual, and the analyst is merely the person who records the omens. Our job is to record the omens that do not appear as well — and to record them at exactly their proper weight: a blank space, nothing more.

---

This analysis draws on public information and personal match-tracking notes. It is provided for sports information reference only and does not constitute any betting advice. Sporting outcomes are highly uncertain; readers should treat analytical conclusions rationally.