Empty Data Is Not Clean Data: Lessons from Anfield 2026 to Kazan 2026

On 27 June 2026 in Kazan, Germany held 74% of the ball, fired 26 shots and generated 1.8 expected goals. South Korea took four shots, produced 0.8 xG, and won 2-0 through stoppage-time goals from Kim Young-gwon and Son Heung-min. I sat in my Los Angeles office, reopened the pre-match file, and saw what I did not want to see: every data cell I needed for that scenario was blank, and I had filled them in with feeling.

That lesson did not start in Kazan. It started at Anfield on 27 August 2026.

Liverpool beat Arsenal 4-0 that day, with goals from Roberto Firmino, Sadio Mane, Mohamed Salah and Daniel Sturridge. The shot count was not that lopsided: Liverpool 18, Arsenal 9. Run through an xG model for the first time, the numbers read 3.6 against 0.3 — more than a tenfold gap in chance quality. As someone who audits data for a living, I did not believe it immediately. I logged everything, verified it across the next ten Premier League rounds, and the model called roughly 80% correctly. Before you trust a number, ask where it was born — I repeat that to myself every time I open a new dataset.

Faith in xG did not save me in Russia. The model broke in the group stage, and it broke in the most irritating way: it returned results that looked entirely reasonable. Germany dominated the ball, dominated the shots, posted a high xG. Every indicator said Germany. The match said South Korea. It took me weeks to name the problem properly: raw data cannot measure deadlock, cannot measure the psychology of a side pinned back for 90 minutes, and cannot measure an opponent deliberately conceding territory. A season is a scripture, each match is a verse — do not rush to chant half of it.

After Kazan I added two variables to every football wager: the opponent's PPDA and the actual intensity of the match. PPDA measures how many passes an opponent is allowed before being pressed — in other words, who genuinely controls the tempo. A team with 74% possession that lets its opponent build freely from the back is not in control. It is holding the ball in despair.

Then 2026 broke the model a different way.

When football returned behind closed doors, every home-advantage coefficient in my formula skewed badly. I logged 157 Bundesliga matches from May 2026 and found the home win rate had fallen from 43% to 36%. I split the sample by month and by team ranking to re-test it. The trend held. I added an audience variable to the formula and cut the home-advantage weight in every wager. The model was not wrong; the world changed while I was not looking.

In 2026 I was handed the full European Championship forecasting job. I backed Italy despite their lack of standout stars, on a single figure: 0.6 xG conceded per match in qualifying. Italy reached the final and beat England despite losing the xG battle in that last game — 1.1 against 1.9. Gianluigi Donnarumma saved two spot-kicks, which no model predicted. xG is not the truth, it is only a mirror — but a mirror does not know how to lie. A warped mirror still beats a compliment.

Those three stories taught me three things, and the most uncomfortable one is the one I have to write down.

A model is only as good as its inputs, and inputs always have holes. The question is whether the analyst is willing to write those blanks down. A table with enough rows and columns looks like an analysis even when every cell has been filled with guesswork. Small data is what big data always exposes. A group stage is three matches. A knockout round is one. A model calibrated on 380 league matches behaves very differently when it has seven games to talk with.

The most dangerous risks in sport also do not announce themselves. A muscle injury, a crack in the dressing room, a club that has not paid wages, a referee's new reading of the handball law — all invisible unless actively screened for. Their absence from a dataset is not evidence they do not exist. In esports the logic is harsher still: a patch that shifts a champion's power, a transfer completed three days before a tournament, a server build different from the practice build. An analyst who does not screen for those does not get a clean report. They get an unscreened one.

There is a temptation bigger than guessing a result: substituting the subject. When data is missing, the analyst is tempted to infer a team, a player, a patch version from surrounding context, then write about it with confidence. I have seen those reports. They fail by analysing the wrong match, the wrong tournament, the wrong moment — and still look thoroughly convincing. That is the costliest error in this trade, because it leaves no trace in the data.

xG has its own limits, and I state them every time I cite it. Italy won the Euro 2026 final while losing the xG count 1.1 to 1.9. No model explains luck, shootout nerve, or the instant a goalkeeper guesses right. An index only answers the question it was built to answer. It does not answer questions about refereeing standards, about psychology, or about who will miss a penalty in the 88th minute.

That is why I keep an odd habit: reading the footnotes before the scoreline. Footnotes say how the data was collected, by whom, under what assumptions. A scoreline only says what happened. The Liverpool shock did not make me afraid of data; it made me afraid of confidence.

For the next round I carry four tasks. For every claim, define the sample size and state the error margin openly. Screen the high-severity risk list before writing anything positive — injuries, bans, unpaid wages, rule changes. Check whether the model's home-advantage coefficient still fits current conditions. And when a data cell is blank, leave it blank. Before you fight, read last season again — and read the footnotes carefully.

A table with blanks looks unfinished. A table filled with guesswork looks far more finished. Between those two, I choose the one that looks unfinished. The question I carry into every round is no longer what the model predicts, but what the model cannot see.