World Cup 2026 and the First Failed xG Model: When a Prediction Model Reads the Wrong Question

On the night of June 27, 2026, in Kazan, I sat in front of a screen with a spreadsheet already open. My xG model — a tool I had spent six months building during my sophomore year of university — gave Germany 1.9 expected goals against South Korea. The actual score: 0-2. Germany did not lose because of bad luck. They lost because my model could not see what was happening beyond the touchline. I reopened all 64 matches of that tournament, re-ran the regression from scratch, and discovered that the flaw was not in the data — it was in the question I was asking the data. A wrong model does not mean wrong data — it means I have not yet read the right question.

Context: Six Months Building a False Confidence

In 2026, I was 22 years old, a second-year data science student in Lyon. The World Cup in Russia was the first time I applied my xG model to a major tournament. My method at the time was fairly simple by today's standards: collect shot data from public sources, assign weights by distance and angle, run logistic regression to estimate goal probability, then aggregate into team xG. I was confident in the model because it had correctly predicted roughly 70% of qualifying results in European leagues.

The first problem appeared in the Germany–Mexico match. My model gave Germany 2.1 xG; they lost 0-1. I noted it as "opponent luck." Germany–Sweden, the model gave 1.8, Germany won 2-1 in stoppage time. I told myself the model was working. By Germany–South Korea, the accumulated error was too large to ignore. Germany created 26 shots, held 74% possession, but lost 0-2 and were eliminated in the group stage — the first time in their World Cup history.

That was when I stopped and asked the right question: what is my model actually measuring?

Core Analysis: Two Variables the Model Ignored

I re-ran all 64 matches with two additions. First, the opponent's PPDA (Passes allowed Per Defensive Action) — a measure of pressing intensity. Second, the blocked-shot rate, which my original model treated as noise and discarded.

The result forced me to rewrite the algorithm in three days.

Opponent PPDA explained 34% of the variance in a team's shot-conversion performance. In other words, when South Korea pressed at PPDA 7.2 — their lowest of the tournament — Germany's shot quality dropped not because Germany shot poorly, but because the space to shoot well disappeared. My model assigned the same weight to an 18-metre shot in open space and an 18-metre shot with two defenders closing in. Mathematically, those two situations are entirely different.

Second, the blocked-shot rate at World Cup 2026 averaged 26.4%, but for low-block defensive teams like South Korea, that figure rose to 41%. My model removed those shots from the dataset, meaning I was measuring Germany's attacking quality after deleting precisely the situations South Korea's defence created to stop them. I was measuring the visible and ignoring the submerged.

When I replaced "number of shots" with "effective shots" — defined as shots not blocked and taken when the distance to the nearest defender exceeded 1.5 metres — the model re-predicted Germany–South Korea as Germany 0.8 xG and South Korea 1.1. The actual 0-2 score fell within the 90% confidence interval.

What I learned was not "add variables." What I learned was that every number in a model must answer the question: whose action is it measuring, under what conditions. xG measures chance quality. But chance quality is the product of an interaction — between the shooter and the blocker. If a model only looks at one side, it is measuring half the match.

Contrarian Angle: Correlation Is Not Causation, and xG Is Not Truth

After World Cup 2026, I saw many analyses using xG as a final verdict. A team with higher xG that lost was called "unlucky." A team with lower xG that won was called "lucky." That reading ignores a basic fact: xG is a probability model, not a measure of justice.

The deeper problem is that xG is built on historical data from major leagues, primarily European. When applied to Southeast Asian football — where set-piece shot density is significantly higher, where pitch quality varies widely between stadiums, where temperature and humidity directly affect movement — the European model is not merely imprecise, it can reach the opposite conclusion.

I verified this when I began working in Vietnam. In a sample of 120 V.League 1 matches I tracked directly, goals from set pieces accounted for 38% of all goals — compared to 22% in the Premier League over the same period. An xG model calibrated on the Premier League will significantly undervalue a corner kick in the V.League context. And when a model undervalues a situation, it is not just wrong about the number — it is wrong about the tactics. It leads readers to believe a team is playing inefficiently, when in fact they are playing exactly to their conditions.

Numbers never lie, but they are very good at telling half-truths. The half-truth xG tells is: this team creates many chances. The half-truth it omits is: under what conditions were those chances created, against which defence, at what minute of the match, and in what psychological state.

There is one detail from Germany–South Korea I have never forgotten. In the 92nd minute, when Germany already knew they were eliminated, Manuel Neuer — the goalkeeper — pushed up to the opponent's penalty area. My model recorded that as a German shot. Technically, correct. Tactically, it was a sign of collapse, not attacking strength. A model that cannot read context will always confuse the two.

Takeaway: Signals for the Next Cycle

Three days after Germany–South Korea, I rewrote the entire algorithm. Not to make it more complex, but to make it more honest. I added opponent PPDA, added the blocked-shot rate, and most importantly — I added a line at the top of every report: "This model measures chance quality under average conditions. Any conclusion outside average conditions requires independent verification."

That line has followed me for seven years. It is why I never publish a prediction without an uncertainty interval. It is why when working with V.League data, I always start by building a league-internal model before comparing to any European benchmark.

And it is why I believe a repeatable process matters more than inspiration, because a process can be checked while inspiration cannot.

The question I pose for the next cycle is not "which model is more accurate." The question is: when a model predicts wrongly, are we fixing the model or fixing the question? Seven years after World Cup 2026, I still do not have a definitive answer. But I know one thing for certain: every time my model fails, it is an opportunity to understand the match more clearly — provided I am honest enough not to blame the data.