Three Knockout Matches and the Limits of Early Conclusions in Football Data

Minute 92 in Kazan, 27 June 2026. The ball left Son Heung-min's foot from roughly forty metres, cleared Manuel Neuer, and rolled into an empty net. In Busan I was sitting in front of a laptop with a half-finished Python window. Germany, the defending world champions, had just been eliminated in the group stage. I did not open the scoreboard. I opened the raw data file. Germany's twenty-three shots in that match. I entered every one into my own xG model: coordinates, distance, angle, shot type, build-up, shooter. The model returned 1.32 expected goals and zero goals scored. That night in Russia, I saw a number that could hurt for the first time. The number that kept me awake was not 1.32. It was 18 out of 23 — 78% of Germany's shots came from outside the penalty area. The commentators that day said Germany were stuck. The data said something else: Germany were not stuck, Germany picked the wrong zone. A team that pushes the ball wide and fires from distance is not necessarily desperate; it is executing a plan with no exit. I retell that old story because it is the first brick of a principle I have held across eleven years in this industry: before arguing about wins and losses, I ask the numbers first. Not because numbers are always right. Because numbers can always answer a question the naked eye cannot: where did this data come from, and how many matches are in the sample? In 2026 I was nineteen, a second-year student in Busan. The xG model I wrote in Python was nothing sophisticated: a simple logistic function, a few variables for distance and shot angle, trained on European league data I collected by hand from public statistics pages. Its error margin was certainly larger than that of commercial models. But that crude model taught me that most arguments online are not arguments about football. They are arguments about sample size. By 2026, that principle nearly made me throw the whole model away. K League 1 returned to play in front of empty stands — the first top-tier league in the world to do so during the pandemic. I sat in Busan, watched the matches on streams, and annotated frame by frame. The 2026 model began to drift in one consistent direction: home advantage disappeared. I collected data on 152 matches. The home win rate fell from 46.2% in the 2026 season to 31.6% in 2026. I finished a forty-page report. The conclusion sat in a single line: every 10,000 spectators is worth roughly +0.08 expected goals for the home side. The 0.08 coefficient does not measure the silence; it measures what we lost. Nobody commissioned that report. I wrote it because I knew that if the foundation was not fixed, every later analysis would be wrong. Two years later, that report took me to Morocco. December 2026. I was twenty-three, a new hire at a sports desk. Morocco became the first African team to reach a World Cup semi-final. My assignment was three knockout matches. I compiled the full dataset, benchmarked it against the tournament average, and wrote out three numbers. Morocco conceded 71.6% of possession. They conceded only one goal across three matches. Their opponents' combined xG totalled 4.02. The number that silenced the newsroom was PPDA: 25.1, nearly double the tournament average of 13.2. PPDA is the number of passes a team allows the opponent before making a defensive action of its own. A figure of 25.1 means Morocco let opponents string together roughly twenty-five passes before making a tackle or interception. Read the usual way, that is laziness in pressing. Read correctly, it is a choice: press only where you want to. PPDA 25.1 — sitting deep is not a concession, it is stretching the pitch. I wrote the piece and replaced "pinned back" with "deliberately sitting deep". Korean media pushed back hard at the time, mostly in objection, partly over wording, partly out of the habit of reading football by feel. But in the final section I did something many outlets skip: I stated the sample size. Three matches. Three matches are not enough to call Morocco the best defensive team in the world, and I refused to sign that conclusion. Three matches are only enough to demonstrate a mechanism: a team can concede territory and still control the terms of the exchange. That is the difference between a conclusion and a mechanism. Conclusions sell articles. Mechanisms outlive tournaments. In 2026 I was twenty-five. The Morocco piece connected me with a sports data company in Lisbon. From that source I found a Korean midfielder at a mid-table club who had played only 564 minutes the previous season, while his contract listed 1,200 — a gap of 41%. I sent his agent a six-page metrics report: minutes, progressive passes, duels won in the opposition half, and a comparison group of players in the same position. On 8 June 2026, I was the first to report the loan move with a €2.8 million purchase option. Transfer fees do not measure talent, they measure the buyer's desire. The €2.8 million figure does not say how good this player is. It says a club is willing to pay €2.8 million for a man whose contract says 1,200 minutes while the data says 564. There are times when I have nothing. An empty file, an unverified source, a sample too small to carry weight. In this profession, the easiest conclusion to write is the one drawn from an empty sample, and it is also the one most often wrong. I have typed "insufficient information to conclude" dozens of times in internal reports. It is the hardest sentence to write, because it gives you no headline. There is one temptation I fight every time I sit down to write: turning a correlation into a principle. The 0.08 coefficient is the clearest case. It was calculated under very specific conditions: empty stands, compressed schedules, differing health protocols across countries, weather, and a single season of 152 matches. I published all those limitations inside the report. Yet once the report was cited on a few forums, it became a different sentence entirely: fans do not matter. The 0.08 coefficient measures a correlation under specific conditions, not a law of football. The same distortion happened with Morocco. After the semi-final came a wave of writing about "the Morocco system", as if three knockout matches were a doctrine. Eighteen months later, when results dipped, those same articles asked where the system had gone. The system had not gone anywhere. The sample had changed. Three knockout matches and ten qualifying matches are two different worlds in terms of underlying conditions, and any comparison that spans them must label the bridge. The third blind spot sits with the data people, myself included. Analysts are moving into the dressing room, but their conclusions often detach from the actual rhythm of a match. Models grow more sophisticated yet still fail to capture a midfielder running the exact line the coaching staff asked for, dragging a defender out of position, opening the gap that the decisive pass only reaches two beats later. The rating system hands him a 7.2 and calls it a solid shift. The model is not wrong. The model simply was not designed to see that. The next tournament cycle is already in front of us. The first signal I will track is not goals. It is the possession-conceded rate and the PPDA split across the first two group matches, alongside minutes distribution across the squad. A team that concedes 60% possession in match one and 38% in match two is telling you about its capacity to adapt. A team that concedes 60% in both is telling you how it intends to win. The distance between those two sentences is my entire job. And the sample will still be small. Three matches, four, five — not enough to be certain. That is fine. If data could give you certainty, nobody would watch this sport. But if you intend to draw a conclusion from an empty sample, try answering this first: are you analysing the match, or are you analysing your own impatience?