When Data Falls Silent: The Craft of Football Analysis and the Trap of Certainty

Tuesday night, 11:47 PM Rio de Janeiro time, my work inbox lit up with a 34-page PDF. The title was clear: "Deep Analysis Stage 2." I brewed a black coffee, sat down at the old wooden desk where I have written every article for more than ten years, and started turning the pages. Page one. Page two. Page three. No team name. No player name. No score. Not a single metric — no xG, no pass count, no possession share, no pressing index. Only a line repeated roughly nine times, once for each heading: "insufficient information to assess." I sat there for a long time, staring at pages as blank and milky as early morning mist over Guanabara Bay. In this profession there is a moment that sends a chill down your spine, and for me it was the moment I realized that everything I was supposed to analyze was empty. The report did not lie. It was simply honest to the point of cruelty. And that honesty — not any tactical error — was what deserved to be dissected in a long piece. I called the old colleague in São Paulo who had sent me the document. He laughed awkwardly: "I thought you'd figure it out. Stage one has nothing. Stage two is built on the void." We were silent for a few seconds. Then I said what I now think was the most important sentence of that week: "Then you didn't send me an analysis. You sent me a broken pipeline." The call lasted forty minutes, and it taught me more than any tactical meeting that month. This story matters not because it is rare. It matters because it is so common that we have grown used to papering over it with rhetoric. Football analysis over the past two decades has undergone a data revolution: every match in a major European league now generates millions of positional data points, thousands of labeled events, hundreds of composite metrics from Expected Goals to Passes Per Defensive Action. Alongside that boom has come a new occupational disease: the belief that more data means complete data. In Rio de Janeiro, where I live and work, people often assume Brazil is the paradise of inspired football while Europe is the paradise of data football. Seen from the bench, both are half wrong. Brazil has analysis rooms as rigorous as anything in the Premier League — Fluminense, Palmeiras, Internacional all invest in camera systems and full analyst teams. Europe has clubs that still buy and sell mainly on the intuition of a scout. The problem in both football cultures is not the volume of data; it is the input-verification stage. And that is where the empty 34-page report touched my craft. At Fluminense in 2026, I watched the coaching staff present a high-pressing model built on GPS data from twelve matches. Twelve matches. The number sounded ample when printed on a handsome slide, with colorful heat maps and direction arrows. But when I suggested testing the stability of the data across three seasons, nobody wanted to bother because it would "waste time." I did it alone. Three weeks later, I discovered that the team's defensive system only truly functioned when opponents had a lateral passing rate above 62 percent. Below that threshold, the model collapsed. We were not playing with a model. We were playing with a precondition nobody had written down. That is why I always tell younger colleagues: "Numbers tell the first part of the story; the rest is flesh and sweat." That sentence is often read as literary encouragement. It is not. It is a professional rule. The first part that numbers tell can be very long, but it can never tell the whole thing. The rest — the pressure on a full-back when the stands roar behind him, the decision of a holding midfielder when his partner has lost position, the breath of a center-back in the 88th minute — lies outside every chart. And when the first part is empty, the rest has no anchor to attach to. I have spent more than ten years working with models, and the biggest lesson I have learned did not come from a complex algorithm. It came from a simple principle: before asking "what does the model say," ask "what is the model fed with." If the answer is "insufficient information," then every conclusion that follows is an illusion. This sounds obvious. Yet in practice, analysis pipelines — at clubs, at broadcasters, at newsrooms — are usually designed to always produce an output, not to be honest about their own gaps. To understand why an empty pipeline is more dangerous than a wrong model, look at the structure of the craft. A decent piece of football analysis passes through five layers: raw data, context, analysis, cross-verification, and judgment. When the first layer is empty, the later layers do not disappear. They amplify the emptiness. I call this the amplification effect of emptiness — a concept rarely discussed because it has nothing glamorous to present on a conference stage. Take the 2026 World Cup, a tournament that taught me the humility lesson I still repeat. In the round of sixteen in Moscow, I predicted Japan would collapse under Belgium's physical pressure. My basis was solid: average height, muscle mass, and duels won all favored Belgium. But Japan led 2-0 after more than fifty minutes with lightning-fast transitions, and I had to sit still while a Brazilian colleague turned and asked where my prediction was. After the match, I rewatched the tape five times. Five times. And I found something my model did not measure: the space between the lines. What I overlooked was that Belgium was not losing because they were physically weaker. Belgium lost for about thirty minutes because the gap between their midfield and defense was stretched wide. Eden Hazard and Kevin De Bruyne still carried the ball well, Romelu Lukaku still held the center-forward position, but the space behind the midfield became a corridor that Takashi Inui, Genki Haraguchi, and Ritsu Doan exploited ceaselessly. When Japan won the ball, they did not need many passes. They only needed one into that gap. Belgium's Expected Goals that night were still higher, their possession still dominant, their pass count still three times Japan's. Every traditional metric said Belgium played better. But the match nearly took a different path, because of a metric that did not exist in my charts. The space between the lines is measurable by video, not by GPS. And because I was used to trusting charts, I did not look at the video carefully enough. The lesson is not "don't trust data." That is a lazy and equally dangerous conclusion. The lesson is: every model has a blind spot its builder cannot see, because the builder is the one who decides what gets measured and what does not. "The 2026 World Cup taught me: every model needs a humble seat." After that, I spent three months rebuilding my analytical framework — not to remove data, but to add a new column: "factors not yet measured." But those three months taught me something else, and this is where the empty 34-page report struck my craft precisely. As I refined the framework, I realized every model borrows from another model. My prized space-between-the-lines metric is computed from positional data, which is generated by a camera system, which is calibrated by a human being. This chain of borrowing is so long I cannot trace it fully. And if a single link breaks — a blocked camera angle, a mis-calibrated algorithm, a tired labeler who mistypes — the entire chain downstream warps. This is exactly what I saw in that 34-page PDF. It was the first link. When the first link breaks, the later links do not know. They keep running. They keep producing reports. They keep presenting to the coaching staff. And the coaching staff, having trusted the pipeline, may make a tactical decision based on a number that does not exist. In professional football, such a decision does not show up as an own goal. It shows up as a midfielder asked to push five meters too high, a full-back left alone for thirty seconds, a striker receiving the ball in a position he should never occupy. No scoreboard records those mistakes. I once sat in a room like that. In 2026, a Fluminense coaching meeting ran three hours. On the board was a formation bristling with pressing arrows. I raised my hand and asked something I later recounted with embarrassment: which reference frame is our GPS data calibrated to? Silence. A young assistant said: the software default. I pressed: and what is the software's default reference frame here? Nobody knew. Three weeks later, when I checked, the software's default reference frame differed from the stadium's by a small error. That small error did not collapse the model, but it skewed the boundary metrics. And the boundary metrics are precisely what decides how many meters high a wide midfielder should push. We changed the model. Not because the model was wrong, but because it was fed by a mis-calibrated reference frame. That season's result: Fluminense finished sixth, four places better than the previous season, based on analysis of forty-seven matches instead of twelve. We kept the 4-2-3-1 and merely intensified pressing on the right flank. There was no miracle here. Only a cross-verification process carried through to the end. What I want to stress — and this is the most easily misunderstood part — is that cross-verification is not negative skepticism. It is a positive act. When I ask what the data source is, I am not sabotaging the meeting. I am trying to protect the coaching staff's decision from its own confidence. In football, overconfidence does not show up as throwing the ball into your own net; it shows up as making a tactical decision on a false basis. "The model is not wrong — it just has not learned how to speak." I wrote that after the 2026 World Cup, and I still use it. A good model is not one that is always right. It is one that knows how to say "I don't know" when there is a gap. In 2026, when the pandemic forced the entire Brasileirão to play behind closed doors, I was assigned to analyze thirty spectator-less matches for a sports magazine. It was the largest natural laboratory our craft had ever had: for the first time, we could isolate crowd pressure from every other variable. The home team win rate fell from 48 percent to 39 percent. High-pressing teams lost 12 percent of their effectiveness on average — not because they ran slower, but because the noise that pushed opponents into errors was gone. "The spectator-less match is the flattest mirror football has ever held up to itself." We looked into that mirror and saw that something we had called home advantage for years was really crowd advantage, and it does not live on the scoreboard. "Home advantage is not on the scoreboard; it is in the players' eardrums." That thirty-match report ran forty pages. The editors objected at first that it was too long, and I understood why — readers do not read forty pages. But I persisted, and in the end it was split into three installments. This story is not about me winning an editorial argument. It is about a principle: complex subjects need time to develop, and developing time cannot be compressed into a single headline. When someone asks why I do not write shorter, I usually answer with a question in return: would you like me to cut the cross-verification section? Nobody has ever said yes. Since then, in every analysis, I always check the environmental context before offering any tactical judgment: crowd or no crowd, weather, pitch, kickoff time, travel schedule. This is not formal caution. It is a direct consequence of those thirty matches. A pressing team loses 12 percent effectiveness without a crowd. So how much does a pressing team lose in January rain in Rio, after a flight from Manaus, with a center-back dealing with personal problems? Nobody can measure that number. But the analyst has a duty to acknowledge that the number exists, even when it does not appear on any chart. I still remember a debate on a Brazilian television set. A young colleague asserted that Team X would certainly be relegated based on a points model. I asked him three questions: how many seasons was the model trained on, how many matches, and was the context the same? He blushed. Not because he was wrong, but because he had never been asked those questions. Our craft lacks people who ask, not people who speak. "Tradition and data do not oppose each other; we use the latter to keep the former." I wrote that for a São Paulo magazine, and it captures my view: data does not replace professional intuition, it protects intuition from delusion. Here I must say something that may annoy many colleagues, and I say it with genuine humility. We are trending toward sanctifying the completeness of data as if it were a prerequisite for analysis. This view is no less naive than sanctifying the model. The truth is that football data has never been complete and never will be. Every match is a bundle of countless variables, most of them unmeasurable. Waiting for complete data before making a judgment means never making a judgment. This is the execution blind spot I want you to see. When I presented the space-between-the-lines idea, someone said: then your model is also poor. Correct. Every model is poor in some dimension. What I defend is not the perfection of the model, but honesty about its limits. Analysis is not the craft of building perfect models. It is the craft of living with imperfection and still deciding. There is a paradox I have never seen anyone fully resolve. The more data we have, the more easily we confuse the precision of a number with the precision of a conclusion. An Expected Goals figure accurate to two decimal places can still lead to a wholly wrong conclusion if context is ignored. Numerical precision is not tactical precision. This is the trap that newcomers — and even people twenty years into the trade — can fall into. I have committed both types of error in my career: saying too much from too little data, and saying too little from too much data. The best analyst is not the one who avoids both, but the one who knows where he stands between those two shores in each piece. In today's transfer market, this trap is more dangerous still. Hundred-million-euro deals for players who have not played fifty top-flight matches are usually justified by potential models. I do not oppose using models. I oppose using models as a shield to hide a lack of verification. A nineteen-year-old who scores fifteen goals in the Portuguese second division might be worth five million euros, fifteen million, or fifty million, depending on how you read the data. But how you read the data depends on whether you check his context: against which defenses he scored, at what point in matches, with what quality of teammates, under what crowd pressure. Ignore those questions and the number is just a pretty number. I will not end with advice. I leave a question I am still carrying, and probably will carry for years. If football data has never been complete and never will be, what standard should make us trust a conclusion? My answer, at this point in my career, is: trust the process, not the number. A process honest about its gaps is more trustworthy than a complete but fabricated chart. When an analysis says "insufficient information," that is not failure. That is maturity. Numbers tell the first part of the story; the rest is flesh and sweat — and the gaps we must learn to live with.