When the Data Table Comes Back Empty: Data Integrity and the Limits of Sports Models

Kazan, the night of June 27, 2026. The clock at Kazan Arena ticked into the third minute of stoppage time, and across the stands most of the crowd in white shirts was already on its feet. Joachim Löw stood with his arms folded at the technical area. Germany held 74 percent of the ball, had taken 26 shots, and the model I was running on a workstation in Los Angeles — ten time zones behind Russia — returned an expected-goals figure of 1.8. South Korea: four shots, 0.8 xG. Then Kim Young-gwon put the ball in the net in the 90th-plus-third minute. Three minutes later Son Heung-min ran alone toward an empty goal and made it two. Final score: 2-0, to South Korea. I stayed up until nearly dawn. I was not rewatching the goals to find defensive errors; I had already done that four times. I stayed up because of something else. Nowhere in my spreadsheet was there a cell that read "Germany is stuck." Nowhere was there a column measuring the feeling of being pinned against a wall for the final thirty minutes. The model returned a number that was right about chances and wrong about the match. That was the first time in my career a model told me it had nothing to say. CONTEXT: A TRADE BUILT ON TRACING NUMBERS I work as a betting and sports data analyst in Los Angeles, Vietnamese by birth, and my daily work circles one deceptively simple question: where did this number come from. Before you trust a number, ask where it was born. Who collected it, how, what counts as a shot, and what was left behind in the labelling process. In 2026 I was a mid-level analyst at a sports data company. On August 27 that year, Liverpool hosted Arsenal at Anfield in the Premier League. I watched as an apprentice. Liverpool won 4-0, with goals from Roberto Firmino, Sadio Mané, Mohamed Salah and Daniel Sturridge. What stayed with me was not the score. The shot counts were close: Liverpool 18, Arsenal 9. Read only that column and you have a roughly balanced match decided by a few moments. The first time I ran expected goals on that match, it came back 3.6 for Liverpool and 0.3 for Arsenal. The gap was so wide I did not believe it. Being an empiricist, I did what an auditor would do: I logged everything, then tested it across the next ten matchdays. The xG model called the direction correctly roughly 80 percent of the time. I had to change how I looked at matches. The Liverpool shock of that year did not make me afraid of data; it made me afraid of confidence. Eighteen months later, that confidence was what failed me at Kazan. And between those two markers I learned something no classroom taught me: the null report. In a professional pipeline, when the input data cannot answer the question, the valid output is a document that states plainly "insufficient information," with a list of the minimum data required. It sounds like paperwork. In practice it is the last line of defence for data integrity. CORE: NUMBERS DO NOT SPEAK; PEOPLE MAKE THEM SPEAK Start with the footnote. I read the footnote column while everyone else reads the scoreboard. An expected-goals figure is not a physical measurement like temperature. It is a probability model trained on historical data, with a specific definition of what a quality chance is. Each major provider — Opta, StatsBomb, Wyscout — defines it slightly differently. A shot from the edge of the box with a clear sight of goal may be scored 0.08 by one provider and 0.12 by another. Nobody is wrong. But mix two sources into one chart without a note and you have created a new number that does not exist in reality. This sounds like dry technical detail until it touches a decision worth millions of dollars. At club level, recruitment departments use data to price players. At league level, broadcasters use data to build narratives. At fan level, a headline based on a number from the wrong source can turn an average player into a star, or the reverse, inside a week. In 2026, when I tested the xG model across ten matchdays, I did not only learn that it was useful. I learned that it was useful under a specific condition: when the sample is large enough and the comparison group is placed in proper context. A single match proves nothing. A season starts to have a voice. Three seasons start to carry weight. Small data is what big data always exposes. That is why I am so careful with small samples. A striker scoring four goals in five games is not a world-class finisher; he is a striker on a lucky run inside a small sample. A goalkeeper keeping four consecutive clean sheets is not a wall; he is a goalkeeper playing behind a well-organised defence across four games. The difference between those two readings decides the entire value of an analysis. The same thing happens in esports, where I spend most of my time covering the US market. A team that wins a regional event does not automatically become a world-title contender. A player with a high rating in a summer tournament does not automatically keep that form after a major patch rewrites the tactical system. THE MODEL WAS NOT WRONG; THE WORLD CHANGED WHILE I WAS NOT LOOKING In May 2026 European football returned after the shutdown, and stadiums welcomed crowds into empty stands. I had a model built on a simple but powerful variable: home advantage. That variable had been right for years. European domestic leagues typically show a home win rate around 43 percent, a figure stable enough that many models treat it as a constant. I tallied 157 Bundesliga matches from May 16, 2026, and found the home win rate had fallen to around 36 percent. At first I did not believe it. I split the data by month, by team standing, by matches with and without crowds. The trend held. Only after confirming it did I add a "crowd" variable and reduce the home-advantage weight in every model. The model was not wrong; the world changed while I was not looking. That lesson travels beyond football. In esports, a balance patch can destroy a model built on the previous tactical order overnight. A character rated weak can become the centre of an entire strategy because of a few small stat changes. A map once played one way can demand a completely different approach. And if an analyst keeps using last season's dataset without stating the version, every conclusion becomes meaningless. My trade has a particularly hard-to-detect error class: context expiry. It produces no error message. It produces a number that looks entirely reasonable. That is also why I always check one question before analysing any tournament: does the competition build match the practice build. In football, the variant is whether refereeing rules changed between rounds, and whether the season was interrupted in a way that distorted the comparison data. XG IS NOT TRUTH; IT IS A MIRROR In the summer of 2026 I was assigned to predict the entire European Championship. I backed Italy even though the side had no standout global superstar. The basis was not the attack. It was the qualifying defensive data: roughly 0.6 xG conceded per match, the lowest among the entrants. Italy reached the final against England on July 11, 2026, and it unfolded in an unlikely way. England took the lead in the second minute through Luke Shaw. Italy equalised through Leonardo Bonucci in the 67th. On expected goals England edged it, 1.9 to 1.1. On the final result, Italy won on penalties. xG is not truth; it is only a mirror — but a mirror does not know how to lie. That mirror showed that across the tournament Italy controlled the quality of chances opponents created. It did not show that a penalty shootout is a near-independent game of probability, detached from everything before it. It did not show that a missed penalty in the 88th minute may have less to do with technique than with the weight of a nation. In esports, the equivalent moment is a deciding final series. A team can win a whole event on one correct play in the last second of game five. If the analytics department concludes that they were "better" because of that result, it is selling a story, not an analysis. THE QUALIFYING-ROUND TRAP AND THE LIMITS OF PROBABILITY Back to Kazan. What I got wrong in Germany's defeat was not the formula. What I got wrong was ignoring data that never sat inside the model: the actual intensity of the match. A side holding 74 percent of the ball can take 26 shots without creating a single genuine chance, if the opponent drops into a block that is compact and patient enough. Germany's xG was double South Korea's, but it was generated in a state of deadlock, where every shot was fired from distance into traffic. The fix I adopted afterwards was to read the opponent's pressing metric — the number of passes allowed before possession is regained. A high figure for the opponent means the possession side is being pushed back. Combined with context, the picture changes completely: Germany controlled the ball but not the space, and the model only saw the first half of that sentence. At short-tournament level the error is larger still. A World Cup gives a finalist seven matches. Seven matches is a small sample. In small samples, variance dominates. A strong team can be eliminated by a penalty, a red card, a refereeing error, or accumulated fatigue from a compressed schedule. A model tuned on a long season does not automatically transfer to a short tournament. I added a permanent section to every short-tournament forecast: a list of small-sample risks, with a recommendation to down-weight any high-confidence conclusion. In esports the problem is sharper. Single-game group stages carry enormous noise. A team far stronger in theory still has a meaningful chance of losing simply because one game did not go to plan. Best-of-three and best-of-five formats do not erase variance, but they reduce the influence of luck on the final result. Any analysis that concludes something about a team's true strength from a single-game result is misreading the data structure the tournament itself provided. THE NULL REPORT: WHEN THE CORRECT OUTPUT IS SILENCE This is the least-discussed part of my job. A professional analytics pipeline has layers. The first collects raw data. The second checks integrity. The third analyses. When the first layer returns an empty dataset — no events, no entities, no timestamps — the analysis layer is not permitted to infer. The correct output is a report that states plainly: insufficient information to conclude, with a list of what must be supplied. That sounds obvious. But in sports media, where performance is measured in articles published per day, a null report is the last thing anyone wants to file. That pressure creates a dangerous habit: filling the blank with speculation dressed as analysis. I have seen it in many forms. A player is injured, there is no official recovery timeline, and immediately articles appear asserting he will return in six weeks. A club has not published its transfer budget, and immediately a figure is offered as if it exists. A tournament has not announced its format, and immediately an analysis appears about the impact of that format. Every time, a blank gets filled with a source-less assumption. And every such assumption, repeated often enough, becomes a citable fact. In my own work I keep one rule: if there is no source, I say there is no source. If the data is insufficient, I say it is insufficient. If a conclusion rests on a single match, I state that the sample size is one. This rule makes my writing slower and less attractive to some editors. But it is my entire credibility, and credibility is the only thing in this trade that cannot be bought back once lost. There is a logical trap even careful people fall into: conflating absence of evidence with evidence of cleanliness. No information about wrongdoing does not mean wrongdoing is absent. No reporting of an organisation's financial problems does not mean the organisation is healthy. The silence of data is silence, not confirmation. I stress this because it bears directly on the integrity of the industry. A well-run sport is not one without scandals; it is one whose detection processes are strong enough not to depend on whether someone happens to speak up. THE CONTRARIAN ANGLE: THE PROBLEM IS NOT THE MODEL When a model fails, the default public reaction is to distrust models. That reaction is understandable and mostly correct in the specific case — a bad model produces bad conclusions. But at industry scale, the biggest problem sits on the other side of the equation: the demand for certain answers. Sports markets consume certainty. An article that says "I do not know" generates no engagement. An article that says "there are three scenarios with probabilities of 45, 35 and 20 percent" does not spread. An article that says "this team will definitely win" spreads, and if it is right the author is a genius, and if it is wrong the author already banked the audience. That incentive structure pushes analysts toward declarations. Each time, the gap between the strength of the claim and the quality of the evidence widens a little. Eventually the whole industry forgets that accuracy rate and the loudness of a claim are two entirely different quantities. The counter is to make humility part of the product rather than an apology attached to it. Publish the sample size. Publish the uncertainty range. Publish the assumptions. And most importantly, publish what the model cannot see. TAKEAWAY: SIGNALS TO WATCH NEXT ROUND A season is a scripture and each match is a verse — do not rush to recite half of it. Over the coming cycle, three signals are worth tracking. The first is how data providers publish their definitions and methods, because that is the foundation on which cross-provider comparisons become either valid or meaningless. The second is how tournaments publish formats and competition builds, because the noise of a single game will decide the real value of every power ranking. The third is how many analyses dare to state "insufficient information" instead of filling the blank with speculation. I am not waiting for a perfect model. I am waiting for an analytics culture more honest about what it does not know, because that is the only thing that holds trust across seasons.