Category Mislabeling: When an Algorithm Files a Family Story into a Football Data Sheet

For more than twenty-eight years of tracking the sports data industry, I have grown used to opening a file and matching its content against the label attached to it. But on a recent morning, at my desk in Shanghai, I had to read a file three times because the label and the content were so far apart. The label stated clearly: Domain Football. Yet what appeared as I scrolled was a conversation with Angelina Jolie, her stories about her sons, about Maddox and Pax working as assistant directors on set, about the afternoons she spent waiting while her children took flying lessons. Not a team. Not a player. Not a table. Not an xG figure, not a PPDA index, not a transfer fee. Only one very simple truth that anyone in this trade must remember: a wrong label can turn an entire downstream chain of analysis into a pile of meaningless inference. Do not trust a number before it has told its story from the beginning. That is what I keep telling young colleagues whenever they hand me a report labeled in haste. And this time, I was the one who had to inspect my own system again. When I went back through the entire two-layer analysis attached to this piece, I noticed something more interesting than the error itself. Its structure. A deep professional analysis divided into nine sections: tactical and technical analysis, club finance and the transfer market, results and the public-opinion cycle, league landscape and team positioning, rules and governance compliance, management and the dressing room, risk profile, media narrative and expectation, and transmission across the football industry. Nine frameworks, nine lenses. And in all nine, the conclusion was the same: not applicable. There was no football content to evaluate. This is not an article about football that was analyzed incorrectly. It is an article that never belonged to football in the first place, pushed by an automated system into exactly the slot it did not belong to. In other words, we hold a rare specimen: a clean example of how an automated data pipeline can slip, and what that slip leaves behind for everyone who comes after and trusts the label. Let me tell you an older story, so you understand why I am so obsessed with labels. In 2026, when I analyzed Hulk's transfer from Zenit to Shanghai SIPG at fifty-five million euros, I used a cumulative xG model and showed that his actual finishing output was only about 0.28 goals per match, nearly forty percent below the expectation the media had painted. The article was attacked fiercely by fans. But three scouts from other clubs contacted me for the detailed report. I learned one thing that day: accurate numbers will find the people who need them. But I also learned something else, only later: an accurate number placed under a wrong label finds no one. It only creates noise. That is why, when I faced this mislabeled file, I did not dismiss it as a minor administrative error. I treated it as a clinical case. Start with the technical context. In today's sports data industry, most articles do not reach analysts by hand. They pass through an automated pipeline: source collection, content extraction, topic labeling, entity labeling, domain labeling, then dispatch into the analysis queue. At the entry layer, domain labeling is done by a text classification model, usually a machine-learning version trained on a fixed dataset. The model does not understand football the way we do. It only counts keyword frequency, recognizes distinctive phrases, and assigns probability to each label. The problem lives right there. An article about a famous mother bringing her sons into a film crew can contain words that overlap with keywords from another domain. In the older training structures of some models, the phrase assistant director once sat in the same cluster as assistant coach and academy. The word sons, in some outdated datasets, sat near descendants or pupils. And so the model labeled football onto an article football had never entered. I stress this because it is not rare. Across the years I have spent producing internal reports for clubs, I have encountered enough labeling errors that I wrote a cross-checking protocol for myself. An article about a player leaving the pitch injured can be labeled health. A piece about a stadium sponsorship contract can be labeled real estate. A story about a youth development program can be labeled education. These errors do no serious harm if the final reader is a sober editor. But when the final reader is a forecasting model, or an automated evaluation process, the labeling error becomes a chain error. When probability collapses, what remains is the essence of the match. But if you filed the wrong match from the start, that essence has no place to appear. Now walk through each framework in this document, to see that a labeling error is not just one wrong line in a spreadsheet. It is a systematic chain of loss. The first framework is tactical and technical analysis. In a proper football analysis, this is where I spend most of my space. I will talk about formation structure, how two midfields close the gaps, how an advancing full-back leaves a corridor, how a deep-lying striker dilutes the opposing defensive line. But in an article about the Jolie family, there is no formation to draw. No gap to measure. No structure to dissect. Tactical sophistication: not applicable. Execution: not applicable. Personnel fit: not applicable. Key data: not applicable. Four lines in a table designed to hold quantitative values now sit empty. That is the first loss. The second framework is club finance and the transfer market. This is the terrain I know best after eleven years living and working in China, where every transfer between the giants is a brand arms race more than a purely sporting calculation. I usually start with revenue structure, the share of broadcasting income, commercial income, the wage bill, net debt. But in this mislabeled article, broadcasting revenue: not applicable. Commercial revenue: not applicable. Wage expenditure: not applicable. Net debt: not applicable. No deal to assess. No financial structure to challenge. What this document has, if we must name it, is a form of intangible capital: family brand equity, professional relationships, career opportunities for children. But that is social capital in the film industry, not financial capital in football. I remember once when a second-tier club asked me to analyze a loan deal in the winter window. I spent three days tracing the origin of the fee figure and discovered that the fee published in the media was nearly forty percent higher than the actual fee in the contract, with the difference pushed into performance payments that were, in practice, nearly unattainable. That is the kind of number I hunt. But to hunt it, I must have a correct label. If I read a film article thinking it is a transfer story, I will waste three days of my life chasing a ghost. The third framework is results and the public-opinion cycle. In a football analysis, this is where I compare a team's current position against preseason expectations, look at recent form, weigh the fixture factor. But here, the sample is zero. No matches to assess. No data to compare with results. The only thing left is a different kind of public pressure: pressure on a public figure asked about balancing family and career. Pressure level low. Source of pressure, media. Likely consequence: a continued positive news cycle. That is a healthy opinion cycle, but it is not a football opinion cycle, where every defeat can push a manager to the brink of dismissal within forty-eight hours. The fourth framework is league landscape and team positioning. Here I usually draw a competitive map: what tier this team sits in, who its direct competitors are, the squad-value gap, financial power, academy output, the direction of talent flow. But in this article, there is no league. No hierarchy. No competitive map. Every entity is an individual in the entertainment industry, and though cinema has its own talent pyramid, from entry-level crew to director and producer, comparing it to the football pyramid is only an analogy, not an analysis. And analogy, in my work, is the most dangerous thing after a mislabel. The fifth framework is rules and governance compliance. This is where I check financial fair play, transfer registration rules, disciplinary sanctions, competition eligibility. But with this article, everything is not applicable. No governance issue is raised. No precedent to cite. The only thing that could possibly be relevant, if we wanted to discuss child labor law and union rules in the film industry, becomes moot because both of her sons are adults. Again, nothing to comply with, nothing to violate, nothing to judge. The sixth framework is management and the dressing room. This is the framework I enjoy most when analyzing a club, because it is where data meets people. The quality of recruitment decisions, structural stability, manager-player relations, generational transition, media pressure on individuals along the age curve. All of that, in the mislabeled article, is not applicable. What remains is a public-relations statement about family cohesion and career opportunity, staged with skill. The line I am just Mom, placed in a dressing-room context, would be a highly notable signal about how a powerful figure wants to position herself. But placed in a film-interview context, it is just a pretty sentence. The same words, two reading modes. That is the power of a label, and also its danger. The seventh framework is risk profile. In a football report, I sort risk into five groups: sporting, financial, personnel, rules, public opinion. Here, the first four do not apply. The fifth, public opinion, leaves exactly one line: the risk of labeling this article as football, probability medium, impact low, mitigation is to correct the label. That is a short but accurate conclusion. And it reminds me that in our industry, the biggest risks usually do not come from the pitch. They come from the data lines we type at eleven at night, when our eyes are tired and our hands have left the keyboard. The eighth framework is media narrative and expectation. This is where I read the temperature of public opinion: what the current story is, whether it is accelerating or decaying, whether it has solid foundations, whether the sample size supports forecasting, and how long it will live. In this article, the current story is about a famous mother balancing family and career. Narrative sustainability is medium. Expected lifespan is short, under a month, until the next news cycle about the figure. And the ratio of social-media heat to fundamentals is low, because the piece has no viral hook. All of that is valid, but all of it belongs to another section. Placed next to transfer numbers, they do not speak the same language. The ninth framework is transmission across the football industry. This is the framework I consider most important in any strategic report. It draws a transmission diagram: from the talent supply chain in academies, through clubs and competitions, to the broadcasting, commercial, and derivative markets. But in this article, that diagram has no single connection. Impact on the talent supply chain: neutral, none. Impact on the agent ecosystem: neutral, none. Impact on broadcasting and commercial markets: neutral, none. Impact on capital networks: neutral, none. Impact on derivative markets: neutral, none. Impact on the national-team ecosystem: neutral, none. Six lines, six zeros. That is the clearest sign that we hold an article that does not belong to the field it was assigned to. And what is frightening is that, if no one checks, this article will sit in our football database forever, waiting for some algorithm on some day to extract it and drop it into a transfer report. Empty stadiums, but data has never been without an audience. That line holds here in a different sense. There is no stadium here, no audience here, but the data still sits there, still waiting to be read. And if it is read by someone who believes the wrong label, it will generate wrong analysis. People will start asking whether her two sons are a kind of trainee of a special academy. People will start comparing their career opportunities with the path of a youth player promoted to the first team. People will start using football's economic models to analyze a family relationship. And so a small labeling error becomes a large chain of distortion. I have watched the same thing happen in our industry many times. In 2026, when competitions paused and then played behind closed doors, I collected Premier League data from 2026 to 2026 and compared it with the post-lockdown sequence. The home-win rate fell from 46.2 percent to 38.4 percent, while average goals per match rose by 0.6. I sent a forty-page report to a club fighting relegation. They hired me as a set-piece analysis consultant, something independent of crowds. But to produce that report, I spent the first two weeks doing exactly one thing: rechecking whether every match in my dataset was truly played behind closed doors. Some matches were recorded as crowdless though a limited number of spectators were present under health rules. Some were recorded as having crowds when the number was so small it could not be treated as an effect. Had I skipped that label check, my entire model would have been wrong. And wrong in a way that is very hard to detect. That is why I say a labeling error is the most serious of all data errors. It is not a value-layer error. It is a meaning-layer error. You can fix a wrong number by rerunning the model. But you cannot fix a wrong meaning with any algorithm, because the algorithm itself has been placed in a context that does not belong to it. This is the point where I want to pause and challenge myself a little, because I know a reader might say: yes, but this is just a small error, one article filed in the wrong section, what is the big deal. My answer is this. In one season, each club in a top national league plays about fifty matches across all competitions. Each match has thousands of data points. Combined, a club can generate hundreds of thousands of data points per season. If the labeling error rate is only one percent, we are talking about thousands of misread data points. And in an analytical environment, a thousand misread data points do not sit scattered and harmless. They cluster. They drag other points with them. They skew the model. And when the model is skewed, decision-makers look at a picture that does not exist. That is the real cost of a supposedly small labeling error. But I must be fair as well. Not every labeling error comes from carelessness. A large share comes from the nature of language. Natural language has no labels. Labels are what we impose on language. And when we impose a labeling system on something fluid, there are always blurred border zones. An article about sport can contain a passage about business. An article about business can contain a passage about sport. Where is the boundary. There is no absolute answer. Only pragmatic rules, tested and adjusted over time. I do not look at the price board, I look at the signature of the money flow. That line is usually used for the transfer market. But it also holds for the data flow. Do not look at the number, look at the signature of the number. The signature of a number includes its origin, its timestamp, who created it, how it was labeled, and why it exists in the dataset. If the signature of a number does not match the label it carries, you must stop. That is the first principle I teach any colleague entering sports data analysis. Back to the mislabeled article. What I want to stress, as a data person, is that we need a human check layer in every automated pipeline. I know that sounds contradictory coming from me, a man who lives on data and models. But precisely because I live on data, I know its limits better than anyone. Models are good at recognizing patterns. Models are poor at recognizing the anomalous. And the anomaly here is an article about a famous family, full of emotional and human-relationship elements, sitting in a dataset designed to measure defensive-line distances and pressing speed. A human looks at those two things and instantly sees the mismatch. A model, if trained well enough, might also see it, but it will not know what to do next. It will not know that the mismatch should be handed to an editor, not pushed further down the analytical layer. That is the gap only a human can fill. There is something interesting I found reading this nine-framework analysis. The writer never flinched. They walked through each framework, marked not applicable wherever it was needed, and kept discipline to the end. That is the right attitude. In our trade, there is a great temptation, when data is absent, to start inventing data. People write sentences like the momentum is unclear or the team has good spirit, sentences with nothing to verify. This analysis did not do that. It preferred blanks to decoration. And in a sense, that very honesty turned a labeling error into a lesson of value. I want to pause here and speak about a theme I consider most central to this period of the year. We are in the middle of the regular season. This is when teams begin to reveal their true nature. No more early-season excitement and promises. Only numbers remain. The teams contending at the top have shown they can endure the rhythm. The teams struggling at the bottom have shown they cannot. And in the space between those two groups, some teams are quietly changing. Teams the table cannot explain. In the last three matches of a few mid-table teams, I noticed an index few track: the number of times a team recovers the ball in the final third of the opponent, compared with the number of times it loses the ball in that same third. This ratio, when it shifts from above one to below one, often announces a collapse in results four to six matches later. Not because it is a perfect index. Because it reflects something the table hides: an imbalance between ambition and execution capacity. A team that pushes many players high to recover the ball but lacks the structure to hold it after winning it will soon pay the price. And when the price comes, it usually arrives as goals conceded from situations nobody remembers the start of. I say this not to predict any team's result. I say it to stress that every good conclusion must begin with a correct label. If you mislabel your data, you will misread the trend. And if you misread the trend, you will make wrong decisions. A manager who believes his team is pressing well while in fact it is being pierced through midfield will choose the wrong training plan. A sporting director who believes a player has good xG when in fact he is only shooting a lot will pay the wrong price. An owner who believes his club is developing while in fact it is falling behind will delay change. All those mistakes begin with a single label. A match lasts only ninety minutes, but its story is longer than a season. And the story of a data point is the same. It begins the moment it is created, passes through many hands, many systems, many uses. If at any point along that journey someone slaps a wrong label on it, the whole story downstream drifts. And in the end, when you look back, you no longer know what that number truly wanted to say. That is why I keep the habit of tracing the origin of every number I use. Before using any statistic, I flow back to where it was born. I want to know when it was collected, by whom, under what criteria, and how many relabelings it underwent. This is a time-consuming habit. It makes me slower than those who simply open a spreadsheet and read. But it also keeps me from mistakes others make. And in this trade, avoiding one serious mistake is worth far more than writing one fast news item. Data never tires, only the reader of it tires. I say that to colleagues whenever they prepare to submit a report at day's end. But it is also true for me. After twenty-eight years, I still have not stopped tiring when reading data. Only, after twenty-eight years, I have learned to tire usefully. I spend my fatigue on the numbers that matter and ignore the numbers that merely decorate. That is how I keep my sanity in an industry where everyone wants fast answers. Back to the clinical case. After confirming the labeling error, I did three things. First, I corrected the article's label. Second, I logged the event into an error diary I have kept throughout my career, so it can be traced later. Third, I sent a short note to the pipeline operations team, asking them to review the labeling process for articles containing phrases related to assistant and sons, since that was the direct source of the error. This is not a grand action. But it is a systematic one. And in data work, a small systematic action beats a large unrecorded one. I tell this story not to criticize anyone. I tell it because I want you to see that even the best systems have days they slip. And when they slip, how we respond matters more than the error itself. If we correct the label, humbly look at the fault, and improve the process, a small labeling error becomes an investment in reliability. If we ignore it, treat it as unworthy of concern, it will return, elsewhere, in another form, with greater force. There is one aspect I want to dig deeper into, because it relates directly to my daily work and to anyone reading this with a professional purpose. The problem of duplicate entities. A proper name can belong simultaneously to a footballer and to a figure in another field. A club's name can match a company outside sport. A job title can be used in two contexts with two meanings. This is the most common source of entity-labeling errors, and it is especially dangerous in our industry, where an article about a business operating a club can be confused with an article about the club itself. I once analyzed a deal and discovered that two articles compiled from two sources were describing the same event but with two completely different transfer fees. One said eighteen million euros. The other said twenty-three million. Both cited the club as the source. When I traced it, I found that one figure was the base fee and the other the total including performance payments. Both articles were right in their own way. But if someone aggregated the two without careful labeling, they would create a third, entirely wrong fee and push it into a database no one could later fix. That is why, whenever I use a transfer figure, I always state whether it is the base fee or total fee. I always record the source and publication date. I always place it in the context of other transfer figures in the same period. None of those steps is superfluous. Each is a layer of protection against labeling error. And in an industry where information moves faster than verification, those layers are the only thing keeping us from drifting. I want to close this analysis with a thought on the nature of probability. Every number we publish carries a degree of uncertainty. That is natural and unavoidable. But there is a kind of uncertainty we can control: uncertainty of meaning. What this number means. Where it belongs. What it speaks about. If we control that kind of uncertainty well, the other kinds become easier to accept. We can say our model is seventy percent accurate at predicting outcomes, rather than saying we do not know what we are predicting. With this mislabeled article, we had a chance to look back at our process. An article about a famous family does not belong in a football data sheet. Confirming that is not a failure. It is a success of the checking process. And in an industry where we are often too busy with the next match to look back at past mistakes, spending time to fix one label is an action worth recording. When probability collapses, what remains is the essence of the match. But when a label collapses, what remains is the truth about your data. And that truth, however small, is worth recording honestly. I will track whether similar labeling errors appear in other data pipelines this season. Because I believe that in the coming months, as the title race and the relegation battle tighten, the pressure of speed will make automated processes slip more often. And what to do then is to stand still for one second and ask a simple question: does this number truly belong where I am placing it.