Mislabelled: How a Fuel-Price Report Ended Up in a Tennis Database

INTRODUCTION Last September, as the hard-court season wound down and the players eased into the brief pause before the Asian swing, I sat in my small Paris apartment to do something most people in this trade consider pointless: rereading every data label in the injury archive I had built over seven years. The work is like taking stock of a warehouse. You are not looking for anything new. You only want to be sure everything sits where it should. But just as I was about to close the laptop, a familiar tag with unfamiliar contents appeared on the screen: tennis. I opened the record and found fuel prices inside. Not the price of a racket, not the price of a ticket, but the price of petrol: a 4.42-rupee rise per litre for petrol, a 6.10-rupee rise per litre for diesel. That was the first time I encountered an energy record sitting inside my tennis database. It made me repeat something I often tell my sports-analytics students: data never lies; only the way we read it is wrong. This time the person reading it wrong was not me. It was the machine that had labelled the record. CONTEXT: THE LABELLING MACHINE Before going further, I need to be clear about the environment I work in. Modern sports analytics runs in three layers. The first collects raw data: match results, injury records, medical reports, workload metrics, calendars. The second labels and classifies, attaching a subject to each record: football, tennis, basketball, or energy, finance, politics. The third is where people like me read, cross-check and draw professional conclusions. The problem sits in the second layer. Most labelling systems today run automatically. They scan keywords, measure the frequency of certain terms, and infer a subject. When they work, they save thousands of hours of labour. When they fail, they create a silent form of pollution that end users struggle to detect, because the faulty record sits among correct data wearing a label that looks legitimate. The record I found is a perfect example. It was about a fuel-price adjustment in Pakistan, about Brent and WTI crude, about a petroleum regulator and a national pricing mechanism. Inside it were a few phrases that happened to overlap with the keywords my tennis system uses for recognition: rising figures, a sixth consecutive cycle, the largest adjustment since some point in time. The machine does not read meaning. It only counts patterns. And it labelled an energy report as tennis. I tell this story not to blame a particular algorithm. I tell it because I once made exactly this kind of mistake, only in the opposite direction. In 2026, while interning at the Paris FC youth academy, I trusted a data pattern without checking where it came from. Since then I have understood that bad data is more dangerous than no data. ANALYSIS: DISSECTING ONE ERROR ACROSS NINE LAYERS To prove how destructive a single stray record can be, I ran it through the same nine analytical layers every tennis expert uses today. The outcome was identical at each layer in one respect: the system returned conclusions that looked professional but were entirely meaningless. Layer one, technical and tactical analysis. This is where playing style, surface adaptability and clutch-point ability are assessed. When the input is a fuel-price report, this layer has nothing to read. No player, no match, no rally. If anyone forces the system to generate a technical verdict from that data, the result can only be fabrication. And the frightening part is that a system can fabricate fluently. Layer two, data and form. Here, first-serve points won, return points won and break-point conversion are calculated. The energy record has none of these. What it has is a 4.42-rupee per litre rise for petrol, a 6.10-rupee per litre rise for diesel, Brent up 2.6 percent to $107.33, WTI up 2.5 percent to $102.56. If an automated model mistakes this for player form, it will draw a six-cycle rising curve and call it a winning streak. A commodity streak is not a human winning streak. Layer three, tournament system and schedule. Tiers, points, mandatory-entry rules and calendar position are assessed. The energy record carries two dates: an effective date and an earlier review date. These are pricing-administration cycles, not competition cycles. But a weak system will read those two dates as two consecutive rounds and imagine a player who has just completed a four-day run. Layer four, the tour landscape and player positioning. Players are sorted into contender, seed, backbone and fringe tiers. The energy record contains no player. The named entities inside it are an energy regulator and a government energy division, two administrative bodies, not two players. This is the most dangerous kind of error: the system sees no human name, yet stands ready to create a fake character profile. Layer five, rules and governance. Officials check medical time-outs, off-court coaching, the serve shot clock, anti-doping and match integrity. The energy record concerns a domestic pricing mechanism, a matter of domestic energy regulation. Not one tennis clause appears inside it. If a tournament's compliance unit received this record by accident, it would hunt for a violation that does not exist and miss the ones that do. Layer six, team and player management. Coaching quality, support-team completeness and agency roles are evaluated. The energy record has none of these. The decision-makers named are state institutions, not coaching staffs. This shows how a labelling system can turn a policy decision into a tactical decision with one wrong click. Layer seven, risk. This is the layer I care about most, because it is my livelihood. Analysts build a risk matrix: injury risk, points-defence risk, career risk, rules risk, media risk. The energy record carries macro-economic and supply-chain risk: supply disruption up to four percent of global supply, attacks on shipping in the Middle East. These are real risks, but they are not a player's risks. Mixing them into an injury matrix is poisoning the model on purpose. Layer eight, media narrative and expectation. Frenzy levels, the gap between market expectation and real strength, and narrative sustainability are measured. The energy record tells a cost-of-living and energy-security story, a sixth consecutive adjustment, global crude above one hundred dollars. It is a complete media narrative, except it has nothing to do with tennis. If the system pushes it into sports coverage, it will create a false wave of expectation around a character who does not exist. Layer nine, industry transmission. This is the only layer where I can imagine a link, however weak. Fuel costs can affect tournament travel, ticket prices and the budgets of smaller federations. But the energy record I found never mentions tennis. Any link is speculation beyond the source. When the data has not spoken, a good analyst must know how to stay silent. WHAT HAPPENED IN THE LABELLING LAYER The key point is that a labelling system does not understand content; it recognises patterns. An article about fuel prices can contain a whole set of keywords that a tennis filter accidentally catches: rise, fall, record, consecutive, streak, highest since. These words appear densely in both energy reports and sports reports. The filter counts them, finds the density high enough, and assigns the label. I call this a semantic-homonym error at the system level. Two documents from entirely different fields can share up to forty percent of their vocabulary. A machine that relies only on vocabulary will never tell them apart. To tell them apart, it needs context; to understand context, it needs a human verification layer. I do not believe in luck; I believe in verified numbers. The fuel-price record slipping into a tennis archive is not a rare accident. It is a symptom of a common disease across the whole sports-analytics industry: we invest heavily in collecting data and very little in checking what that data actually belongs to. WHY DIRTY DATA IS MORE DANGEROUS THAN NO DATA Paris FC taught me that bad data is more dangerous than no data. I wrote that line after an afternoon in 2026 I will never forget. Back then I was twenty, a third-year sports-analysis student interning at the Paris FC youth academy. I was assigned to review the medical files of the U19 squad. Among those files, one player caught my eye: an eighteen-year-old midfielder named Lucas Moreau. In his previous fourteen matches he had suffered three hamstring flare-ups, yet the coaching staff kept starting him. Read only the results column and everything looks normal. His run of appearances was impressive. But when I charted injury frequency against training load, the picture changed. I calculated that if he kept playing at that rate, his risk of a muscle tear would reach eighty-seven percent. I presented the finding, and the coach reluctantly gave him a week off. As a result, Lucas avoided a serious injury and scored twice in the next three matches. The lesson was not that I was clever. The lesson was: had I looked only at the results column, I would have missed the entire story living in the health column. A beautiful table can hide an approaching disaster. And a contaminated table can invent a disaster out of nothing. The 2026 World Cup is the second example. When Germany were eliminated in the group stage, the football world poured over tactical analysis. I did not follow the crowd. I dug into the fitness file of Mesut Özil, who started all three matches despite signs of wrist tendon inflammation and ankle pain. Cross-referencing the data, I found he covered only sixty-eight percent of the distance he had covered in the previous season at Arsenal. Germany collapsed not because of tactics, but because physical warning signs were ignored for five months. Both stories lead to the same conclusion: the quality of the data determines the quality of the conclusion. A model running on dirty data will produce conclusions that sound highly convincing and are entirely wrong. That is why I begin every analysis by checking the data's provenance, not by running the model. 2026: BUILDING A CLEAN MODEL MYSELF In 2026, when the pandemic paralysed football across Europe, I was twenty-three, freshly graduated and working as an analyst's assistant at a sports-data company in Paris. Everyone at the firm focused on tactical analysis, on vague predictions about how the league would return. I chose a different direction. I proposed building a model for re-injury risk after a disruption, based on data from past seasons that had been interrupted, such as the 2026 Ligue 1 strike. I gathered one thousand two hundred medical records from five clubs. The result showed a twenty-three percent rise in muscle tears in the first four weeks after football returned. My boss approved it, and the model later became a diagnostic tool for several lower-tier clubs. What I learned from that project was not the twenty-three percent figure. What I learned was that when data behaves abnormally, people are eager to believe anything that looks scientific. That is the ideal condition for dirty data to breed. Had I let a record that did not belong to football slip into my dataset back then, my model would still have run, still produced a figure, and that figure would have been used to decide the health of real footballers. WHY I CANNOT ACCEPT THIS ERROR Some will say: one stray record among hundreds of thousands is no big deal. The error rate is under one in a million. I disagree, and my reason is structural, not emotional. First, a labelling error does not stay put. It replicates. A mislabelled record becomes a training sample for later models, then generates new labels from those models, and those new labels become samples for the next cycle. A single speck of dust can ruin the whole machine if the machine has no filter. Second, a labelling error makes no noise. A wrong injury model does not shout. It only returns a probability. And decisions about player health are often made on probabilities, not on bare facts. When a probability is distorted by a stray record, the consequences fall on specific human beings: a player pushed back onto court too soon, a knee damaged further, a career shortened by a few years. Third, a labelling error destroys trust. When fans discover that a sports report of theirs actually contains fuel-price data, they will doubt every other report. And once trust is lost, it is very hard to regain. In my trade, a reputation is built through thousands of correct calls and can be destroyed by one wrong one. What I am after is not a perfect system. What I am after is an honest one that knows how to say it does not know. WOMEN'S SPORTS DATA: WHERE LABELLING ERRORS HURT MOST If there is one field where data-quality problems hurt more than anywhere else, it is women's sports. For years, women's tennis databases have been far thinner than men's: fewer matches encoded in detail, fewer workload metrics recorded, fewer medical files fully digitised. When the source data is already scarce, each stray record carries a larger share, and its impact on conclusions weighs more heavily. I once watched an analysis group use a women's tennis dataset to predict shoulder-injury risk. The first results looked very tidy. But when we reviewed the source, we found that a sizeable share of the records actually came from men's doubles matches that had been mislabelled by gender. Those records were not technically bad data. But they did not belong to the set, and their very presence skewed the entire baseline. This is why I argue that the fuel-price story is not merely a story about a technical bug. It is a story about how we are building the future of sports analytics on foundations we have never properly inspected. THE CONTRARIAN ANGLE: WE DO NOT LACK DATA, WE LACK DOUBT There is a widespread belief in sports today: the more data, the better. Clubs hire more specialists, install more sensors, sign more deals with data vendors. Everyone believes volume will lead to quality. My experience runs against that belief. The industry's problem is not a lack of data. The problem is a lack of doubt. We have too many records and too few questions about where they come from. We are good at measuring and poor at verifying. We are good at collecting and poor at filtering. Distance covered and sprint counts are packaged as effort metrics, but running in vain also produces beautiful numbers. A player who covers twelve kilometres in a match may have spent the whole match in the wrong positions. A vast dataset may be nothing but a pile of records no one has checked. If you hand me a large dataset without telling me how it was collected, I would rather work with a smaller one that has been verified. There is another paradox. Video-review times that run too long are shredding the rhythm of matches. Two minutes of waiting is enough to cool a goal. This is another expression of the same disease: we add tools without weighing their cost. An automated labelling machine is like an over-slow review system. It makes us think we are more accurate, when in truth we are slower and more fragile. So I propose a small but systemic change: every record, before it is used, must answer three questions. What is its provenance. What is its unit of measurement. What would render it invalid. If a record cannot answer all three, it should be quarantined from every decision-making model. WHAT IS RIGHT INSIDE A MISTAKE I do not want this piece to read as an indictment. A labelling error, in the end, is a free test. The fuel-price record slipping into my tennis archive taught me three things. First, my system needs a human checkpoint at the labelling layer, even if that checkpoint is just one person skimming ten records a week. The cost is tiny; the benefit is large. Second, labelling must rest on context, not only on keywords. An article about fuel prices and an article about tennis can share many words, but they speak about two worlds. Context is what separates them. Third, every database should keep a public error log recording not the correct calls but the wrong ones. My industry learns a great deal from injury cases but is reluctant to record data-error cases. I would say a database with no error log is an immature database. Here I must also turn the lens on myself. I once published a verdict on a player's re-injury risk based on data I later discovered had been collected during an abnormal competitive period. My verdict was not technically wrong, but it lacked a disclaimer it should have carried. Since then I add a line to every report of mine: data may change under abnormal circumstances. I wrote that line not to protect myself, but to remind myself that every figure comes with conditions attached. CONCLUSION The fuel-price report sitting in my tennis database is not a freak occurrence. It is an alarm bell for an industry that is growing very fast and checking very slowly. After seven years of decoding athletes' injuries, I have learned one simple thing: the flaw is rarely in the human body; it is usually in how we measure the human body. And sometimes the flaw lies in forgetting to ask where a record belongs before using it to judge someone. A risk model saves no one; it only tells you where to look. Our job is to look closely, and sometimes to stop in time. I do not believe in luck; I believe in verified numbers. But before I trust any number, I check whether it truly belongs to the story I am telling.