Injury-Data Labeling Failure: When a Drone Report Lands in a Tennis File

On Thursday evening I opened a batch of files during an injury-data audit for a tennis analytics system. The eleventh file carried a clear label: tennis. I opened it.

Inside was a military communiqué. The Saudi-led coalition said it had intercepted a Houthi drone near Makkah. Spokesman Turki al-Malki stated the incident was stopped. A Houthi political bureau member countered through the group's own news agency. A little further down, one paragraph described the 1,200 km (745 mile) East–West Pipeline linking the Gulf oil fields to the Red Sea. Another mentioned that up to 4% of global oil supply could be at risk.

No player. No set. No ranking, no schedule, not a single line of medical record.

I closed the file and sat still for a few seconds. The weight did not come from the content inside. It came from knowing exactly what would happen if this file kept flowing downstream.

In seven years of working with sports data, I have never seen an injury-risk model collapse because of missing data. I have seen them collapse because of mislabeled data.

My job is to analyze injuries and the return-to-play process. Every model I build starts with an operation that sounds dull: labeling. A training session labeled high intensity. A match labeled hard court, three sets. A hamstring flare labeled recurrence. Labels are the skeleton of the whole system. Without correct labels, every calculation behind them is arithmetic on sand.

Sports data companies now process hundreds of thousands of records a day: match logs, medical files, training diaries, workload trackers. Nobody reads each record by eye. Most of the work goes to automated labelers. Those labelers are excellent at classification and terrible at self-questioning. They never ask: does this record actually belong to my world?

In 2026, as a third-year student interning at the Paris FC youth academy, I was assigned to review the U19 medical files. I found midfielder Lucas Moreau, 18, had suffered three hamstring flares across fourteen matches while the coaching staff kept starting him. I charted injury frequency against training load and showed he faced an 87% risk of a muscle tear if he kept playing. The staff reluctantly gave him a week off. Lucas avoided a serious injury and scored twice in his next three matches.

I retell that story because the lesson sits somewhere else. What I did that day was simply read the label on a file correctly. If Lucas's file had been mislabeled "healthy," no model on earth would have saved him.

That gave me a principle: the biggest vulnerability in any sports data system sits at the labeling stage, not the calculation stage. I find the gap not in the athlete's body but in how we measure it.

So what did that file contain, and why does it matter?

In content, it is a geopolitical and energy report. The Saudi-led coalition says it intercepted a drone near Makkah, the site tied to the Hajj pilgrimage. The Houthis say otherwise. There is an infrastructure detail: the East–West Pipeline carrying Gulf oil to the Red Sea. There is a market detail: a share of global oil supply placed at risk. Names of senior political figures appear too.

In data terms, the file carries four warning signals any sports analyst should recognize, because they are identical to signals I meet in injury records.

Signal one: interested-party sourcing. The interception claim comes from the spokesperson of the very coalition that carried it out. The opposing side speaks through a news agency it controls. In an injury file this is the club saying the player is "fine" while the player says "it hurts." Both sides have reasons to say what they say. A one-sided claim without independent sourcing is not data; it is testimony.

Signal two: internal contradiction. One passage refers to a "six-month US–Iran war." Another refers to "nearly seven months of fighting." The figures do not match. In injury analysis, that kind of duration conflict is a classic marker of a file blended from multiple sources, or of a compilation that was never cross-checked. I have seen medical reports saying "three weeks out" on the first page and "six weeks out" on the second. That is rarely a typo. That is two different files merged into one.

Signal three: no publication date. The report uses the present tense yet references July 2026. Without an absolute timestamp an event cannot be placed on a timeline, and without a timeline there is no sequence analysis. To me, an injury without a date is an injury that does not exist.

Signal four, and this is the fatal one: the wrong label. The entire file is tagged tennis while its content is armed conflict and energy markets.

Four signals. One alone and I would take a note and move on. All four and I halt the process.

What happens if I do not halt? Picture an automated system at a sports data company receiving this file. The labeler sees the tennis tag and pushes it into the training set of an injury-risk model. The model learns that somewhere, at some point, an "event" existed with attributes: a 1,200 km pipeline, 4% of supply, nearly seven months. None of those attributes map to hamstrings, ligaments or training load. But the model does not know that. It only knows weights.

This is where I want readers to slow down. The harm is not that the model mispredicts one injury case. The harm is that the model becomes more confident for no reason. Injected noise dilutes the real signal and raises the model's apparent reliability on the validation set. I saw this during a 2026 project building a model for injury recurrence after disruption. We collected 1,200 medical records from five clubs, drawing on seasons that had previously been interrupted, such as the 2026 Ligue 1 strike. The result showed a 23% rise in muscle tears in the first four weeks after football resumed. But to reach that 23% figure we had to discard hundreds of mislabeled records. Keeping them would have produced a prettier number. And a wronger one.

Since 2026 I write every report as a weighted scenario, never saying "certain." I always add a disclaimer: data can shift when abnormal conditions arise. It sounds cautious, but it is the only way an analysis avoids turning itself into a prophecy.

Based on my experience watching matches across many seasons, broadcast data suffers from exactly the same disease. Once I saw a statistics table list a clay court as a hard court, and that single mislabeled cell was enough to make an entire spin-rate comparison between two players meaningless. Viewers noticed nothing. The numbers still looked fine. But the conclusion drawn from them was wrong at the root.

During that audit I distinguished two kinds of error. The first is missing data: empty files, absent dates, untracked players. That error is loud; everyone sees it. The second is mislabeled data: complete files, standard formatting, impeccable appearance, but content belonging to another world. That error is silent. It raises no alarm. It goes straight into the model.

The Makkah file is the second kind in its purest form.

The sports data industry carries a dangerous bias: the belief that more data will fix everything.

I understand why the bias exists. Machine learning models often improve with more samples. Sponsors like big numbers. And in a major tournament season, the pressure to publish fast pushes label checking into the paperwork category rather than the analysis category.

But distance covered and sprint counts have long been packaged as effort metrics, while ineffective running produces equally handsome numbers. We are used to doubting the metrics. We rarely doubt the label attached to them.

The common mistake is blaming the model. When an injury prediction system fails, the default response is to gather more data, more variables, more seasons. But if the water source is contaminated, pouring in clean water does not make the reservoir cleaner. It only makes it bigger.

Paris FC taught me that bad data is more dangerous than no data. With no data, you know you are blind. With bad data, you believe you can see.

I do not believe in luck; I believe in numbers that have been verified. And for me, verification starts with the smallest question: is this label correct?

A risk model saves no one; it only tells you where to look. If it tells you to look at a file about a drone near Makkah, it is wasting your time and, worse, eroding your trust in the system itself.

The question I carried away from that audit was not how to build a better model. It was: who is checking the labelers? And in this major tournament season, as every data stream flows faster than ever, that question still has no answer.