The File That Wore a Tennis Label: Why Dirty Data Is More Dangerous Than Empty Data

At 6:14 in the morning, a new file landed in my monitoring folder. The label field read, plainly: tennis. I opened it. Inside were 8,752.80 points, down 1.1% on the session, led by banks and miners. The first four names up were NAB, CBA, Westpac and ANZ. Then BHP, Rio Tinto, Fortescue. Then Woodside and Santos. Not a single player. Not a single set. Not a single first-serve percentage.

I read it three times, slowly, the way I read a box score after a match: once to find the anomaly, once to check whether the anomaly was mine, once to see whether it repeated. By the third pass I understood that what I was holding was not a typo. It was an Australian equity-market record wearing a tennis shirt, and it had passed through the domain-tagging layer without anyone stopping it.

The first thing I wanted to know was not how to fix the file. It was how many records of the same kind had already reached the models we place our trust in.

I called the person who owns the data pipeline. The answer was familiar: the system tags automatically by source, nobody inspects individual files. I asked who inspects the labels. Three seconds of silence. Those three seconds were worth more to me than a full season of statistics.

This incident sits at the lowest layer of a pipeline most viewers never see. A live sports broadcast, counted from the moment you hear a commentator read a number to the moment that number is produced, typically passes through five layers. Layer one is the raw source: an official data provider, a tracking camera system, or simply a wire story. Layer two is normalisation: bringing everything to one format, one unit, one time zone. Layer three is domain tagging: tennis, basketball, football, finance, weather. Layer four is the model: computation, comparison, forecasting. Layer five is air.

The first four layers can all fail. But only layer three fails in total silence. If layer two gets a unit wrong, you see it immediately: a player listed at 1.88 metres suddenly becomes 188 metres. If layer four gets a formula wrong, you see percentages that do not add to 100. When layer three is wrong, the system runs perfectly, still produces numbers, still draws clean charts. It is simply talking about something else.

A label, in the end, is a promise. When I tag a file 'tennis', I am promising every downstream layer that inside there are players, matches, surfaces, scores. The modeller downstream does not open the file to check. He trusts me. That is the entire foundation of modern analytics, and its entire vulnerability.

The actual contents of that file were an Australian equity-market report. The S&P/ASX 200 closed at 8,752.80 points, down 1.1%, with banks and miners as the two main drags. Among the banks, National Australia Bank announced roughly 2,500 job cuts as part of a restructuring. Among the miners, BHP reported lower copper output at the Escondida mine, pushing the supply outlook for the metal into a more cautious stance.

In energy, Woodside and Santos came under pressure from softer commodity prices. On the macro side, Australia's third-quarter inflation data and the Reserve Bank of Australia's meeting on 28 October were the only calendar markers the record mentioned. The market was pricing roughly an 89% probability of a rate move at that meeting.

Every one of those numbers is real, verifiable, and neatly presented. And not one of them carries any meaning for tennis. No player, no tournament, no coach, no governing body appears anywhere across the source's 25 information points.

When I ran the nine-dimension analytical framework over that file, the output was nine consecutive verdicts of 'insufficient information, cannot assess'. There was no serve to evaluate. No break-point conversion rate to compare. No ranking-points structure to decompose. No calendar, no surface, no head-to-head, no form curve, no injury risk, no media narrative, no industry value chain to trace.

Most people would read that output as a failure. I read it as the one moment in the week when the system did the hardest thing it can do: it refused to answer.

A model is only trustworthy when it is capable of saying 'I do not know'. If you build a system in which every input must produce a conclusion, you have not built an analytical system. You have built a fabrication engine with a nice interface. And in sport, that engine has an enormous market: millions of people waiting for someone to tell them what will happen tomorrow night.

Three modes of data contamination I have encountered in 25 years in this trade, and every one of them begins at a tagging layer nobody audits.

The first is mislabelling, exactly as happened this morning. A finance file is marked as tennis, or the reverse. The cause is usually mundane: a metadata field filled in wrong, a record pasted into the wrong batch, a manual entry process with a copy-paste error. The consequence is not mundane at all, because the modelling layer downstream will learn from it without knowing.

The second is cross-contamination. Two different data sources get merged into one table, one identifier key, one column. In sport this shows up most often when a league changes data provider mid-season. You have half a season computed under the old provider's definitions and half under the new one's, and a single column holding both. The average still produces a number. That number is meaningless.

The third is silent drift. Nothing is mislabelled, nothing is merged. The definition of a metric simply changes over time while its name stays the same. A pressing index is recalculated a different way. An expected-goals model gets a version upgrade. The column name is unchanged, the unit is unchanged, only the meaning has moved. Every cross-season comparison becomes a comparison between two different things wearing one name.

What makes all three hard to catch is that they do not produce obvious errors. The numbers stay in plausible ranges. The charts stay smooth. The model still produces forecasts that sound intelligent. Mislabeled data does not generate a display error; it generates confidence — and misplaced confidence sets off no alarm at all.

I have been on the other side of this lesson, which is why I no longer treat source-checking as a chore.

On the night of the Euro 2026 semi-final between Italy and Spain, I sat in the studio with two colleagues. Around the 60th minute, working off real-time data from the channel's tracking system, I said on air that Italy's pressing numbers were falling sharply and the coach would have to withdraw Chiesa within minutes. Chiesa was withdrawn shortly after. A colleague beside me blurted out a line of surprise, it was cut into a clip and spread across social media, and over the following two days I took dozens of calls from other networks.

What I remember from that night is not the applause. It is my boss's warning the next day: do not turn yourself into a prophet, because the audience will hold you to a standard you cannot sustain. He was right. A correct prediction does not prove a correct method. It proves that on this occasion, the data and reality happened to meet.

Since then, whenever I use real-time data on air, I attach the limits: this data does not measure a player's psychology, does not measure an unexpected tactical decision, does not measure a coach reading the game his own way. Stating limits does not weaken me. It turns my next error into a measured error rather than a stain.

In 2026, when every league stopped, I sat at home and did something nobody assigned: I collected data from 312 matches across the Premier League, La Liga and the Bundesliga in 2026-20, split into two groups — with crowds before the pandemic and without crowds in the closing stretch. The result made me reread the spreadsheet four times: home win rate fell from 46% to 38%, while average goals per match rose slightly, from 2.67 to 2.81.

Those two movements in opposite directions taught me something no software could. The world had changed; the data had not. Same league, largely the same players, same stadiums — but the social context around the match had been inverted, and every forecasting model built on historical data was now systematically off. Had I not split the groups by hand, I would have blended them into one average forever and called it 'a decline in home advantage'.

That is the third mode of contamination in its purest form: nobody mislabelled anything, the world simply changed while the label did not.

Conversely, in 2026, when I watched Josef Martínez's footage from Atlanta United fourteen times, I was not looking for a number to prove he was good. I was looking for why his conversion rate was so abnormally high, around 23.4%. The answer lay in his finishing style: very little wasted backlift, early contact, striking inside spaces defenders believed they had already closed.

I wrote a 1,200-word analysis. The content director called me in, offered one line of praise and one instruction: stop writing like a thesis. The next week I was given lead commentary on Atlanta United's match, Martínez scored twice, and I called him 'the quiet predator'.

The lesson there runs opposite to this morning's incident. In that case the data did exactly its job: it flagged an anomaly and forced me onto the pitch to find the cause with my own eyes. The number did not replace observation. The number pointed to where observation was needed. Statistics are only seasoning. People are the main course.

The modern sports data pipeline has become an industry in its own right, with exclusive contracts, camera infrastructure, and official data distribution deals for betting markets. Deals like the Genius Sports–NFL agreement on official data distribution exist for a very practical reason: sports data is worth money, and the accuracy of a label is what determines that value.

In tennis, automation has moved the same way. Hawk-Eye arrived at Wimbledon in 2026, and by the 2026 season the ATP rolled out Electronic Line Calling across its tour. Every step like that increases the data generated per match, increases the number of processing layers, and therefore increases the number of places where a bad label can hide.

The cost of a bad label scales. One contaminated file in one morning is one person's problem. One contaminated file inside a pipeline serving hundreds of broadcasters, dozens of forecasting models and millions of viewers is an industry's problem. And the uncomfortable part is that at that scale, the failure does not surface as an obviously wrong number. It surfaces as a new trend, a 'breakthrough finding', an analysis shared ten thousand times.

Sport is building an enormous data nervous system, and almost nobody is doing the pathologist's job: opening the tissue to see which cell does not belong.

The correct handling of this morning's file is not to try to extract something about tennis from the ASX 200. There is no player to assess technically. No form data to build a curve from. No tournament to assess structurally. No governing body to assess for compliance. No industry value chain to trace. Any attempt to fill those nine boxes would be fabrication, and fabrication in sports analysis has one distinguishing feature: it always sounds plausible.

And this is where I want to go against the grain of my own profession.

Sports analytics currently rewards confidence, not calibration. Someone who says 'I put 70% on this scenario, and here are the three assumptions it rests on' is read as indecisive on air. Someone who states it flatly and happens to be right once gets cut into a viral clip. The incentives lean heavily toward the second, even though over the long run it is worse for the audience.

I learned this the uncomfortable way. At the 2026 World Cup, before the quarter-final shootout between Russia and Croatia, I said on air that Russia had drilled penalties heavily, but Croatia had a goalkeeper who had already made saves in an earlier shootout, and I predicted Croatia would win 5-4. The result was 4-3. Right team, wrong score. A young colleague texted to ask why I had not committed to a sharper number. That question bothered me for a month, and I spent that month rewatching the tournament, logging every phase I had judged wrong, building a reconciliation sheet between prediction and outcome to find the blind spot in how I think.

The blind spot turned out to be simple and unpleasant: I had made a safe prediction out of fear of being wrong, not because the data only permitted going that far. I changed my method. I still make predictions, but always with a numeric confidence level and always with the assumptions that would collapse the call.

This morning's incident is the technical version of the same problem. A system that cannot say 'insufficient information' is a system that cannot audit itself. And a system that cannot audit itself will one day speak with great confidence about a data file that has nothing to do with what it thinks it is looking at.

Silence is the right answer here, and it is also the hardest answer to give on camera. Silence is not the absence of an answer — it is the answer, for those who know how to listen.

There is a deeper layer the industry has not looked at squarely. In many sports we have handed decision rights to invisible systems nobody elected. A patch in an esport can decide a championship more than a player can. A line-calling algorithm version can decide a set. A ranking model can decide who enters a main draw. These are invisible referees, and invisible referees do not hold post-match press conferences.

If we have handed power to invisible systems, auditing those systems has to become part of the job rather than an add-on. The commentator at the top of this profession now needs to read not only the match but the data pipeline behind the match — to know what it takes from where, and where it might be lying entirely by accident.

After all these years I still hold one belief that is hard to shake: most of the best answers in sport do not come from a spreadsheet. They come from sitting in a stadium long enough to see what no column records — the half-metre a player stands out of position, the way a coach checks the clock at minute 70, the way a crowd goes quiet before a free kick. A spreadsheet does not know what longing is, and we should stop pretending otherwise.

But the spreadsheet is also the only thing that tells us where we went wrong. It is not the enemy of observation. It is observation's error log.

On this morning's incident, the fix is clear and unglamorous: build a validation gate at the ingestion layer that checks the domain label against the actual entity types inside the file. If the label says tennis and the contents are ticker codes, index levels and margins, the gate must block it before any model touches it. The cost of that gate is measured in hours. The cost of a season of analysis built on contaminated data is measured in credibility, and credibility cannot be bought back with a correction notice.

One detail in this incident is worth more than the rest, and it sits on the other side of the picture. The nine-dimension framework did not try to force a conclusion. It returned insufficient-information verdicts at every position, cited which information points the evidence came from, and recommended fixing the label at the source rather than patching the symptom. That is the behaviour of a mature system. A mature system is not measured by how many conclusions it produces, but by how many times it dares to refuse one.

In the analytics room we tend to raise models like children. Feed them clean data, train them carefully, praise them when they get it right, excuse them when they get it wrong because 'the data wasn't enough'. The analytics room's favourite child eventually has to stand on its own two feet. And the first step in standing on your own two feet is knowing what ground you are standing on — rather than standing on an ASX 200 file labelled tennis.

So the question I am carrying into next week is not how to forecast better. It is this: if tomorrow your pipeline hands you a file with the right format, the right structure, every required field present, but whose contents belong to an entirely different world — do you have any way of noticing?

If the answer is no, then every number you confidently quote is standing on the same patch of sand. And in sport the sand is always shifting underfoot; we just tend to notice only after we have fallen.