The Empty Data Table and the Silent Trap of Tennis Analytics

Sydney at night, the second screen still glowing as I reopened the data extraction file for a tennis piece. Title: N/A. Source: N/A. Article type: unclassified. List of information points: empty. Only one label survived — "tennis." A data table stripped of all content, still carrying the tag of a sport.

I sat still for a few minutes. What chilled me was not the empty file but the possibility that followed: if I let the analysis system downstream keep running, it would still produce a report that looked entirely reasonable. It would have players. It would have first-serve percentages. It would have a week-by-week form chart. It would have conclusions about the decisive rallies. Not one line of it would come from real data.

Numbers whisper. Those willing to listen will hear an entire match. But when the numbers fall silent, the hasty will invent a voice on their behalf.

The story sounds like a dry technical error, yet it touches the weakest point of professional tennis analytics. Since Hawk-Eye covered nearly every major court, since each serve was broken into dozens of variables, analysts have lived inside a feeling of safety: everything can be measured.

Tennis has a peculiarity compared with football. Every point begins with a serve, so the data is highly structured and repetitive. That makes missing data harder to detect. When a serve table has all its columns, very few people check whether each cell was actually filled from the match.

I once believed in that feeling of safety. In 2026, I wrote a World Cup prediction piece based on xG and was called a bookworm who knew nothing about football by a group of people on Reddit. When Croatia reached the final, a journalist from The Athletic contacted me to ask how I calculated "defensive xG prevented." I spent two weeks writing Python, cross-checking against StatsBomb data, and sent back a 17-page breakdown.

The lesson was not "data wins." The lesson was that readers' trust comes from my explaining every step that produced the data, not from a tidy conclusion.

In tennis, the process is stricter. A single match generates thousands of data points: first-serve speed, second-serve speed, first-serve points won, second-serve points won, return points won, break-point conversion, the ratio of winners to unforced errors. Each metric has a different origin — some recorded by Hawk-Eye, some manually entered by the tournament's statistics team, some estimated by algorithm. Blend them without noting the source and you get a pile of numbers that looks consistent but is actually out of phase.

Around 2026, some organizations began standardizing data for the top-tier events, such as the Tennis Data Innovations joint venture between the ATP and its media arm. But smaller events, qualifying draws, and matches away from centre court still regularly fall into a grey zone. That is where data is most easily filled in with assumption.

Take a verifiable example. The Isner–Mahut match at Wimbledon 2026 lasted 11 hours 5 minutes, spread across three days, with the fifth set ending 70–68. Isner fired 113 aces. That figure is fully sourced: Wimbledon's organizers published it, and the major statistics providers all cross-checked it.

But if you take 113 aces alone and conclude "Isner served far better," you have just skipped an important variable: Mahut also served 103 aces in that same match. And both men served on grass, where the ball travels fast and low. A serve metric only means something when read alongside the surface, the opponent and the physical state — a core principle that every tennis data table quietly accepts but rarely states.

Now back to the empty file on my screen. My analysis system was built on the principle of an "evidence chain": every conclusion must trace back to an original information point. No information point, no conclusion. That is a fence against fabrication.

The problem is that the fence only works if someone checks it. In practice, modern models tend to be designed to "try to answer" rather than to "refuse to answer." Give them an empty input and they will still find a way to fill the gap with familiar assumptions. They know roughly what first-serve percentage a top-10 player usually posts, so they assign a plausible value. They know a grass match usually produces fewer break points, so they insert an estimated range. The final report is still packed with numbers.

This is the crux: a model designed to always answer will turn the emptiness of the data into the emptiness of its conclusions — while dressing them in a look of completeness. Readers have no way to tell a measured metric from a guessed one.

Based on my experience following matches, I always have to note which parts are naked-eye observation and which are measured data. That note makes my reports uglier, shorter, less "judgmental." In return, the demanding reader trusts me a little more.

Before you trust a number, ask where it was born. That question applies to my own numbers too.

The instinct of analysts is usually this: more data, more safety. That instinct is half right.

The wrong half lies in the fact that data has not only volume but structure. An empty table that still carries the "tennis" label is the most dangerous kind of failure, because it triggers exactly the processing mode of a tennis piece — except it has no guts. The system does not see that it is missing anything. As long as the label survives, everything else looks intact.

In tennis, this kind of failure appears in unexpected places. A post-match stat table with all its columns but half its cells defaulted to 0. A model predicting break-point conversion that cannot tell the difference between "there were no break points" and "the break-point data is corrupt." A headline about a "form slump" when the losing streak is only three matches long, two of them against top-5 opponents.

Plenty of data that stays silent about how it was made is more likely to lull you than a scarce table with sourced notes.

A season missing detail is like a match missing stoppage time. You think you have the whole match, but the decisive part lies in the very gap you skipped. For a player, that decisive part can be a fifth set after four hours of battle — something no aggregate statistics table can capture without point-by-point data.

I once turned down a request to write a piece explaining "football without fans" during the pandemic, because I needed three more weeks of data before I could be sure. My home-advantage model then priced it at 0.45 goals per match; after nine rounds without crowds, it fell to 0.08. I had been wrong to ignore the crowd variable. The editor wanted the piece to run immediately. Caution sometimes costs you an opportunity, but it keeps you from losing credibility.

Mixing up one variable is like losing your bearings for an entire year. With that empty data file that night, the only choice I considered correct was to tag it "not for downstream use" and run it again from scratch.

The fascination lies in a question without a complete answer. If tennis is moving toward using language models to summarize matches, write automatic commentary and build on-demand stat tables, then who will be the gatekeeper stopping those models from inventing a match that never happened? Organizers already have Hawk-Eye and the statistics providers to verify every serve. But no one has built a similar verification layer for the words that run on top of the data.

Perhaps the most interesting work of the coming season is not measuring one more metric but designing a gate: an automatic refusal layer that says plainly, "I don't know," whenever the input data is empty. I don't yet know who will lead that trend. But I am fairly sure it will come — because models keep growing more confident, and data is not always full.