When the Data Pipeline Returns Zero: Why an Empty Result Is Tennis's Most Honest Analysis

2:47 a.m., Brisbane time. The second monitor was still on. The spreadsheet sat open with twenty-seven column headers: first-serve points won, return points won, break-point conversion, net-point success rate, rally-length distribution, pressure index at deciding games, nine columns broken down by set. Every column had a name. No row had a number.

The data pipeline had returned zero. A dropped connection was not the cause. The API had not expired. The first layer — the layer that turns raw text into entities, into player names, tournament names, match dates, statistics — ran through its loop and returned an empty set. No names. No event. Not a single information point to hold on to.

I stared at that empty frame for ten minutes. The only thing I thought about was not how to fix the error. It was how long it would take before anyone noticed if I simply kept writing.

On average, in this profession, the answer is about forty minutes — long enough for a popular analysis piece to travel from a draft screen to a newsfeed, and not long enough for anyone to check a single number against a single source.

I shut the machine down at 3:10 a.m. The next day I published nothing. What follows was born out of that night of publishing nothing.

Context: the extraction layer and the price of an empty row

Tennis is the most densely measured sport per minute of play. Wimbledon installed Hawk-Eye in 2026 to adjudicate in and out. The Australian Open moved every court to Electronic Line Calling in 2026, and from the 2026 season the system covers almost the entire ATP circuit, removing line judges from the court at most events. Every serve now generates a velocity vector, a landing point, a spin rate, a flight time. Every rally generates length, direction changes, contact positions, court-opening angles. Tournament organisers publish millions of data points per edition, and most of that is pushed onto broadcast dashboards within two seconds.

Raw material is abundant. Raw material is not information.

My work sits in the middle layer. A decent analytical pipeline has three tiers. Tier one extracts: it turns text, match logs, scoreboards and press releases into queryable entities — who, where, when, what did they do. Tier two cross-checks: it pulls data from at least two independent sources, compares, and hunts for divergence. Tier three is writing. When tier one returns an empty set, tier two has nothing to compare, and tier three — if it runs anyway — has exactly one thing left to do: invent.

That is the whole problem of sports analysis right now, compressed into a single technical fault at tier one.

This industry pays for certainty. A headline with a percentage gets shared more than a headline with a confidence interval. A declarative sentence is read faster than a conditional one. The incentive structure of the content market pushes writers toward the far end of honesty, and most writers never notice the push, because it happens every day, every piece, every headline, until it becomes reflex.

I entered the profession in 2026 at the fact-checking desk of Sports Illustrated. The job was boringly simple: take a piece, find every number, call or email the source, confirm, flag. On my first day I sent back seven pieces out of eight. My manager called me in and said something I still remember: "You're not wrong. But you're doing a job the newsroom only pays for when it fails silently."

I understood that sentence differently later. Fact-checking only has value when it blocks an error. A blocked error is invisible. An unblocked error gets hundreds of thousands of reads.

The Vietnamese tennis market sits in a particular spot in that picture. Vietnamese fans watch Grand Slams seven to nine hours out of sync with Melbourne, which means an Australian Open final lands around 3:30 p.m. Hanoi time, while a US Open final lands around 3 a.m. A large share of the audience follows through licensed paid platforms, the rest through unofficial streams and three-minute highlight reels. Both groups receive the same product: a result plus a story. Very few outlets supply verifiable numbers.

Vietnamese tennis has a data gap of its own. Lý Hoàng Nam won the Wimbledon boys' doubles in 2026 alongside Sumit Nagal, collected multiple SEA Games golds, and broke into the world's top 250 in singles. Nguyễn Thùy Linh reached the WTA top 120. But no open database in Vietnam stores shot-level statistics for these players match by match. There is no domestic equivalent of the rally-charting projects international analysts use for free. Without data, a writer has two options: write with the eye, or write with memory. Both are polite forms of invention.

The evidence chain: four kinds of empty input, and only one worth writing

In my files, empty inputs fall into four categories, and they are not equal in value.

The first is empty because of a technical fault. A scraper breaks, an API key is revoked, a source format changes. That emptiness carries information about the system, not about the match.

The second is empty because the source does not yet exist. The match has not been played. The ranking has not updated. That emptiness is timely, and the correct response is to wait.

The third is empty because the subject has no data. A player ranked outside the world's top 300 competes at a low-tier ITF event that no shot-tracking system covers. That emptiness is a finding: it tells you where the sport's data map ends. That boundary is not technical. It is commercial.

The fourth is empty because the extraction layer ran correctly but found no entity to extract. That is the one I met that night. And it is the only one whose correct answer is to publish the emptiness.

An empty result is not an analytical failure. It is one of the most valuable outputs available, because it forces readers to face the question the whole industry avoids: what are we measuring, and how far does our measurement reach?

The 2026 Wimbledon final: 218 against 204

On 14 July 2026, Novak Djokovic beat Roger Federer 7-6(5), 1-6, 7-6(4), 4-6, 13-12(3) in the longest Wimbledon final in the tournament's history, lasting four hours and fifty-seven minutes.

In the fifth set, with Federer leading 8-7, the Swiss player served the sixteenth game at 40-15. Two championship points. Djokovic saved both, held, and won the tiebreak 7-3.

The notable number sits elsewhere. Federer won 218 points across the match. Djokovic won 204. The man who lost fourteen more points than his opponent lifted the trophy.

Feed that dataset to a pre-match model and ask it to price the title after being told the total point count, and the model says Federer wins. The scoreboard says Djokovic wins. Both are correct. Neither is sufficient.

This is where most tennis analysis stops looking. Total points won is an accumulated metric, and in a match decided by set structure, points are not equal in value. A point at 40-15 on serve in the fifth set at 8-7 carries far more weight than a point at 2-1 in the second set. Every statistics table adds points together as if they were identical. Nothing in that table says they are not.

That is why I built a pressure index as the twenty-seventh column in that spreadsheet. It weights each point by situation: break point, set point, championship point, tiebreak point, and the stage of the set in progress. The method is not new. It has existed for years in basketball and American football analytics. Tennis adopted it a decade late, partly because the sport has so many discrete points that people assume adding them up is enough.

Adding them up is never enough.

Roland Garros 2026: three points at minute three hundred and twenty

On 8 June 2026, Carlos Alcaraz beat Jannik Sinner 4-6, 6-7(4), 6-4, 7-6(3), 7-6(10-2) in a French Open final lasting five hours and twenty-nine minutes — the longest final in the tournament's history.

Alcaraz lost the first two sets. In the fourth set, at 4-5 and serving to stay in the match, he fell to 0-40. Three championship points for Sinner.

Alcaraz saved all three, held, took the fourth-set tiebreak 7-3, then took the fifth-set tiebreak 10-2.

I keep this match separate because it demonstrates a mechanism that data cannot defend itself against. The same raw dataset — similar match length, similar total point counts, similar error density — differing only in the outcome of three specific points, produces two entirely opposite stories.

Had Alcaraz lost one of those three points, the next day's story would read: Sinner has completed the transfer of power on clay, Alcaraz lacks the nerve for the biggest moments, and a chain of analysis would be written to prove it using the very numbers that already existed beforehand.

Had Alcaraz won all three, the story reads: steel nerves, extraordinary willpower, and another chain of analysis is written to prove it using the very numbers that already existed beforehand.

Same data, same method, two contradictory conclusions. The only error lies in someone taking the final result as a hypothesis and working backwards to find evidence.

In my tracking log, that match is filed under "deciding points over total points played": roughly one percent. One percent of the data generates one hundred percent of the story.

This is where I want to be direct with people who do this work: if a conclusion of yours disappears when you rotate the dataset by three points, that conclusion never existed.

The empty-stadium laboratory, and what it actually measured

In June 2026, when tennis returned after the pandemic shutdown, I began a before-and-after comparison that ran for nearly a year. I placed Grand Slam and Masters 1000 matches played with crowds in 2026 alongside their equivalents played without crowds in 2026 and 2026, and tracked seven metrics: first-serve percentage, first-serve points won, double faults per hundred service points, break-point conversion, break points saved, tiebreaks per hundred games, and unforced errors per hundred points.

Football had already taught me a large-scale lesson about how the competitive environment changes behaviour. The crowdless season is the cleanest laboratory elite sport has ever had, and tennis is the only sport where that laboratory ran almost intact: same rules, same surfaces, same scoring system, with a single variable removed.

My results were unremarkable. The crowdless group showed first-serve percentage rising slightly and consistently, fluctuating around one to two percentage points depending on the tournament. Double faults fell. Tiebreaks per hundred games rose. Break-point conversion fell among lower-ranked players and barely moved among the top tier.

The conventional reading of these results produces an easy-to-sell conclusion: crowds distract players, so without crowds serving improves.

I did not write that sentence. Because my data cannot support it.

At least four other variables moved during that window and I cannot isolate them. The calendar was compressed, and many players arrived with far fewer preparation matches than in a normal year. Match balls were changed at some events for supply-chain reasons. Weather and scheduling were disrupted, particularly at the French Open, moved to September instead of May. On-court staffing rules, including reduced line judges and ball kids, changed the rhythm between points.

Those four variables are enough to destroy any causal claim. What I have is a correlation, measured over a short window, in a context that will not repeat. That correlation has value as a hypothesis. It has no value as a conclusion.

This is why I place a "model limitations" section at the end of every analysis, and make it at least a third the length of the conclusions. That section is not a ritual of modesty. It is a map of the holes a reader needs to know about before quoting me.

The small-sample trap, and why break points are the most misunderstood metric

Break points are the most quoted and most misunderstood metric in tennis.

Here is why. A player in a three-set match might create only four to six break points. Across a five-match tournament, three sets each, that might reach twenty-five to thirty. At that sample size, the gap between a 30 percent conversion rate and a 50 percent rate amounts to roughly five or six points. Five or six points in a sport where a single point can turn on a net cord.

Which means: the "best break-point converters" tables you see on statistics sites after every tournament are almost always noise rankings. The sample is too small to separate signal from randomness, and those tables never print a confidence interval next to the number.

I once offered a simple test to editors I worked with: take a tournament's break-point conversion table, shuffle the order of matches, recalculate, and compare. If the ranking scrambles substantially purely because the order changed rather than the data, the table has no predictive value.

Most of those tables scramble substantially.

This does not mean break points are unimportant. It means the meaning of a break point lives in its situation, not in its count. A break point on your serve at 4-5 in the fifth set is a fundamentally different object from a break point on your serve at 5-1 in the first. Adding them together and dividing does not produce information. It produces a number.

This is where I share ground with football analysts, and also where I separate myself from the crowd. Transfers are where people pay hundreds of millions to buy a row in a spreadsheet, and tennis has a smaller version of the same disease: academies, player managers, sponsors, all making decisions on rankings built from small samples over short windows.

The data stream flowing toward bookmakers

There is a part of this story the sports analytics industry rarely discusses, and I consider it the darkest part of sport's digitisation.

Shot data, positional data, serve-speed data, per-rally data — all of it is collected by the operators running measurement systems at tournaments. Those operators sell access. One of the fastest-paying and highest-paying customer groups is betting companies and in-play odds providers.

Which means: precisely the data I use to analyse a match, in raw form and at real-time speed, is delivered to the betting desk before I finish downloading my own file.

I have no objection to data analysis. I object to the incentive structure it creates.

Over the past decade, integrity reports from international tennis governing bodies have repeatedly placed tennis among the sports with the highest number of suspicious match alerts, with most alerts concentrated at lower-tier events — where prize money for a single round is many times smaller than the amount that can be wagered on that same round.

That is a structural asymmetry. A player ranked outside the world's top 350 at an ITF event might earn a few hundred dollars for a win. Global betting turnover on his match can exceed his entire annual income. When that gap is wide enough, data stops being a tool for describing a match. Data becomes a commodity for pricing someone else's risk.

This is why I write about the third layer of my pipeline — the layer of data that does not serve viewers. Very few tennis analyses mention it, because it has no images, no moments, nobody to interview.

Vietnam's tennis data gap and what it costs

Back to Vietnam.

When a Vietnamese player competes at a Challenger in Asia, statistics for that match barely exist in public form beyond the final scoreboard. No serve-placement distribution. No rally lengths. No average contact positions. Fans at home follow a live scoreboard, then read a short match report.

This gap has concrete consequences. It turns every debate about a Vietnamese player's progress into an emotional debate. When Lý Hoàng Nam loses, people say he lacks fitness. When he wins, people say he was inspired. There is no dataset to separate those two hypotheses.

I tried to fill part of that gap by charting matches myself. For three consecutive years I rewatched footage of Vietnamese players at regional events and hand-logged every rally into a spreadsheet. Each match took four to six hours for a two-hour contest. I completed more than forty matches before abandoning the project because I could not sustain it alongside my main work.

Forty matches is a dataset smaller than any academic threshold. But it was enough to produce one observation I have not read anywhere else: among the Southeast Asian players I charted, the share of points won from the fifth stroke onward was markedly lower than from the third stroke, and that gap was wider among Vietnamese players than among Thai or Filipino players in the same sample.

I draw no conclusion from it. Forty matches, three countries, one chartist, no cross-checking. But it is a testable hypothesis, and testability is the minimum standard. The claim "Vietnamese players lack fitness" is not testable. Mine is.

The contrarian angle: the empty result is the most suppressed product in the industry

I once thought the problem with sports analysis was a lack of data. After years in the work, I believe the problem lies elsewhere: the industry has no mechanism for publishing empty results.

There is no slot for a piece headlined "I don't know". There is no slot for an analysis that ends by admitting the dataset cannot support a conclusion. The incentive structure of content platforms punishes that product on every measurable axis: read time, shares, comments, click-through.

The result is a system generating thousands of conclusions every day from datasets that cannot support conclusions, with no accountability when those conclusions turn out wrong — because by the time they are wrong, the reader has moved on to the next one.

At a deeper level, this is a form of organised intellectual arrogance. Writers believe they have the right to conclude because they have data, when most of the time they have data rather than enough data. The distance between those two things is the entire gap between analysis and fortune-telling.

There is a lesson I learned in 2026 and still repeat in every conversation with young people entering the field. In 2026 I learned that a 95 percent probability still contains a 5 percent that knows how to laugh. I built a prediction model for a major tournament, gave the favourite a title probability above 23 percent, and wrote a piece declaring that the data had identified the champion.

That team went out in the quarter-finals. The champion was the team my model ranked fourth, at just over 11 percent.

The error was not in the model. A 23 percent probability for the strongest team in a knockout tournament is a perfectly reasonable forecast, and it was numerically correct. The error was in my sentence. I turned a probability distribution into a declaration, and that declaration had no basis in any model I ran.

It took a month to rebuild the algorithm after that tournament, adding variables for club minutes played before the event and the mental state of key players. But the biggest change was not in the algorithm. It was at the top of every piece, where I now force myself to state the confidence interval before stating the conclusion.

The first data rebellion was never about overthrowing anyone — only about proving that the numbers deserved to be heard. But numbers deserve to be heard only when the person presenting them accepts stating the part the numbers do not cover.

There are three temptations a sports analyst always meets, and I see them in myself every week.

The first is turning data into a weapon. When you hold a dataset and the person opposite does not, using it to end the argument is a natural reflex. But a dataset does not end an argument. It moves the argument from a dispute about judgement to a dispute about method, and most writers are not trained to argue about method.

The second is treating every surprising result as a contrarian signal. Going against the crowd once earned me my most-read pieces, which is exactly why I set myself a threshold: a phenomenon is only called a signal when it repeats across multiple independent samples, or when it has an explicable causal mechanism. One surprise is noise. Three surprises with the same mechanism is a signal.

The third is treating a dissenting opinion as evidence of missing data. This is the most dangerous temptation for someone with a bias toward order and neatness. My critic may be right for reasons I cannot measure, and my inability to measure it does not make it disappear.

What current data cannot tell me, I set aside and state plainly. For tennis, that list currently has four items. Public data does not reveal a player's actual injury status, only whether they withdrew or played on. Public data does not reveal sleep quality, flight schedules, or accumulated fatigue. Public data does not capture ball and court conditions at a level of detail sufficient to compare across tournaments. And public data barely exists at lower-tier events, where most of most players' careers take place.

Those four items are not an apology. They are the boundary of what I am permitted to say.

What to track in the next cycle

From the courts, I hear the breath of a match more clearly when the stands are empty. That was true in 2026, and it remains true in a different sense now: when the noise outside is stripped away, what is left is structure.

Three signals I will be tracking in the coming cycle.

First, the transparency level of Electronic Line Calling data. Since the system spread across almost the entire ATP circuit, a vast amount of ball-position data is generated at every match. The question worth tracking is how much of it will be released to the public and how much will flow only into commercial channels. That boundary will shape who is permitted to analyse tennis over the next decade.

Second, the periodic integrity reports. The number of suspicious match alerts and their distribution by tournament tier is the only indicator showing where the data stream is flowing and what consequences it is producing. I will compare that distribution between reporting periods, not just read the headline figure.

Third, the emergence of an open database for Southeast Asian tennis. If that happens within a few years, the entire debate about Vietnamese players changes in nature. If it does not, we will keep reading analyses written with the eye and presented in the voice of statistics.

Data does not lie; it is the people reading it who make excuses. And the only way to tell an analyst from a storyteller is to see what that person writes on the days when the data pipeline returns zero.