The Empty Spreadsheet in Brisbane: Why Tennis Analytics Must Learn to Say 'Insufficient Information'

It was 6:40 in the morning in Brisbane. The sky was already bright and the air so humid I had to raise my monitor contrast by two notches. The left screen, as on every January morning, held the 52-week cycle tracker for the 32 players who would contest the Australian Open. The right screen held the nine-section analysis framework I use for every major report.

That morning the framework came back empty. Nine sections, eighty-seven cells. The only cell with text in it was the domain label: tennis.

I sat still for nearly four minutes. It is the kind of silence a nine-year veteran still has to relearn.

Filling it is dangerously easy. "This player won 68% of his tie-breaks last season." "His second-serve defence on the deuce court ranks among the tour leaders." Nobody checks. Nobody cross-references. A sentence like that reads as professional, and it can be entirely invented.

Context: a blank cell with a signature

My data pipeline runs in four layers: fetch, parse, classify the domain, then extract into structured fields. If the classifier dies, the domain label comes back blank. But the label said "tennis". That means the machine read the genre, knew this was tennis content, and still failed to retrieve a single line of the body text.

This is a failure signature, not a finding. The fetch layer broke: a paywall, a JavaScript-rendered page, or a truncated connection. All of those produce the same shape — a structurally intact skeleton with nothing inside.

Something similar once happened to me, in a different sport with different consequences. In 2026, before the World Cup in Russia, I built a model from six major tournaments of historical data, using Elo ratings and qualifying records. The model gave Brazil a 23.4% chance of winning, the highest of the 32 teams. I wrote a long piece declaring that the data had identified the champion. France, whom my model ranked fourth at 11.2%, lifted the trophy. Brazil went out in the quarter-finals. In 2026 I learned that a 95% probability still has a 5% that knows how to laugh.

The lesson was not that the model was wrong. Every model is wrong. The lesson was that I presented a probability as a promise. I later added variables for club minutes played and player mental state, then rebuilt the algorithm from scratch. Since then, every analysis I publish ends with a section titled "model limitations". Demanding readers go there first.

Core: telling "no data" apart from "low risk"

In medicine, a patient who never shows up for an X-ray does not get stamped "healthy". The chart says "not imaged". In tennis analysis, that distinction is erased almost daily.

If I have no injury report on a player, the correct sentence is "no medical data available". The wrong sentence is "no sign of injury". The wrong one sounds safer, more professional, and it converts the absence of information into a positive conclusion. That is the most dangerous error in this profession, because it does not incriminate itself.

I use a rule I call the minimum-information gate. An analysis may only leave my desk when it carries three things: at least one named entity, at least one verifiable fact, and an absolute date. Without a date, every figure in the piece becomes a rumour with a decimal point. "He wins 71% of tie-breaks" means nothing without a date: which season, which surface, across how many attempts.

To see the gate at work, take a real case. Daniil Medvedev reached the Australian Open final in January 2026, leading Jannik Sinner by two sets before losing 3-6, 3-6, 6-4, 6-4, 6-3. A Grand Slam runner-up finish is worth 1,200 ranking points. Twelve months later, at the Australian Open 2026, Medvedev lost to Learner Tien in the second round in five sets. A second round is worth 45 points.

The difference: 1,155 points.

On the ranking page, that fall looks like a form crisis. On the calendar, it is arithmetic. The same backhand, the same serve — but the 52-week cycle turned once and reclaimed last year's reward. No shot got worse. Only the ledger got struck through.

That is why I separate two concepts most coverage blends together. The ranking is an accounting book, not a measure of level. Elo is the measure of level. The ranking counts what you collected over 52 weeks, with deductions. Elo estimates how strong you are right now. A player can drop four places while playing better than he did last month. Another can climb three places without changing anything, simply because the man above him lost points. Data does not lie; it is the people reading it who make excuses.

In tennis, points-defence pressure has a very specific shape. It is not spread evenly across twelve months. It piles into a cliff: a large block of points expiring within two weeks, usually landing right at a surface transition. A player who made the semi-finals of a hard-court Masters in March must defend those points exactly when his body needs to adapt to clay. Add a flight, a time-zone shift, and a minor ankle complaint, and you have the formula for a season derailed by administration more than by technique.

The same logic applies to tools viewers rarely notice. Protected ranking lets a player returning from long-term injury enter the main draw using the ranking held at the time of injury rather than the current one. Lucky loser entry backfills the main draw after a late withdrawal. A wild card goes to a home player with no explanation attached. Combined, those three mechanisms can reshape an entire draw section, and none of them appears on a scoreboard.

Then there is the human element. A medical time-out is a legitimate recovery tool and an equally legitimate tool for breaking an opponent's rhythm. To find out whether it is being abused, you do not read the match report. You count: which set it appeared in, how many consecutive points had been lost, and what percentage of service games the player won immediately afterwards. When I ran that count across a sample of 60 matches, the signal fell inside the noise. I wrote it into the model limitations: sample too small, no conclusion.

The 2026 season gave me a rare clean sample. When tournaments returned to empty stadiums, the measurement conditions became ideal: same rules, same surfaces, same players, only the noise differed. From those empty stands, I could hear the match breathing. What I found in my small sample was a slight rise in serving advantage and fewer breaks. But I have to be explicit: that is correlation, not causation. The absence of crowds may not be the cause. A compressed calendar, autumn weather in Europe, heavier balls in cold damp air — any of those could be the cause. I have not isolated the variable.

And I have to add something this profession usually avoids. The same live point-by-point feed that runs into my analysis room also runs into betting exchanges in the same second. Live data supplied to betting companies is the darkest side effect of sports digitisation: it turns a medical time-out, a shrug after a missed point, into a trading signal. Insiders do not fully see what they are selling. I offer no betting insight of any kind, and I write this not to sound virtuous, but because match data deserves to be used for something else.

Contrarian: not every anomaly is a trend

There is a temptation that twins the temptation to invent numbers: the temptation to treat every anomaly as a discovery.

In 2026, a piece of mine was spiked for going against the newsroom consensus. A national team lost its opening match and was written off as tactically cowardly; my data showed they had generated the highest volume of high-quality chances in the group stage. The editor rejected the piece on the grounds that it contradicted the general feeling. A week later that team reached the semi-finals, and my article became the most-read piece of the month. But the thing I remember is not the vindication. It is the uncomfortable confidence of holding a beautiful number next to a conclusion I had already reached. The first data rebellion was never meant to overthrow anyone — only to prove that numbers deserve to be heard.

Counter-intuitive data is only worth publishing when the phenomenon repeats across samples, across surfaces, across seasons. A player winning three five-set matches at one tournament is luck. The same player winning eleven of fourteen five-setters across three seasons is a technical trait, and it has to live somewhere in his body: the legs, the breathing rhythm, or the ability to hold shot structure when tired. The gap between those two things is not a matter of feeling. It is a matter of sample size.

At the same time, I have to hold space for what the data cannot say. My data counts points, not fear. It knows a player lost a tie-break; it does not know how many hours he slept the night before. It knows how many seconds a medical time-out lasted; it does not know what the doctor told him. Recognising that boundary does not weaken this profession. It makes it more honest.

Takeaway

That morning in Brisbane, I did not fill in the blank framework. I called operations, asked for the fetch log, checked the status codes and the full page-render trace. Forty minutes later the source material appeared in full. The fault was a single line of JavaScript — and had I not stopped, I could have published a fluent nine-section analysis of a match my system had never read.

The signal I will track next cycle is not the form of any player. It is something smaller and harder: whether we still have the nerve to hand back an empty cell. In a season where every feed needs a new hero each week, the ability to say "I have not measured that" is a professional skill, not a weakness.

And if your tracker has never had a blank cell in it, you may be filling it with things you never measured.