When the Spreadsheet Returns Blank: The Limits of Sports Data

3:47 a.m., in a small apartment on Lạch Tray Street, Hai Phong. I sat in front of the screen, waiting for the data pipeline I had built over two weeks to finish its final pass. It finished. No errors, no crash, no red warning line. Just a nearly blank page: tournament name empty, athlete name empty, information points empty, source empty. In the bottom corner, a green line reported success.

That kind of failure is far more unpleasant than a system crash. A crash tells me where to fix things. Returning an empty result that still looks valid means the system is lying through silence — and the reader on the other end has no way to detect it. I call it silent failure. The most dangerous data pipeline is not the one that breaks, but the one that returns blank and then reports success.

Context: a profession built on verification

I entered the trade at twenty, starting as a fact-checker at a sports magazine, then spent fourteen years at a large newsroom before moving into data journalism. Those years taught me something simple: most errors in sports journalism come not from malice but from confidence that arrives too early. A reporter watches a match, sees the home side take eighteen shots, feels the momentum, and writes that the home side deserved to win. That feeling may be correct. But it has not been verified, and the gap between those two states is where I work.

In 2026, midway through the V-League season, I published the first series applying an expected-goals model to Vietnamese football. In the match between Hai Phong FC and SLNA at Lạch Tray, the home side generated 1.92 expected goals but lost 1-0 through an individual error. The media called it a slump. I called it random injustice, because the opposing goalkeeper made eleven saves, 3.8 times his own season average. The piece was mocked for two weeks. In the third week, the Hai Phong head coach publicly cited my data table at a press conference, and the story turned.

From then on I set myself an inviolable rule: no verified data, no conclusion. That rule sounds simple until you have to enforce it on a night when the spreadsheet returns blank.

In June 2026, before the Germany–South Korea group-stage match at the World Cup, I published an analysis built on two metrics: Germany's PPDA pressing coefficient had fallen from 8.1 in 2026 to 12.6 in 2026, and average distance covered per match had dropped by 6.2 kilometres. I wrote that Germany trusted possession too much and had forgotten how to win the ball back early. Germany held 74 percent possession, lost 0-2, and went out in the group stage. Germany collapsed inside my spreadsheet before it collapsed on the pitch. But that story left me with a different scratch, which I will address later.

The core: anatomy of a silent dataset

That night in Lạch Tray, my pipeline had three layers. Layer one collected raw data: match sheets, serve statistics, break points, ball-in-play time. Layer two normalised and labelled. Layer three interpreted and produced conclusions. That night, layer one ran fully and returned complete data. Layer two ran fully and returned an empty table. Layer three received the empty table, had nothing to interpret, and returned exactly what it received: blank. And because layer three had no alarm for empty input, it recorded the status as complete.

The frightening part lies there. A system that cannot distinguish between "no data" and "no finding" will always return the same result for two entirely different situations. To a reader, blank easily reads as "no problem." To an analyst, blank means "I do not yet know." Those two readings lead to opposite actions: one sits still, the other goes looking for the cause. Throughout my career I have seen many sporting decisions made on precisely this confusion.

My fix for silent failure was to build a fourth layer, the quarantine layer. It does one thing: it counts. If the number of input records is greater than zero but the number of output records is zero, it halts the entire pipeline and applies the label insufficient evidence. That label is allowed to appear in the final article. I once published an analysis consisting of exactly one sentence: not enough data to conclude. A colleague called it evasion. I left it as it was, because a piece that admits it is empty is better than a piece that fills the gap with guesswork.

In tennis, this failure mode appears more often than people think. A player wins 78 percent of first-serve points in a match; that figure sounds dominant. But if the first-serve in percentage is only 48 percent, then most of those points were built on the second serve, and the second serve is the metric that most exposes pressure. To judge properly, I have to lay three data layers side by side: first-serve in percentage, points won on first serve, and points won on second serve. A single layer always tells a distorted story — not because the data is poor, but because the person reading it skipped the other two layers.

Break-point conversion is even more troublesome. A player can create fourteen break opportunities in a match and convert only two. Looking at a 14 percent rate, people conclude the player is mentally weak. But the denominator of fourteen opportunities is rarely uniform: some come after a burst of high-speed first serves, some come when the opponent has just double-faulted twice in a row. They are all called break points, but their difficulty differs so much that lumping them into one percentage is a methodological error. Every shot is a hypothesis. xG is how we test it. And like any hypothesis, it has value only when we know the conditions under which it was measured.

I always arrange evidence in a three-tier structure, like a court file. The first tier is precedent: historical metrics, head-to-head records, last season's trend. The second tier is current evidence: match data, surface context, schedule. The third tier is the verdict, accompanied by an explicitly stated margin of error. This order must not be reversed. When I see an analysis that opens with a conclusion and then goes hunting for data to back it, I know it will collapse at the second tier.

One example of how much context matters. In the 2026 season, when Wimbledon was cancelled for the first time since the Second World War, most remaining events had to be played in empty stadiums or with limited crowds. Roland Garros moved to late September, and Rafael Nadal won the title there on 11 October 2026 after beating Novak Djokovic in three sets. That same year, the US Open took place in New York without spectators, and Dominic Thiem won after overturning a two-set deficit against Alexander Zverev. The empty stadiums of 2026 were not an exception, but the cleanest laboratory modern football has ever had. What happened in that laboratory deserves comparison against every other season, rather than being set aside as an anomalous year.

Based on my own experience watching matches, one thing has repeated often enough to become a rule: crowds affect serving more than returning. A server in silence tends to find rhythm faster and hesitate less at important points. A returner loses the visual signal from the stands that normally helps keep footwork rhythm between points. If that holds broadly, then every statistic on serve points won collected in 2026 needs its own correction factor. I have not published that factor, because the sample is still too small and I do not want to repeat an old mistake.

The contrarian angle: when silence is more trustworthy than noise

The most counter-intuitive thing I have learned in twenty-five years is this: in sport, an empty conclusion is often more trustworthy than a full conclusion assembled too quickly. The sports analysis industry is governed by the pressure to have an opinion. After every match, people need a story, and if the analyst does not supply one, someone else will. That competition pushes everyone toward the conclusions that sound most plausible rather than the ones that are most correct.

The paradox is that my own data systems are caught in the same pressure. When I write an automation script, I must choose what happens when the input is empty. The easiest option is to let it pass and keep running. That option produces a report that looks complete, with full headings and full sections, missing only content. That is exactly what happened on the night in Lạch Tray, and it is exactly what happens to a great many analytics tables circulating online.

There is a thin line between humility and avoidance, and I have stood on the wrong side of it. In 2026, I delayed publishing a projection about a player's form sequence for three weeks simply because the data was not clean enough. By the time I published, the event had already happened. Readers lost the value of a well-timed forecast. So I set a second rule to complement the first: after presenting the data and stating the margin of error, I must still deliver a judgement. Refusing to rule is not neutrality; it is letting someone else rule in your place, by a lower standard.

Coaches believe in reputation. Data believes in repetition. The 2026 World Cup delivered the verdict. But one thing deserves adding, and few mention it: a metric carries weight only when the sample is large enough and the context similar enough. Two matches do not make a trend. Sometimes one season does not either. If I repeated the Germany analysis using two friendly matches, the conclusion would be meaningless even though the arithmetic would still be correct.

This is where I think most analytics tables are stuck. They build excellent processes for calculating, but almost no process for declining to calculate. A system without a silent mode will always generate a voice, even when it has nothing to say. In a media environment, that voice spreads faster than silence, and by the time people discover it is hollow, it has already been written into the digital memory of an entire community.

What I do not know

I do not know whether an expected-goals model can measure a player's tightening hand on a decisive break point. I do not know whether distance-covered data captures the fear of elimination. I do not know whether the press conference is where data gets distorted or where it gets corrected. My spreadsheet counts double faults, and I admit it does not count the reasons behind them.

What I do know is that every conclusion of mine has an expiry date and a margin. In this analysis, that margin is wider than usual, because the original source records no publication date, names no tournament, names no player, and contains no information points. It is a genuinely empty case, and the most honest way to handle it is to record it as an empty case. People remember results. I remember the conditions that produced them. And when the conditions do not exist, the only thing worth remembering is that they do not exist.

Signals for the next cycle

My next cycle begins with a short checklist. Every data pipeline I build from now on must return two separate states: finding present, and insufficient evidence. If a pipeline cannot distinguish those two, it is not finished. For readers, I suggest a similar habit: when you encounter an analysis that looks too tidy, check whether it contains a section stating its own limits. The presence of that section is a stronger signal of trustworthiness than any accompanying chart.

Data is never in a hurry. People in a hurry are the ones who get it wrong. I shut the machine down that night at almost five in the morning, and the biggest lesson was not in what I found, but in what I did not find and nearly wrote up as though I had.