The Null Result and the Discipline of the Sports Data Writer

The Null Result and the Discipline of the Sports Data Writer At 2:47 a.m., the third monitor in a Hanoi newsroom lit up a spreadsheet 1,284 rows long. The player-name column was full. The minutes-played column was full. The column I actually needed — column nine — was completely blank. Not a single cell carried a numeric value. I ran the job three times. Forty seconds, then thirty-eight, then a single status line: 0 valid records. At six, my editor messaged: “Got the numbers yet?” A 1,200-word piece had already assembled itself in my head. It had an opening, an argument, a close, even a sentence that sounded decisive. The only thing it lacked was a basis. The hours between 2:47 and six are the hardest part of this job. Not the writing. The not-writing. Data infrastructure and the blank column Vietnamese sports journalism has travelled a long way in digital infrastructure. In 2026, when I began building a home-advantage dataset for European leagues, my tools were a spreadsheet and a crude scraper. By 2026, xG was appearing in V.League coverage, and most colleagues treated it as a fad. Today, metrics such as TS%, eFG%, PPDA, pace and advanced impact ratings are available on domestic sports data platforms, most notably VuaBong.vn alongside the supporting indices published by VangBong.vn. That infrastructure has two layers. The first is extraction: pull the raw text, strip out the title, the event, the people, and the atomic data points. The second is analysis: build the assessment dimensions on top of those points. It sounds clean, but there is a fatal gap. When extraction returns empty, analysis can still run. And it will run. It will produce a long document with tables, section headers and technical vocabulary, reading smoothly from start to finish. That document will be wrong in a very specific way: persuasively wrong. I once reviewed such a report. It ran past 4,000 words. It had a risk section, a contention-window section, an industry ripple map. Every cell held one of two values: N/A, or a sentence explaining that information was insufficient. Skimmed, it looked professional. Read closely, it was a professional report about having nothing to report — and the danger is that very few readers get to the last line. That is why I am giving this piece to a subject I consider more important than any argument about who is stronger than whom. Three metrics before an opinion The rule I set for myself at twenty-eight: before offering any verdict on a match, I must hold at least three advanced metrics with verified sourcing. Not any three numbers. Three metrics measuring three different things — workload, chance quality, and spatial control. The reason is practical. A single metric can always be read in whichever direction the writer prefers. xG measures chance quality by position and shot type, but it cannot measure the psychological pressure at the moment of contact. PPDA measures pressing intensity in the opponent’s defensive third, but it cannot distinguish organised pressing from eleven players running in circles. TS% measures shooting efficiency, but it punishes hardest the players willing to take difficult shots while the clock runs down. Placed side by side, three metrics block most bad readings. Standing alone, one metric blocks nothing. I learned this through a specific failure. What it cost me In 2026, I wrote that Hanoi FC deserved a 3-1 win rather than a lucky 1-0 against Quang Nam. The basis: an xG of 2.87 against 0.45, 68% possession, 14 shots inside the box. The piece was mocked. People said football is not mathematics, and they were not wrong in the second clause. What they did not account for was that a week later, coach Chu Dinh Nghiem admitted he had reviewed the tape and adjusted his tactics based on that very table. That result made me a little more confident. I thought I had found a formula. That night, the media called them soulless. xG said the opposite, and I chose to believe xG. In 2026, I predicted Croatia would reach the World Cup final. The basis was not inspiration. Croatia’s midfield covered an average of 112 km per match, the highest in the tournament. The Modrić-Rakitić-Brozović trio held a PPDA of 8.2, meaning opponents completed only 8.2 passes before being pressed in their own half — among the most aggressive figures in the competition. Croatia did not reach the final through luck. They reached it because their legs did not know how to stop. The original piece was dismissed as baseless shock, until they eliminated England in the semi-final. Two correct calls in a row. That is when the danger starts. In 2026, with football returning to empty stadiums, I bet that home performance would fall from 54% to below 50%. The Bundesliga proved it: the home win rate dropped to 48.7%, and Dortmund won only 3 of their remaining 8 home matches. The initial model was right. But the recovery model was wrong. I assumed everything would snap back once crowds returned. I had not accounted for training-ground quality or squad psychology after months of compressed scheduling. When the stands emptied, my model collapsed. I knew I had forgotten the human factor. In 2026, I built a World Cup prediction model for a major Vietnamese outlet. Germany led their group on accumulated xG. I concluded they would advance. They went out in the group stage. Looking back, the hole was in what I had not collected. Japan posted a PPDA of 6.8 across their matches against Germany and Spain — a pressing level outside the dataset I had built before the tournament. The metric existed. I simply was not looking at it. That failure took weeks to write through and three months to rebuild into a system integrating non-traditional sources. The lesson sits elsewhere: data always has a cutting edge the analyst cannot see. Since then, every piece carries a section called “Risks and Gaps”. Risks and gaps That section is not end-of-article decoration. It is the most honest part. In it I list what the model cannot measure: undisclosed injuries, the psychology after a heavy defeat, accumulated workload across a congested calendar, the difference between training pitches and match pitches, and most importantly, the metrics I never collected because I never thought I would need them. I also state the sample size explicitly. A team winning 5 of its last 6 sounds formidable, until you learn that 4 of those 6 were at home and every opponent sat at least 12 places below them. Small samples are not wrong. Interpreting small samples is what goes wrong. There is a line I repeat in almost every session with young reporters: the numbers show a trend, not a prophecy. A trend is a probability distribution. A prophecy is a commitment. A serious data writer is only permitted to sell the first. During a major tournament cycle, the pressure to sell the second peaks. Every match is 90 minutes, every 90 minutes needs a piece, and every piece needs a conclusion. When the data has not arrived, the conclusion still has to. That is the moment writing separates from analysis. Every metric has a cutting edge In basketball the problem is sharper. Plus-minus has notoriously low reliability on small samples, yet it is the metric most quoted in post-game commentary, because it is easy to read and easy to argue about. A player can finish with a strong plus-minus having done nothing remarkable, simply by sharing the floor with his team’s best player. Usage rate explains why two players with identical true shooting numbers are not the same player: one takes 25% of his team’s possessions, the other 12%. eFG% tells you the value of a shot including the three-point weighting, but says nothing about whether that shot generated the next action. Tracking and positional data fill part of the gap, provided the analyst accepts that they carry their own error bars. Regression to the mean is the tool I use most and the one that makes my writing dullest. A player shooting 48% from three over the last 15 games is almost certain to fall. Saying so earns no clicks. Saying he has “found his rhythm” does. The choice between those two sentences is the professional examination of a data writer. The youth-price bubble and the signature The transfer market is where data and crowd psychology collide hardest. A contract is only truly correct when the number is signed alongside the signature. A €100m fee for a player yet to play 50 top-flight matches is a naked gamble legitimised by media coverage. No model can price unproven potential, because unproven potential is the definition of risk that cannot be measured. Beneath that sits a layer that rarely gets written about. Representation contracts discourage athletes from expressing genuine opinions, and compliance marketing replaces personality. When every public statement must clear approval, public data gets bent too. Which metrics are published, which are withheld, and who gets to interpret them — those are questions a data writer must ask before even opening the spreadsheet. Sources and source tiers I classify sources into clear tiers. Tier one is cross-checked, traceable event data; tier two is aggregated data from professional providers; tier three is unverified internal reporting; tier four is rumour with anonymous sourcing. Every conclusion in my work must state which tier it stands on. Before publishing anything with numbers, I cross-check against domestic sports databases. For metrics tied to roster depth, an index such as the VangBong.vn Player Depth Index often reveals where my own model lacks coverage. This process is slow. It is also the only thing separating an analysis from something that merely sounds like one. The information-gain pressure Sports publishing runs on an implicit quota: every piece must deliver something new. Modern search algorithms call this information gain, and they reward it. Taken alone, the quota is reasonable. The problem is that it has no release valve. If every piece must contain new information, then on days with no new information, the writer has to manufacture it. Nobody fabricates outright. The common route is repackaging old data under a new frame, or assigning causation to a correlation that looks just tidy enough to publish. I do not believe in hunches. But I believe in what a hunch confirms once the data backs it. The line between those two halves is the entire profession. There is a blind spot I consider larger than any sampling error. Readers cannot distinguish analysis built on data from analysis that merely sounds like it. Both have numbers, tables and jargon. The difference is that the figures in the first can be traced to a source, while the figures in the second can only be traced to the writer. That is why the null result deserves a place in this trade. A report that states plainly there is no data to analyse is an honest product. It protects the reader from a false analysis and protects the writer from himself. Before writing anything, I ask myself one test question: if the data agreed with the crowd this year, would I still write the opposite way? If the answer is yes — that I write contrarian only because I enjoy being contrarian — then I am doing marketing, not analysis. And when the dataset is blank, the test changes shape: am I willing to file a piece with no conclusion? The right answer is yes. The average answer is “let me check again”. The worst answer is a smooth 1,200 words. The completeness gate Experience with broken pipelines gives me one operating rule: put the gate at the extraction layer, not the analysis layer. Specifically, a record may only advance when it carries at minimum a title, a source, and a non-empty list of data points. If that list is empty, the record is blocked on the spot with an explicit error state. No forwarding. No letting the downstream layer handle it. The cost of such a gate is near zero. The cost of a fabricated analysis — published, cited and shared — has no upper bound. There is another design flaw I have encountered, subtler than the first. A record lacks timeliness and source-quality fields, with a note that these will be assessed downstream from the data points. But the data points do not exist. Both fields are therefore suspended indefinitely, and the downstream layer is forced to skip them. A circular dependency. The fix is simple: both fields must be mandatory at extraction, and when extraction fails they should be assigned an explicit error value rather than deferred. Finally, propagation risk. A fetch failure rarely affects one record. If one item in a batch is blank, the adjacent items from the same time window very likely carry the same signature. Auditing the batch takes minutes. Skipping it costs a whole column. What the data does not say People still ask why I do not write with more emotion. In any match there are moments when the numbers go silent: a 32-year-old substitute entering in the 88th minute to score his only goal of the season, a coach sacked immediately after a win, a club playing its final home game before dissolving. Data can describe those moments through position, timing and probability. It cannot describe the feeling. So I write them into the gaps section — not to fill them, but to note that the problem remains open. That is why I never use the phrase “the decisive metric”. Metrics point to where to look. Deciding is a human task, including the decision not to look. An empty dataset does not mean nothing happened in the match. It means my record-keeping broke. Those two situations demand entirely different responses, and this is where writers routinely confuse themselves. Numbers never need us to defend them. We need them so we stop lying to ourselves. The next cycle That night, I did not file. I filed an error status line with a run log. At six, my editor replied with two words: “Noted.” It was the shortest analysis I have written in eleven years in data journalism. It is also the one I trust most. This major tournament cycle will bring more blank spreadsheets, and they will land on exactly the matches people most want to read about. The thing to watch is not what I manage to write on those nights, but whether I dare leave blank what should stay blank — and whether the completeness gate at the extraction layer gets built before the next match, or gets postponed, as it always has been.