Empty Data Sources, Full Conclusions: The Dangerous Habit in Sports Analytics

"In the Shanghai derby, I chose the numbers instead of the whole city." I wrote that line in July 2026, after Shanghai Shenhua beat Shanghai SIPG 2-1 in the Chinese top flight. SIPG produced 20 shots and an expected-goals figure of 2.8. Shenhua managed 6 shots and 0.9 xG. My editor asked for a piece about the home side's fighting spirit. I put the spreadsheet in front of him and said that if the match were replayed twenty times, Shenhua would win three of them at most. But the detail I remember most from that night was not on the pitch. In the data file I received, the source field was blank: no provider name, no publication date, no description of the collection method. I asked about it. The answer was: "Just use it, nobody checks."

That episode shaped how I work. Since 2026, every analysis I publish needs at least three independent metrics — xG, PPDA, distance covered — before any judgment, and I attach the raw table so readers can verify it themselves. The rule is not about showing off numbers. It keeps me from saying things I cannot prove. In March 2026 I analysed ten World Cup qualifiers played by the Germany national team and found their average PPDA was 11.3, while the leading pressing sides sit between 8.5 and 9.5. "In March 2026, I wrote a prophecy. All of Germany laughed." On 27 June 2026, Germany lost 0-2 to South Korea and finished bottom of Group F. The article was shared more than 50,000 times in a single night.

I retell that story for a different reason. Behind every correct prophecy there is a condition few people notice: the qualifier file I used came from a provider that stated the publication date, the match range and the definition of PPDA. Without those three fields, 11.3 is just a string of characters. Three years later I paid for ignoring a different layer of data. In the Euro 2026 semi-final I used my model to argue Denmark would beat England: Denmark averaged 118.7 km per match against England's 112.3, and 18 shots per match against England's 11. Denmark lost 1-2 after extra time. I had ignored bench depth, a variable that sits in none of my columns.

A conclusion is only as trustworthy as the source that produced it, and the sports industry publishes a great many conclusions built on empty sources. This becomes obvious when you work across two markets. In Shanghai I cover esports for Chinese readers, where speed is the only standard. A match ends at 22:00; by 22:20 there are ten analytical pieces, each with a spreadsheet that looks thoroughly professional. In April 2026 I checked three of them. All three cited the same secondary source, and that source never stated how many matches were in the sample.

The problem lies elsewhere: a wrong number is less dangerous than a number with no papers. In esports, where official data usually arrives later than newsrooms need it, metrics such as map win rate, pick-ban rate or average game length get recycled through layer after layer of intermediaries. Each copy drops a field: sample size, time window, game version, tournament tier. By the time it reaches the reader, 62% looks like a fact, when it may only be the sum of four matches played on two different patches.

I call this check the data context, and since 2026 I have been obliged to write it into every piece. That year, as leagues returned after the pandemic, I collected 250 Bundesliga matches and found the home win rate fell from 43% to 31%, with goals per match down 0.4. "No crowd, and football changes shape. I found it — and was rejected." My editor wanted a hopeful paragraph about recovery. I refused, and lost my separate contract. But that research taught me a metric only means something alongside the conditions that produced it: empty or full stands, fixture density, weather, and whether players had to cross two countries inside four days.

For Vietnamese football the lesson costs more. The domestic league plays far fewer matches than European competitions, which means every statistical sample is small. A striker with 5 goals in 4 rounds can be described as being in form, but across 4 matches the confidence interval around that number is so wide it is nearly meaningless. Based on my experience tracking matches in the V.League and in national youth competitions, I see the same error repeating: small samples are given the weight of large ones, simply because the spreadsheet looks tidy.

Deeper down there is an incentive that makes the habit hard to break. The betting market needs continuous numbers and it does not wait for sources. When official data is missing, the market generates substitute data, usually on aggregator platforms that publish no methodology. I have seen an odds table cited as reference data in a professional analysis without a single line naming the publisher. In esports the erosion runs faster than in traditional sport, because regulation on data disclosure and competitive integrity trails market demand by several years.

The counter-intuitive angle sits here: a data-driven article with no traceable source can do more damage than one built on what a reporter saw with their own eyes. The eye-witness knows they are subjective, so readers keep a layer of doubt. The sourceless data writer manufactures false certainty, and the doubt disappears. Correlation is not causation — everybody knows the phrase, but it only does work when we know how the sample was drawn.

I also have to speak about my own side. Every prophecy I make carries a probability of being wrong, including the ones that came true. March 2026 did not prove my model right; it showed that within one particular sample the model was not rejected. The stumble at Euro 2026 is clearer evidence of that limit. After the correction piece, I added a closing section to every article called "Where could my assumptions be wrong?" — not as self-defence, but so readers know which part of my reasoning is most fragile.

"The spreadsheet is an altar, and I offer myself to every number." But an altar is only sacred when people know where the number came from. This big-tournament season, as tables multiply faster than ever, try one small habit: before trusting a percentage, ask how many matches the sample holds, over what period, and who published it. "Every crowd is wrong. The only thing that is not wrong is probability." Yet probability holds only while the source stays intact. When the source field is blank, the right move is not to keep writing, but to send the file back and ask for it to be pulled again.