A dog-and-cat bulletin inside the football data lake: when caseload gets read as prevalence
**Câu trả lời cốt lõi** Consejo Ciudadano para la Seguridad y Justicia de la Ciudad de México đã phát khuyến cáo bảo vệ động vật nuôi trong dịp Día de Muertos, nhưng nguồn này bị dán nhãn bóng đá trong dây chuyền phân tích. Mười bốn điểm thông tin không chứa thực thể bóng đá nào, nên toàn bộ các chiều phân tích chiến thuật không áp dụng được. **Dữ kiện chính** - 59.588 báo cáo ngược đãi động vật được ghi nhận tại Thành phố Mexico từ năm 2019. - 6.616 ca trong tháng 1 đến tháng 6 năm 2026, tăng 28% so với cùng kỳ năm trước. - 203 ca gắn với nghi lễ kể từ năm 2023, không có bảng phân tách theo màu lông. - Mười hai trong mười bốn điểm thông tin truy về một tổ chức duy nhất, không có kiểm chứng độc lập. - Nhãn bóng đá sai khiến chín chiều phân tích chiến thuật, tài chính và chuyển nhượng không thể áp dụng. **Nguồn** Consejo Ciudadano para la Seguridad y Justicia de la Ciudad de México, khuyến cáo dịp Día de Muertos, chương trình CompAnimal. Ngày công bố cần xác minh vì văn bản dẫn số liệu tháng 1 đến tháng 6 năm 2026. **Hỏi đáp liên quan** Hỏi: Vì sao 6.616 ca không phải là tỷ lệ ngược đãi động vật? Đáp: Vì đó là số hồ sơ một tổ chức đã xử lý, phụ thuộc vào mức độ quảng bá kênh báo cáo, không phải tỷ lệ hiện hành trong quần thể. Hỏi: Nhãn sai này ảnh hưởng gì tới dữ liệu bóng đá? Đáp: Nó làm nhiễm toàn bộ các tầng phân tích phía sau, giống lỗi so sánh PPDA giữa hai nhà cung cấp có định nghĩa khác nhau. Hỏi: Chỉ số nào đáng theo dõi sau sự cố này? Đáp: Độ trễ phát hiện, tức số tầng đi qua trước khi có người đọc lại nhãn ở dòng đầu tiên.
2:47 a.m., Seoul. I open a stage-two analysis file and read the first line: domain — football.
Below it are fourteen information points. No club. No player. No match. No table, no contract, no coach, not a single name belonging to the world I make my living from.
Instead there is a public-safety advisory issued by the Consejo Ciudadano para la Seguridad y Justicia de la Ciudad de México, warning of risks to companion animals during Día de Muertos. Black-coated and white-coated dogs and cats are listed as the most vulnerable. Goats, roosters and crows follow.
I read it a second time. Then a third. I try to bend each point into a football story: a club with a dog for a mascot, a league staging an event before the holiday, a player suspended for animal cruelty. Not one point survives the bend.
Tables speak, if anyone has the patience to listen. That night the table said the one thing an entire multi-stage pipeline did not want to hear: the label is in the wrong place, and nobody in the chain stopped long enough to check.
The file traces back to an annual bulletin. The Consejo Ciudadano is Mexico City's citizens' council for security and justice, and it runs CompAnimal, the channel through which animal-abuse reports are received and logged. Día de Muertos, from late October into early November, is when the council repeats its advisory.
The figures it publishes: 59,588 reports since 2026; 6,616 cases between January and June 2026, up 28 percent year on year; 203 cases tied to ceremonies since 2026. The behavioural recommendations include keeping animals indoors, securing entry and exit points, and temporarily suspending adoptions of black-coated animals during the peak period. The council also notes that animal abuse may constitute a criminal offence and urges residents to use the reporting line.
That is the entire content. Nothing else. And that is exactly why the file deserves a piece, even though it contains not a single minute of football.
Because its structure — an institution issuing a calendar-driven advisory, attached to a block of self-reported figures — is the structure our industry produces every day, with only the letterhead changed.
One more detail, and this is the decisive one: twelve of the fourteen information points trace back to a single organisation. No second agency, no academic verification, no independent non-governmental confirmation. The whole block stands on one leg.
I have covered football for fourteen years, and I have learned one thing about data blocks standing on one leg: they do not collapse immediately. They collapse the moment somebody leans on them.
Six thousand six hundred and sixteen is not the number of abuse cases.
It is the number of files one organisation processed. Epidemiology calls the distance between those two things the distance between caseload and prevalence. A hospital that opens a second intake desk will record more cases without a single additional sick person. A reporting line that is better publicised will take more calls, and the number on the report will rise, and nobody will have lied.
The 28 percent increase may be a measure of CompAnimal's communications success. It may equally be a measure of a genuine rise in abuse. The text does not distinguish between the two, and the way it is presented — a rising number placed beside an advisory — implies the second.
I made exactly this error in 2026. I was in my final year of a statistics degree, re-watching every passing data point from a nineteen-year-old midfielder named Kim Min-jae across fourteen K League Classic matches, and I found a chance-creation passing rate of 6.8 percent, below the league average. I wrote a piece whose headline amounted to saying his transfer hype was inflated. Three hundred comments calling me an idiot. Twenty comments engaging seriously.
The 6.8 percent was true. My conclusion was wrong. I had not adjusted for role, for team structure, for his average receiving position on the pitch. A rate without its denominator is just a rumour with a decimal point.
That is precisely what is happening with 6,616.
The division nobody performed
6,616 cases in the first half of 2026, up 28 percent year on year. Run the division backwards: roughly 5,169 cases in the first half of 2026. The two numbers are arithmetically consistent, and that is the entire extent of what we know about them. There is no independent dataset to cross-check against, and no population denominator is given, so no per-hundred-thousand rate can be derived — the only figure that could be compared with another city or another year.
The 203 ceremony-related cases since 2026 are the most interesting set in the document and the most thinly exploited. There is no breakdown by coat colour. Which means the claim that black-coated and white-coated animals are most vulnerable stands in front of those 203 cases with no support from the organisation's own data.
The mechanism is plausible: in ritual contexts certain coat colours carry symbolic associations, and symbolic associations can drive real behaviour. But plausible is not measured. I have seen too many pressing models built on a plausible assumption collapse inside three matchweeks.
And there is one datum that does not reconcile: an article about Día de Muertos, an event in late October, cites figures from January to June 2026. Either the piece was published in the second half of 2026, or the year is a typo. The text does not resolve it. A public advisory carrying an unverified date error is an advisory I cannot yet cite.
The self-confirming loop
One line in the recommendations matters more than the rest: temporarily suspend adoptions of black-coated animals during the peak period.
That is an institutional opinion, not a legal rule, and its proportionality is contested within animal-welfare practice itself. What stopped me was its measurement consequence.
Suspending adoptions of black-coated animals means black-coated animals stay in shelters longer. Staying longer means the length-of-stay figure for that group rises inside the organisation's own data. And when someone opens the table again three years later, they will find that black-coated animals are harder to rehome — and will call that evidence for the original hypothesis, when what they are reading is the footprint of the intervention.
Measurement creates the phenomenon, and the phenomenon comes back to confirm the measurement. This is the hardest error in data science to detect, because it does not produce wrong numbers. It produces correct numbers for a question that was distorted beforehand.
Football runs on this loop.
The same error in a different shirt
I have seen it in the transfer layer. A player appears on forty shortlists at forty clubs, as aggregator sites count them. Those forty shortlists become an indicator of interest. A real club reads the indicator, picks up the phone, and the rumour becomes fact. The transfer market does not sell players, it sells the faith of supporters — and it manufactures that faith by counting what it has just inflated.
I have seen it in the metrics layer. PPDA, passes allowed per defensive action, is defined differently by different data providers: some count only inside the defensive third, some across the whole pitch. Placing two PPDA figures from two providers side by side is doing exactly what the Mexico City advisory does: comparing two caseloads collected under two different processes and calling the result a trend.
I have seen it in the modelling layer. The same shot, two expected-goals models, two numbers. Neither model is technically wrong; they answer slightly different questions and are presented as the same one.
What saved me, by contrast, was process data. At the 2026 World Cup, before South Korea met Germany, the Korean press talked only of a draw or a narrow defeat. I went back through Germany's pressing numbers across two group games and found opponents had taken 245 touches in dangerous areas, roughly 40 percent above the qualifying campaign. I wrote a piece arguing Germany would collapse on a broken press. A thousand people laughed. The score was 2-0 to South Korea. They called me a statistical fraud because they could not call me wrong.
In 2026, with stadiums empty during the pandemic, I had more than 130 K League and Bundesliga matches without spectators. The home win rate fell from 46 percent to 34 percent, and goals per game rose to 3.1. I published it, was accused of inventing the numbers, and released the raw dataset for 48 hours so anyone who wanted could check. When the stadium is empty, the truth begins to fill the space the crowd left behind.
The difference between those two cases and 6,616 is this: Germany's pressing data was process data, measuring what was happening; 6,616 is caseload data, measuring what was reported. Tables speak, but only when someone reads the right column.
And here is where I may be wrong.
Possibility one: the label was right and the extraction was wrong. If the origin was a multi-topic roundup and the pipeline grabbed the wrong paragraph, the error sits in extraction rather than classification. Those two errors need two different fixes, and I have concluded on the second without excluding the first.
Possibility two: 2026 is not a typo. It may be a forward-dated advisory drawing on an unfinished cycle. In that case the anomaly is mine, not theirs.
Possibility three, and the one I dislike most: the organisation may hold a genuine coat-colour breakdown, and it simply was not among the fourteen points I was given. Absence of evidence in a summary is not absence of evidence in an institution.
And there is a professional bias I have to name: I live in Seoul, I earn my money keeping football data clean. A mislabelled case flatters my own profession. Turning an animal-safety advisory into a parable about my own trade is precisely the move I accuse others of making. The crowd is always safe, and that is exactly why the crowd is always mediocre — but the man standing outside the crowd can be wrong in his own private way.
My prediction, and it is falsifiable: within the next two transfer windows, at least one major outlet will publish a story built on aggregated rumour data and issue a correction after the piece has already spread. The metric worth tracking is not the error rate but detection latency — how many stages pass before someone reads the label on the first line again.
Every number I dig up buries a myth the media created. That night, the myth buried was a very convenient belief: that if the pipeline is long enough, it will filter itself.
I do not need anyone to agree with me. I need someone good enough to argue back.
