A 'Football' Record With No Players: The Data-Routing Error Vietnamese Analytics Rooms Need to Face

Last September, during a routine review of a content data stream, a record surfaced carrying the topic label “football”. Inside was an article about Angela Stribling, a familiar face on BET and a Washington D.C.-area radio host, who died at 58. Twenty-two information points had been extracted. The entity list read: Ed Gordon, WJZ-TV, WJLA-TV, Sirius, the programme “Pillow Talk with Angela”, Bill Clinton, Stevie Wonder, Quincy Jones, Janet Jackson, 50 Cent, Brandy and Sterling K. Brown.

Not one club. Not one player. Not one league, contract, coach or governing body. The record still passed through the first processing layer and settled inside the football data stream as a valid entry.

I am not telling this story to catch out a machine. I am telling it because it lands squarely on what analysis rooms in the V-League build every day, and because the fault sits exactly where few people bother to look.

When domestic football imports an entire pipeline

Over the past seven years, the analysis trade at Vietnamese clubs has changed shape entirely. Where one person once sat through video replays taking notes by hand, every match now drags thousands of event-data rows behind it. A single V-League fixture can generate data on ball position, running lines, the penalty area, shot frequency. To feed those models, clubs had to build an extra link few people notice: a content-processing system that reads articles, extracts entities and assigns topic labels.

That link was largely not designed for Vietnamese football. It was inherited from international data, where football is a large, famous topic and therefore attracts anything vaguely shaped like it.

Three keywords in that record explain the mechanism. “Network” means a broadcasting network, but in football it also means a club network or an academy network. “Campaign” means an advertising campaign, but in football it also means a season campaign, a title push. “National” means national reach in media, but in football it means the national team. A keyword-driven classifier reads those three signals, adds the word “football” from somewhere in the text, and assigns the label without ever needing to know whether anyone is running on a pitch.

At many domestic clubs, the workflow still rests on a few people doing several jobs at once: one analyst watches video, enters data and reads the news. None of them is assigned cross-checking as a sole duty. When every stage is urgent, the stage that gets cut is verification — and that is the one stage capable of catching a stray record.

That is the cheap error, and it is easy to fix. The expensive one sits behind it.

Bad data does not crash a system — it drags the conclusions sideways

A record filed under the wrong topic does not stay put. It travels onward into the entity tables.

The mechanism runs like this. After labelling comes entity resolution, which links names into a relationship graph: which player belongs to which club, which club belongs to which league, which broadcaster holds which rights, which owner holds which stake. That graph underpins almost every later query.

When BET, WJZ-TV, WJLA-TV and Sirius land in the entity table of a football stream, they become nodes shaped exactly like rights-holding broadcasters. From there, a simple query such as “which media outlets are attached to domestic football” can return an American television network that has never shown a minute of V-League football. Nobody checks. Nobody spots it. Not until a commercial dossier or a rights report is built on that very data.

Based on my experience following matches, I have met a smaller version of the same failure. In the summer of 2026, I and the numbers went down to the bottom of the V-League to build a statistical table for the 378 goals of the 2026 season. A two-character data-entry error attached a match to Thống Nhất Stadium that did not belong there. One row. But that row skewed the whole distribution of goals by flank that I used for the analysis. I had to re-check six months of notes to find it. Bad data does not bring a system down. It quietly drags every conclusion a few degrees off, and nobody knows.

The Stribling record is more serious because it fails at topic level, not cell level. A wrong cell is fixed by fixing a cell. A wrong topic label means fixing the pipeline.

There is one more layer, and it matters most. The record is an obituary. Its sources were a colleague’s memorial post on Facebook plus the deceased’s own LinkedIn profile. The date and cause of death were both left unstated. For a claim that sensitive, that is a source at the very bottom of the credibility table. And it went straight into the processing stream with no gate to stop it.

Put plainly: our pipeline does check whether content matches its topic, and it got that wrong. The pipeline does not check whether content has any business existing at all. Two failures stacked in one record.

To see how far the mislabelling runs, redraw the article’s real transmission chain: an American broadcast career, moving through BET and local stations, extending into the Washington D.C. area and Sirius, then into voice work for national advertising and public-awareness campaigns. That is a complete media chain, valuable to media studies. It has no joint connecting it to the football chain.

For organisations building prediction models for the V-League, the damage does not stop at one bad query. Every model trained on historical data inherits the error rate of that data. A foreign entity appearing often enough becomes a feature in the model. And once it is a feature, nobody remembers it was ever a mistake.

The fault is not in the machine

The first reaction most people have on hearing this is to blame the algorithm. I do not think so, and that reflex hides the real problem.

If an analytics pipeline is willing to accept a record that nobody in the chain has ever read, the problem lives in the process, not in the model. The algorithm does exactly what it was taught: find keyword signals and label as fast as possible. The human job is to place a gate where nonsense can be caught. An article about an American radio host sitting inside a football stream is nonsense obvious to anyone who skims it.

A gate in the right place does not need to be complicated. One cross-check is enough: if a record carries a football label but contains no known club, player, league or governing body, send it to a manual review queue. The cost of that step is measured in seconds. The cost of skipping it is measured in months of data repair.

I have a harder observation. The pressure to “have more data” is now greater than the pressure to ask whether the data is right. Analytics rooms race to widen their collection sources, while the number of people actually checking source quality barely moves. The result is a system that grows thicker in data and thinner in verification.

Before trusting my own eyes, I choose to trust structure. But structure also has to be checked by human eyes, at exactly one chokepoint. Remove that chokepoint and you do not get a smarter system. You get a system that is faster at being wrong.

Nor should we miss a side lesson from the article itself. The “pioneering” description attached to Stribling is the author’s opinion, explicitly marked as opinion, and backed by no quantitative measure. In data, adjectives like that are traps. It is easy to file a flattering remark into the facts column and then build three more layers of analysis on top of it. Data never shouts, but it whispers loudly enough for anyone willing to listen — and an adjective is not a whisper, it is an echo of the writer.

Three things to do now

The first is to count again. If a sample audit turns up two or more records carrying a football label with no football entity inside, it is no longer an isolated error. It is a system error, and it has to be handled at the classification layer.

The second is to audit the entity tables. Check whether BET, Sirius, WJZ-TV or WJLA-TV exist in any football relationship table. If they do, they must be removed and dependent queries recomputed.

The third is to log the trigger keywords on every miscategorised record. The first three candidates to review are “network”, “campaign” and “national”.

And one task that is not technical: keep this record as a test fixture. One thoroughly documented failure is worth more than ten unexamined successes. Next season, when the pipeline is upgraded, this will be the first test I run.

The paper newspaper closed, but the tactical map began to open. That map is only worth trusting if we stop at one single place: the place where we ask ourselves whether we are reading the thing we actually meant to read.