Tracking Vietnamese Badminton With Data: Where My Model Fell Out of Step With the Court

Tracking Vietnamese Badminton With Data: Where My Model Fell Out of Step With the Court

On the electronic scoreboard, the winner of the men's singles semifinal at the Vietnam International Challenge had nine fewer winners than his opponent. His win rate on rallies longer than fifteen shots was 38 percent. His win rate on rallies shorter than six shots was 71 percent. I sat in row seven, logging every rally in a notebook, and the result ran against almost every assumption I had carried into the tournament: the man who held the long rallies, the man who made fewer unforced errors, the man who stayed calm in the closing points — all of that belonged to the player who lost.

Four days later I still had not finished answering the only question worth asking. Had I recorded it wrong, or had he won with something my spreadsheet had no column for?

This article is the product of those four days. It does not arrive at a clean conclusion. It arrives at a list of places where my model fell out of step with the court — which I consider the most valuable part of three years spent tracking Vietnamese badminton.

Method and data scope

I collect Vietnamese badminton data in three layers. The first is public results from the World Federation tournament system: set scores, match duration, opponents, rounds. The second is video of streamed matches, used to count rallies by hand. The third is a notebook kept at the venue, where I classify every point into four groups: winners, opponent unforced errors, forced errors, and points from service or rule faults.

Over three years I logged 412 matches. That sounds like a lot, but only 168 had complete rally data usable in a model. The rest lacked video, lacked enough camera angles, or simply happened when I was not there. A sample of 168 matches can describe group trends. It cannot describe a single individual, and anyone using it to make claims about one person is reading beyond the limits of the data.

I state that before the analysis because the worst habit in sports data work is presenting a small-sample figure as a law. Russia 2026 was not an anomaly; it was a reminder about small samples. In 2026 I wrote that a team with 87 percent possession would win, based on one tournament's statistic table. Three weeks later I rewatched ten matches, counted every pass, and understood that possession is only surface paint. That lesson crossed over to badminton intact.

Long rallies say nothing about the winner

In badminton analysis there is an almost default belief: whoever controls the long rallies controls the match. It sounds reasonable. It is also very hard to verify, because most badminton statistics platforms provide only scores and duration, not rally distribution.

I counted rally distribution for those 168 matches. The first result: the correlation between long-rally win rate and match win rate in events below Super 100 level is very weak. The coefficient I calculated sits around 0.2, essentially without predictive value. At higher levels the correlation is stronger but still insufficient as a standalone metric.

The reason is that a long rally can arise from two opposite mechanisms. First: a player actively manoeuvres, pulls the opponent out of position, accepts more shots to open an angle. Second: neither player dares to finish, and the shuttle crosses the net simply because nobody dares to come forward. The same rally-distribution number describes two completely different tactical stories.

When I separated the two groups by adding one small column — whether a long rally ended in an active winner or in an unforced error — the correlation changed markedly. Long rallies ending in active winners correlated positively with match wins. Long rallies ending in unforced errors correlated negatively. Combined, the two groups cancelled each other out and produced a nearly meaningless metric.

Every number has a genealogy; I need to know its ancestors. A metric has value only when I know which process produced it, who recorded it, under what conditions, and what it left out.

Points from opponent errors are loans, not income

At lower international levels, the share of points coming from opponent unforced errors is very high. Across the 168 matches, that share ranged from 34 to 52 percent depending on the match, averaging around 43 percent. Nearly half of a player's points did not come from that player's racket.

This is where my prediction model failed badly. I built a metric I called conversion rate: winners divided by total points won. Players with a low conversion rate but many wins were rated low by my model. In fact they won because opponents at that level made many errors. When they moved up a level, where opponents are steadier, the free points disappeared and results collapsed quickly.

I call points from opponent errors loans, not income. They can arrive steadily for a stretch, but they depend on someone else, and they vanish exactly when you need them most. When I re-ranked the Vietnamese group by pure winners instead of win rate, the order shifted by seven places out of twenty players. Seven places is a large enough gap to change how a national squad allocates tournament entries.

The Russia World Cup shock taught me this: skewed data is more dangerous than intuition. When intuition is wrong, people still doubt it. When a spreadsheet is wrong, people carry it into a meeting and make decisions.

How to read a badminton scoreline

A public badminton scoreline contains very little information about how a match unfolded. It says who won which set, by what score, in how long. It does not say where the points came from.

Two matches ending 21-19, 21-19 can be entirely different. Match one: two players trade attacks, rallies run long, and the winner takes points with finishes on the twelfth shot. Match two: the two players commit forty unforced errors between them, and the winner is simply the one who erred less across the final three points. Identical scoreline. Completely different story. A tournament entry awarded on that scoreline could be awarded wrongly.

That is why I always add three columns when I am at the venue: winners, unforced errors, and rally distribution across short, medium and long categories. Those three columns turn a meaningless scoreline into a reasonably honest description of a match. They do not make me right. They just make it harder for me to fool myself.

The women's singles case: where data is thinnest

Nguyen Thuy Linh has been the most closely followed Vietnamese player for years, and she has the fullest public dataset. But full in the sense of results, not process.

For a player competing mainly on the Asian circuit and a handful of Super 300 events, the number of matches per year is small. Each match sits in a different tournament, under different conditions, against a different opponent. When I tried to compute a stable metric for her, I realised I was splitting a small sample into smaller samples. After three splits, each cell held a few matches, and every conclusion drawn was a conclusion drawn from noise.

I once published an analysis of her third-set win rate late in a season, concluding that fitness was a weakness. Three months later I had to correct it. The problem was that I had pooled matches with very different rest gaps into one group. Most of the third-set losses I used as evidence happened in consecutive tournament weeks, when the schedule allowed no recovery. The third-set wins came in weeks when she played a single match. After splitting by days between matches, the gap almost disappeared.

That is the most basic error in sports analysis: pooling things that should not be pooled. A season on paper only looks beautiful while the model has not met reality.

For Le Duc Phat, the dataset is even thinner. In men's singles, opponent density in the region is higher, so a Vietnamese player often meets a higher-ranked opponent in the opening round. That creates an effect that is especially hard to handle: his metrics are driven more by opponent quality than by his own quality. Analysing such a player without normalising for opponents is analysing the draw, not the person.

In both Vietnamese singles draws, the domestic tournament structure creates another paradox. The national championship is where players compete most, and also where data is least recorded. No complete video, no detailed statistical sheet, no rally-counting system. Most of what I know about young Vietnamese players comes from forty-seven matches I attended in person, and that is a sample far too small to conclude anything about anyone.

A season on paper and ranking points to defend

The world federation ranking system counts a player's best results within a 52-week cycle. Current ranking is therefore a photograph of the past, not a forecast of the future. When points from a major event are about to expire, a player's ranking drops even if they lose no further matches.

I once built a ranking projection for several Vietnamese players based on current points and recent form. The model failed exactly here. It had no variable for points to defend. It also had no schedule variable: in Vietnam, players tend to concentrate competition into a few Asian events and domestic tournaments, clustering points into a few weeks of the year. The accumulated error produced a projection that was off by six places.

I published a correction after that mistake, openly, with an error table. I kept the old article up rather than deleting it, because deleting old work is the fastest way to convince yourself you were never wrong.

There is a second data layer Vietnamese media rarely looks at, and it matters as much as ranking points: money. Vietnamese badminton runs largely through provincial and municipal team structures. Players compete for their managing unit at the national championship, and units have very different budgets. Some players change units between seasons, which brings changes in coaching staff, training schedules, and sometimes the ability to be entered for international events.

Tracking money and contracts in Vietnamese badminton is far harder than in football, because most information is not public. But it still leaves traces: entry lists for the national championship, coaching staff composition, the number of international events each player is registered for in a year. I read those three tables together, and they tell a clearer story than any press release about who a unit is investing in.

During transfer and restructuring periods between seasons, the loudest information is usually the least valuable. A rumour about a player changing units can spread fast, while the official entry list stays silent. I learned to wait for the entry list. It is slower, but it does not lie.

Applause at home

The most common assumption I ever carried was that playing at home is an advantage. My data does not support that at international level below Super 300.

I split matches featuring Vietnamese players at home and away, then compared third-set win rates. Third-set win rate at home was roughly seven percentage points lower than away in my sample. I stress that this is a small gap and a small sample, so it is not enough to assert. But it was enough to break my confidence in the original assumption.

At least three mechanisms could explain it. First, physical conditions: high humidity changes shuttle speed, and an indoor venue in Vietnam may not fully match the conditions a player prepared for in training. Second, psychology: at home, a player may play to avoid losing points in front of a home crowd rather than to win them. Third, scheduling: domestic events often come with media duties, opening ceremonies and photo sessions, reducing recovery time.

I cannot separate these three mechanisms with the data I have. Admitting that matters more than choosing one plausible mechanism and writing as if it were proven.

Coaching staff and what data never records

There is one column I have never managed to record, even though it affects almost every match: the quality of the training block before it.

A player entering a tournament after three weeks of specific preparation for one opponent will show a completely different rally distribution from a player entering after two weeks off with an injury. The scoreline does not distinguish between those cases. And in Vietnamese badminton, information about training loads is almost never public.

What I can observe, based on my own experience attending matches at venues in Vietnam, is the difference in how coaching teams respond between sets. Teams with proper opponent analysis tend to shift service direction markedly in the second set, after having one set to read the opponent. Teams without it tend to keep the same service direction to the end. I log that shift, and it is one of the few indicators that reveals preparation behind the court.

I cannot quantify it into a precise number. I only know it exists, and I write it into the assumptions section of every report I produce.

What my model has no column for

Good analysis means asking the right question, not holding a beautiful answer. The right question I took from those four days was this: does my spreadsheet measure behaviour on court, or does it measure the quality of the opponents that behaviour met?

Two players with identical conversion rates may be competing against very different opponent pools. Without normalising for opponent quality, my metric is blending two quantities into one column. That is the most common error in amateur sports analysis, and I made it for almost two years.

Some variables public data never contains. Injury is the largest. A player stepping on court with an unhealed ankle injury will show an entirely different rally distribution, yet on the scoreline they remain last month's version of themselves. The draw is the second: at equal levels, an easy bracket can carry a player to a semifinal without beating anyone in the top eight. Playing conditions are the third: one shuttle box selected at a different speed, and an entire service plan loses its value.

Injury, the draw, a shuttle half a gram faster — variables with no rows and no columns. Not because they are small, but because they are hard to measure. And what is hard to measure tends to be dropped from the report, then dropped from memory.

The instant review system in badminton is another example. It was introduced to reduce disputes over in-or-out calls. My data on reviewed points shows something different: the total volume of dispute did not fall. It simply moved from around the umpire to around the review room, where spectators argue about camera angles and the exact moment of contact. The grey zone of the laws does not disappear with technology; it only changes address. That is why I no longer treat the number of reviewed points as a measure of fairness.

Signals for the next cycle

Three signals I will track in the coming cycle.

Pure winner rate among players under twenty-two. If this group raises conversion rate while holding its win count, that is real progress. If win rate rises while conversion rate stays flat, most of the gain came from opponent errors and will not hold.

Points to defend over the next twelve months for the highest-ranked players. I will rebuild the model with a points-to-defend variable and publish the error margin before publishing the projection, so as not to repeat the old mistake.

Player movement between managing units at the national championship. This is the least noisy channel for seeing who a unit is investing in, and it usually leads international results by one to two seasons.

xG does not sign contracts, but it tells me where I am putting my pen. The same principle applies to badminton: metrics do not decide tournament entries, but they tell me what evidence I am relying on when I recommend a name. I trust data, but I trust process more — because process is the only thing still standing after a model collapses.

Method note

All figures in this article were recorded and calculated by the author from public results of the international tournament system and streamed match video over a three-year period. The sample of 168 matches with complete rally data is small; all conclusions are group-level tendencies, not claims about individuals. The error margin for hand-recorded metrics is estimated at plus or minus three percent due to subjective classification of points. This article is for sports information reference only and does not constitute any betting advice. Competitive outcomes carry high uncertainty; readers should treat the analysis rationally and cross-check against original data sources.