The Empty Data Sheet and the Trap of Confident Judgement in Basketball Analysis

On Tuesday morning, a scouting file landed in my internal inbox. The filename followed my own naming convention exactly: VBA_Scout_Q4.json. I opened it. The first line read "domain": "basketball". Every line after that was empty. No shooting percentages. No player names. No minutes played. No turnover counts. What remained was a label reading "basketball" sitting on top of nothing at all.

In thirteen years on the job I have received every kind of broken file. Files that failed because the connection dropped. Files missing columns. Files with the wrong date format. This was the first time I received a file that was both empty and apparently valid. The label stuck on top made it look like a report that had finished running, only short. A hurried reader would assume the output was clean. A hurried reader would start writing.

I closed the file. Opened it again. Still empty. Then I did something I would never have done thirteen years ago: I wrote nothing.

To understand why an empty file is more dangerous than a missing one, you have to look at how basketball data travels from the edge of the court to the analyst's desk.

At a professional club, that journey has four stages. Stage one is collection: wide-angle cameras, tracking systems, or simply a person writing down every action, second by second. Stage two is extraction: raw data becomes a table with columns, rows and units of measurement. Stage three is cross-checking: someone compares minutes, shot attempts and total points against the official scoresheet. Stage four is analysis. That is where I sit.

When one stage breaks, the consequences differ completely. If stage one breaks, the whole game vanishes and everyone can see it. If stage two breaks crudely, you get a mess of garbled characters, obviously wrong. But when stage two breaks subtly — running the full process, raising no error, then returning nothing — what it leaves behind is not a mistake but a shell. And a shell has a label.

In basketball this kind of failure happens more often than outsiders imagine. A ball-tracking system can misread arena lighting, lock onto the wrong jersey colour, and return a correctly formatted file full of zeroes. At competitions where filming conditions vary between arenas, the rate is noticeably higher. I once received a forty-minute game sheet where the combined shot attempts of both teams added up to three. Three attempts in a basketball game. That figure was not technically wrong. It was correct for the camera that had failed.

The problem lies in the reader's reflex. People tend to trust things that have the right shape. A JSON file with correct brackets, correct keys, correct field names will pass through every check because nobody bothers to open it up. The "basketball" label at the top does the rest: it tells the reader that the subject has already been identified. From that point on, the blanks behind it stop being errors. They become space for the reader to fill in.

This is the part I want to say slowly. Four chains of evidence. Four times the data said one thing and people heard another.

Chain one: the lesson of a striker whose xG exceeded his goals.

In 2026 I was twenty, a third-year student in Da Nang, writing a personal blog analysing expected goals (xG) for a V-League club. I pointed out that striker Gastón Merlo averaged 0.8 xG per match but converted at an actual rate of just 0.4. In other words, he generated chances at the level of a top forward but finished them at an average one. A young coach at another club mocked me online: what does a girl know about tactics, don't read a few numbers and start spouting off.

I did not argue. I published the full dataset for the following twelve matches: shot counts, shot locations, shot timings, shot situations. The club collected only 9 of a possible 36 points, exactly as the model had predicted. That coach issued a public apology.

The point I want to stress is not that I was right. The point is that I was only right because I had a column of numbers to check against. If all I had held that day was a label — "Merlo is a good striker" — with blank space beneath it, I would have had nothing to say. Every argument of mine would have sunk to the same level as that mockery, just more politely worded.

Chain two: Germany in 2026 and the log file nobody opened.

In 2026 I was interning at a sports outlet. Before the World Cup I analysed the German national team using PPDA — the number of passes an opponent is allowed before each defensive action. Germany's qualifying PPDA was 12.5, while the average of the five most recent World Cup winners was 9.8. The team's average running distance was only 98 kilometres per match. I wrote a piece predicting Germany would be eliminated in the group stage. Colleagues called me the laboratory scientist.

On 27 June 2026, Germany lost 0-2 to South Korea, finished bottom of Group F and went home. My article was shared more than five thousand times. But this time I do not want to tell a story about winning and losing. I want to point at something else: for weeks beforehand, nobody reopened the log file. The label "reigning champions" was heavy enough to crush every number behind it.

In 2026 the whole world mourned Germany. I quietly reread the model's log file.

Chain three: empty stands and the data that does not sit on the pitch.

In 2026, when competitions were suspended, I worked as a data analyst for a sports consultancy. I collected figures from three hundred matches across eight European leagues played without spectators and found that the home win rate fell from 45 per cent to 38 per cent. Home advantage, which every coach treats as a constant, turned out to lie mostly in the stands rather than on the grass.

I sent a report to a team near the bottom of the table, recommending they press high from the opening whistle in away matches, because opponents had lost the sound of their own crowd. The head coach was initially sceptical. After testing it in the second half of the season, the team took 12 of a possible 15 points from five away games, having taken only 6 of 15 before.

The lesson is not about pressing or not pressing. The lesson is that data does not stop at the touchline. It also lives in travel schedules, in congested fixtures, in whether the stands have people in them. A model that only reads on-pitch numbers will miss precisely the variable deciding the result.

Chain four: translating metrics from English into practical Vietnamese.

After years of writing I realised something: most basketball fans in Vietnam do not need to know about log loss or overfitting. They need to know what a number says about tonight's game. So I always translate.

Usage rate (USG%) is the share of possessions that end with that player — a shot, a turnover or an assist. A player with high USG% and low shooting efficiency is someone consuming possessions without giving anything back. True shooting percentage (TS%) folds in free throws and three-pointers, telling you how many points each shooting possession is worth. Effective field goal percentage (eFG%) adds extra weight for three-pointers.

Put those three numbers side by side and people stop arguing about whether a player is good, and start asking: which part of his game is good, and is it repeatable? Good data does not end an argument; it moves the argument to the right place.

The trap: three ways of filling in a blank.

Now back to that empty file on Tuesday morning. I asked myself: if I were a less disciplined analyst, what would I fill in?

The first way is to fill it in with memory. He shot well last game, so he will probably shoot well this one. That is data from a single sample, promoted to data for an entire season.

The second way is to fill it in with reputation. He just signed a big contract, so he must be effective. That is a habit of anchoring to price rather than to output.

The third way is to fill it in with season averages. This is the most dangerous, because it looks professional. People take the season average, paste it onto one specific match, and draw conclusions about that match. But a single game is a very small sample, and small samples prove nothing with certainty.

All three lead to the same product: an analysis that sounds extremely confident and has no roots. The most dangerous thing in this profession was never an empty analysis. It is a confident analysis with no roots.

Data is a monastery: the less noise there is, the more clearly you hear something trying to speak. But a monastery so silent that nobody is inside is not sacred. It is just an empty room, and I have to tell the coaching staff that this room is empty.

A label is not evidence.

In this profession there is a mistake I call the comfort error. It happens when an external marker is used as a substitute for internal content. The "basketball" label at the top of the file is one such marker. The competition name printed on the cover of a report is another. The job title of the person who signed it is a third.

None of those markers proves the content was processed correctly. The label sits on the upper layer, the content sits on the lower one, and the two can fail independently. When the upper layer is intact and the lower one is empty, the reader is deceived in the subtlest possible way: they think they are missing a little information, when in fact they are missing everything.

Across thirteen years watching basketball at many levels, from domestic competitions to international ones, I learned one thing: the order of trust lies not in how bright the label is, but in how full the data is. A faint label on a full sheet beats a bright label on an empty one.

What I check before trusting a data sheet.

Every time a sheet reaches me, I run four questions in the same order.

First, do the totals reconcile? If the official scoresheet says Team A scored 78 and my sheet adds up to 71, then everything else is worthless until I find where those seven points went.

Second, which columns are empty? A blank column in a full sheet usually signals a system error rather than a coincidence.

Third, are the units consistent? Minutes, seconds, percentages and integers mixed in one column is the fastest way for a model to produce garbage that still looks tidy.

Fourth, provenance. Who recorded it, when, and with what equipment. A number without a source is decoration.

These four questions need no complex model. They need a person willing to sit still and not rush. Data discipline does not live in the algorithm; it lives in whether you are willing to stop at the blank column.

There is something worth saying about the domestic basketball transfer market, where clubs often judge imports on a handful of exhibition games or on a recommendation. There, the label once again overwhelms the content. A familiar-sounding name, a CV featuring a big league, a three-minute highlight reel full of beautiful shots — that is the label. Actual minutes played, shooting percentage under pressure, turnover rate against physical defence — that is the content. Clubs pay for the label, then are surprised when they receive the content.

At this point I have to say what nobody in sports media wants to hear: sometimes the most correct answer is to say nothing.

The market rewards certainty. Someone willing to assert gets the clicks. Someone who says "I don't know" is treated as weak, unqualified, not hard-working enough. But in data analysis, "I don't know" is a statement with full evidence behind it — it is the result of a process that ran to completion and returned nothing. Saying "I don't know" in this case is not surrender. It is a conclusion.

There is a correlation I keep seeing: the more confident an analysis sounds, the higher its error rate, and the more cautious it is, the fewer people read it. These two variables are not causally related. Certainty does not produce error, and caution does not produce truth. But both correlate inside the same market, and the market is paying for the wrong side.

The counter-intuitive part is this: people think data exists to produce answers. Most of the time, data exists to eliminate answers. Every time I run a model and it returns a weak result, I have not lost an opportunity. I have eliminated a hypothesis. Eliminating hypotheses is a form of progress, just a little less photogenic.

There is another mistake I call the hindsight error. Once a prediction has come true, people retell it as though it had been certain from the start. But every honest prediction carries the conditions under which it would fail. My 2026 model on Germany was only right because it was tied to one specific dataset. If Germany had changed shape, if a key player had recovered from injury, if the first two matches had gone slightly differently, the conclusion could have flipped. Not stating the conditions under which you would be wrong is a very polite way of lying.

Every coach talks about feel. I do not have feel; I have standard deviation. And the standard deviation of an empty file is something I cannot ignore.

So I return to the opening question: where does that empty file go inside a pipeline with no gate? It gets passed to the next step. The next step adds a layer of analysis on top of it. That layer is written into a report. That report reaches the coach. And by then, the error no longer lies in the empty data. It lies in the fact that somebody looked at a void and confidently stated a conclusion.

Contamination through a data chain never happens through one large error. It happens through many small layers, each only slightly more confident than the last. At the end of the chain, an initial blank has become a claim that sounds solid, and nobody can trace where it started.

The signal for the next analysis cycle is not that I found a broken file. It is the gate I will now place before every subsequent step: if the data-point count is empty and the entity list is empty, the process halts with an explicit warning. No inference from labels. No filling blanks with memory.

Numbers do not lie, but they do not tell stories either — and an empty sheet has not even had the chance to speak.

People watch goals to remember a match. I watch xG to understand the match that never happened.

People look at a full scoreboard to believe they understood. This Tuesday morning, I looked at an empty file and understood that I knew nothing at all. The distance between those two states is the whole of my profession, and it is also the only boundary keeping this trade from turning itself into a machine that mass-produces hollow certainties.