When the Basketball Data Sheet Returns Null: Why a Good Analyst Must Learn to Say I Don't Know

On the morning of March 6, I opened the scouting file of a Vietnamese professional basketball club and received exactly one value: NULL. The import in the file had not played a competitive game in fourteen months, since his anterior cruciate ligament surgery. My filter ran across the 2026 VBA season, the 2026 season, and the training camp in Thailand — not a single possession was logged. The assistant coach knocked, pointed at the screen and asked the question I have heard hundreds of times: can he still play? I did not answer immediately. Thirteen years of watching basketball and football in this country taught me something uncomfortable: most answers to that question in the Vietnamese market are built from memory, not from data. Memory always returns a value. It never returns NULL. That is the real problem, and it has nothing to do with whether that player can come back. Vietnamese basketball has a data paradox that outsiders rarely see. The professional league tipped off in 2026, and each season's group stage runs somewhere between 15 and 18 games depending on the number of teams, at a density of two to three games per week. Stat tracking is almost entirely manual: two or three people key in play-by-play from two camera angles, then reconcile it with the referees' paper scoresheet. The result is that we have points, made shots, personal fouls. We do not have thick enough shot-location data, we do not have reliable individual defensive data, and we have almost nothing on distance travelled. For comparison: one modern NBA game generates roughly 1.4 million positional data points from its camera system, sampling 25 frames per second for all ten players and the ball. One VBA game generates a few hundred spreadsheet rows. That gap is not a technology story, it is a budget story, and it determines which questions are even permitted. Data is a monastery: the less noise, the more clearly you hear something trying to speak. But every monastery has empty rooms, and my job is telling an empty room apart from a room nobody has opened. In every internal report I have written, three states get mixed together, and that mixing is the origin of most bad recruitment decisions in small leagues. The first state is zero. A player shoots 0-for-5. That is complete data: we observed, and the outcome was failure. It can be analysed. The second state is NULL. The player has no record at all, because he did not play, or because the tracking system failed, or because the league did not publish. That is not failure, it is non-observation. The third state is sparse: records exist but there are too few to reason from, say 38 minutes scattered across six games in a different role each time. In many of the files I receive, the NULL cell has been filled with a zero. That is the most serious and also the most common error, because it is invisible. It corrupts the average, it corrupts the model, and worse, it corrupts the contract decision. When a cell is NULL, the meeting room never leaves it alone. Humans have an instinct to fill gaps, and in basketball there are three common fills. The first is the coach's feel. Every coach talks about feel. I have no feel, I have a standard deviation. But I have to be fair: the feel of a man with twenty years in the job is a model trained on a body of data I have no access to. He sees how a player lowers his centre of gravity in a sprint, how he changes direction on the third step. The problem with feel is not that it is wrong, it is that it carries no confidence interval. When a coach says the player is back, nobody knows whether that claim is 60 percent or 85 percent likely to be right. The second is recency. People remember the last three games before the injury, or the first three after the return, and turn that into a season description. This bias exists in every sport, but in a league with a 16-game group stage it is many times more dangerous, because three games is already close to a fifth of the season. The third is media narrative. A player is called the missing piece after two good games, and the label sticks to him all season, even after the data has contradicted it for a long time. People watch goals to remember a match. I watch xG to understand the match that did not happen. That principle applies to basketball in its own way: look at the possessions that were never logged to understand what got dropped from the story. There is an easy class of data problem, and I have met it. In 2026, the whole world mourned Germany. I quietly re-read the model's log file. Germany's PPDA in qualifying was 12.5, well above the 9.8 average of the previous five World Cup winners, alongside an average of 98 kilometres run per game. Every variable was present, complete and public. Predicting that team's group-stage exit was just reading a table that had already been printed. Everyone knows the result. But most of the real work of a data consultant does not look like that. It looks like the NULL file on March 6. For a player who has not played in fourteen months, I have no possessions to compute efficiency, no defensive data, nothing on load tolerance. What I have are base rates from the population of players returning from ACL reconstruction: reduced efficiency in the first season, reduced minutes, and a meaningful re-injury risk. But base rates describe populations, not individuals. A 27-year-old is not a 33-year-old. A 110-kilogram centre loads his knee differently from an 82-kilogram ball-handling guard. My approach is to build three scenarios instead of one forecast. The cautious scenario: the player returns at 60 percent of his old efficiency, plays 18 minutes, fits a rotating bench role. The neutral scenario: 80 percent efficiency, 25 minutes, a controlled starting role. The optimistic scenario: 100 percent efficiency, 30 minutes. Each scenario carries a probability and an explicit verification condition: if after eight weeks the player does not clear the lateral-movement threshold in the internal test, the optimistic scenario is eliminated. That is pricing the gap, not filling it. The report to the coaching staff carries three lines: three-point efficiency estimate with low confidence; transition defence, insufficient data to conclude; injury status NULL since January 2026. Those three lines are much harder to read than a single number, which is why they usually get replaced by a single number in meetings. The small-sample problem in the domestic league is harsher still. A local player has four explosive games and the media immediately calls him a phenomenon. Four games at 30 minutes each is 120 minutes played. Given the typical variance of true shooting efficiency at league level, 120 minutes is not enough to separate a 45 percent shooter from a 52 percent shooter. I have computed my own threshold: roughly 700 to 900 minutes are needed before the confidence interval narrows enough for a conclusion about a shooter. In the VBA, 900 minutes is nearly two seasons. Which means most single-season evaluations of local players are describing luck more than ability. I say this as someone who once erred in the opposite direction. In 2026 I published an analysis of Gastón Merlo at SHB Da Nang, showing an average xG of 0.8 per game against an actual scoring output of 0.4. A young coach at another club commented publicly that a girl knows nothing about tactics. I did not argue. I published the full dataset for the next twelve games, with shot counts and shot locations. The team took only 9 of a possible 36 points, exactly as the model projected. But the part I talk about less: had Merlo scored through those twelve games, I would have had to write a different piece, and that piece would also have been methodologically correct. A prediction coming true does not prove the method. It only proves the method has not yet been rejected. There is another variable that gets left out of every domestic basketball model, and it has saved me many times: context. In 2026, when European leagues played behind closed doors, I collected data from 300 matches across eight competitions and found the home win rate fell from 45 percent to 38 percent. I sent a report to a team at the bottom of the table recommending they push the press high from the opening minute in away games, because opponents had lost their crowd. The head coach was sceptical at first. After testing it in the second half of the season, the team took 12 of 15 away points, against 6 of 15 before. The lesson is not that pressing works. The lesson is that when technical variables are thin, context variables are thick. When you do not have enough data about a player, look for data about the environment he operates in: schedule, travel distance, squad depth, even arena temperature. I have watched domestic basketball long enough to know my own limits. Some things I can measure: shooting percentage, possession counts, efficiency by zone. Some things I only infer from experience: how connected a locker room is, how well a player tolerates a bench role. In my reports I mark those two categories as different. That is the minimum discipline, and it is not exciting. At this point the argument has to turn back on me. Saying there is not enough data can be a very sophisticated way of dodging responsibility. An analyst who says that in a meeting never has to own the contract. The staff signs, the player gets hurt again, and the analyst stays clean. I have seen this pattern many times, in both football and basketball, and it damages the profession faster than being wrong does. A good analyst is not the person who says I do not know the most often. That person is the one who can say: I do not know this part, here is the probability I assign to the rest, and here is the decision I recommend. Refusing to decide is also a decision, just one nobody signs their name to. The second half of the argument belongs to causation. When I cite base rates for players returning from ligament surgery, I am describing a correlation within a population. If this club has a better-than-average medical staff, that base rate may be wrong in the optimistic direction. If this player has a history of re-injury, it is wrong the other way. A model is a map, and the map is not the territory. Veterans understand this so well that they sometimes forget to say it out loud, and then a spreadsheet becomes a prophecy. There is one small check I always run before presenting anything: whether the length of the sample matches the length of the conclusion. A league table after four rounds says nothing about team strength. A player with good defensive numbers after six games may simply not have faced a strong opponent yet. I have seen headlines written from exactly that kind of data, and I understand why they exist: they are easy to read, and readers have no tool to push back. Over the next three months, what I will track is not the standings. I will track each team's NULL rate. The team that starts publishing complete play-by-play, minutes broken down by quarter, and separates training data from game data will build a compounding advantage over two to three seasons. That advantage does not come from a smarter model. It comes from having fewer empty rooms for other people to fill with feel. And the import in the March 6 file? I sent the staff three scenarios, one verification threshold, and a note at the bottom: if he clears the lateral-movement test in week eight, call me, and I will update the model the same day. Numbers do not lie, but they do not tell stories either. My job is to keep the story from being written into the places where the data is still empty. When a young coach tells me his feel is enough to decide, I smile. I touch the future with a keyboard, and a keyboard can record even the times I did not know.