Nine Games Is Not Enough to Sign a Max Contract

One weekend afternoon, a request landed in my inbox from a scouting department. Attached was a valuation sheet for a twenty-year-old player, with exactly one question attached: should we commit to a maximum contract or not. I opened my dataset. Nine games. Two hundred and fourteen minutes. Sixty-eight shot attempts. Nothing in the spreadsheet was broken. It was simply missing data. I left the verdict cell empty and typed into it the sentence no scouting department wants to read: insufficient information to conclude.

Three days later the file came back with a red comment. The sender asked whether I understood that the opportunity comes once, that the club could not wait another twenty games to learn whether it should sign. I understood. Everyone in this job understands. But knowing you must decide is not the same as having enough ground to decide correctly. Between those two things lies a gap, and most bad contracts in professional basketball are signed inside that gap.

An industry maturing faster than the market can absorb

Fifteen years ago, a basketball analysis sheet in Vietnam contained points, rebounds and assists. To say anything about shot quality, a writer had to use his eyes. Today every NBA game generates millions of positional data points: camera systems tracking the ball and twenty-two players at dozens of frames per second, classifying each possession into more than a hundred play types, and feeding composite metrics such as EPM, DARKO or LEBRON. A club in Hanoi or Ho Chi Minh City can sit at home and rewatch every shot attempt of any player in any league.

The paradox is this: the more data there is, the harder it becomes to see the line between conclusion and guess. When a spreadsheet has five columns, everyone knows what is missing. When it has five hundred, people tend to believe they have everything.

In basketball there is one thing I always check before saying anything about a player: sample stability. Free-throw percentage tends to stabilise after only a few dozen attempts. Three-point percentage needs roughly seven hundred attempts to enter a stable range. Individual defensive metrics need more still. A player can go through an entire season and still not leave behind enough data to answer the simplest question about his true shooting ability.

Evidence chain: from one game to one signature

A few years ago I spent weeks reconstructing a game that almost everyone remembers only by its result. It was Game 7 of the 2026 Western Conference Finals, when the Houston Rockets missed twenty-seven consecutive three-pointers and lost 101-92 to the Golden State Warriors. That streak entered history as one of the most famous collapses in modern basketball, and it is usually told as proof of how fragile a jump-shooting identity can be.

Using twenty-seven straight misses to reprice an entire franchise is a methodological error. That same season, Houston finished the regular season 65-17, the best record in the league, and that is the sample with weight. The streak lasted twenty minutes. The sixty-five wins lasted six months. Only one of those two slices can be used for forecasting. There is also a detail many people forget: Chris Paul, Houston's primary organiser, did not play in Games 6 and 7 because of a hamstring injury. A variable that large, left out of the story, means the story is being told with emotion.

The same thing happens at team level. In the 2026-16 season the Golden State Warriors won 73 regular-season games, according to Basketball Reference, breaking the record the Chicago Bulls set in 2026, and then lost the Finals in Game 7 to the Cleveland Cavaliers after leading 3-1. There are two ways to tell it. One is to treat the regular-season record as meaningless, with basketball decided over seven games. The one I choose is this: seventy-three wins remains the strongest signal of that team's quality, while the final collapse was shaped by variables outside the model, among them Draymond Green's suspension in Game 5, Andrew Bogut's injury, and a run of one-on-one possessions that went the wrong way in the last two quarters.

I do not believe in hunches. But I believe in what a hunch looks like once the data confirms it. Seventy-three wins is data. Game 7 is the crowd's hunch.

Variance in small samples is far larger than intuition suggests. A player whose true three-point rate is 35 percent can quite easily shoot 45 percent or 25 percent over his first nine games without changing in any way. That noisy curve reads to the eye as a trend, and then becomes a headline. Nine games do not create a new player. They create a sample small enough that anyone can find in it whatever they want to find.

Why the market still pays maximum money for a thin sample

If everyone knows nine games is too little, why does the market still pay? Because the basketball transfer market prices options, not history. A twenty-year-old with nine good games is not paid for what he has done but for the probability that he becomes a star within four years. The buyer is paying for upside variance, and in a league where only about thirty people are considered genuine franchise cornerstones, the reward for finding the thirty-first is large enough that overpaying becomes an acceptable cost.

That is why I do not entirely agree with those who describe the young-player price bubble as evidence of madness. It is the consequence of a specific incentive structure: salary rules cap what a single star can be paid but do not cap how many stars a team may hold, so the price of hope rises faster than the price of achievement. I still believe the bubble is deflating. But I do not think those paying the highest prices are people who cannot read data. They are reading a different kind of data: data about probability.

That makes my job harder. If the question is how good a player is, I can answer. If the question is how much to pay, most of the answer lies outside my dataset.

The VBA problem: when an entire season is a small sample

In the VBA the problem is sharper. The regular season lasts only a few weeks, each team plays far fewer games than an NBA team plays in a single month, and one minor injury is enough to distort a club's entire statistical picture. In a league like that, almost every dataset is thin, and almost every conclusion needs a warning line attached.

My work there takes a different shape. I rarely have the sample to say with confidence which team is stronger. I can say which team is running a sound process: shot quality, free throws generated, turnover rate, and the stability of decisions in the final two minutes. In a league with too few games, process is the only thing thick enough to analyse.

The contrarian angle: the danger is not the empty cell

There is a paradox it took me years to understand, and I paid to understand it. An empty cell makes people cautious. A full spreadsheet makes them confident. Between those two states, confidence is the dangerous one.

In 2026 I held the most complete dataset I had ever built for a group stage of a major tournament: shooting data, possession data, positional data, all of it clean. I concluded that a leading national team would advance. It went out. Looking back, I realised my sheet was missing an entire column: the pressing intensity of its opponents, which at the time I was not collecting at all. The spreadsheet was not empty. It was missing exactly one column, and that column decided everything.

Two years earlier, during the period when competitions had to be played in empty stadiums, I went through something similar. A home-advantage model I had spent six years building suddenly produced systematically wrong numbers. When the stands emptied, my model collapsed. I knew I had forgotten the human factor.

Both times I did not lack data. I lacked the right data. And so, when a younger colleague asks me how to avoid being swept up by a rising player, I do not tell him to wait for a bigger sample. I tell him to list what he is not measuring.

Risks and gaps

I always put this section at the end, and this time it matters more than usual, because the subject of the piece is missing data itself.

The largest risk belongs to the writer: nine games can be used to build a compelling story, and that story will then sustain itself on the reader's belief. On the club's side, waiting for a bigger sample means the market takes the player. And deeper still, my own model carries risk: the metrics I use to evaluate young players were largely built on NBA data, and they transfer imperfectly to a league with a different pace, rule set and level of roster quality.

The biggest gap is injury prevention, and psychology. No spreadsheet of mine measures how a twenty-year-old reacts to being handed maximum money at twenty-two, when everything arrives too fast and there is nothing left to prove. That is a variable I never put in the model, and it has ruined more contracts than every injury combined.

What would change my answer

Back to the file with the empty verdict cell. Insufficient information to conclude is not a final answer. It is a conditional one, and the conditions have to be spelled out.

For a twenty-year-old after nine games, some signals stabilise faster than three-point percentage, and those are the ones I would track over the next twenty games. Free-throw rate and free-throw attempts per shot attempt settle early, and they indicate whether a player can generate contact, a skill far harder to teach than shooting. Decision speed in two-on-two situations settles relatively fast. And most importantly, I would compare the quality of a player's attempts with their outcomes. If a player is shooting 25 percent from three but every attempt is open and on time, that points to a good process waiting for results rather than a bad player. The reverse is also true: a player shooting 45 percent across nine games on mostly forced, broken attempts is the case where I want to rewatch the film before trusting any number.

I am not trying to inflate the sample at any cost. I am trying to work out whether the sample I have is measuring the right thing.

A few weeks later the scouting department sent me a new file. This time it was eighteen games. I still could not sign anything, but for the first time I could write a sentence that did not begin with an if. That is the whole of the progress data brings to my work: questions that get more precise, rather than conclusions that get more certain.

Numbers show a trend, not a prophecy. A number never needs us to defend it. We need numbers so that we do not lie to ourselves. And a contract is only truly correct when the number signs alongside the signature.