That counted as original, in 2006. Morey could see that no one else was using a model to judge basketball players—no one had bothered to acquire the information needed by any model. To get any stats at all, he’d had to send people to the offices of the National Collegiate Athletic Association (NCAA), in Indianapolis, to photocopy box scores of every college game over the past twenty years, then enter all that data by hand into his system. Any theory about basketball players had to be tested on a database of players. They now had a twenty-year history of college players. The new database allowed you to compare players to similar players from the past, and see if there were any general lessons to be learned.
A lot of what the Houston Rockets did sounds simple and obvious now: In spirit, it is the same approach taken by algorithmic Wall Street traders, U.S. presidential campaign managers, and every company trying to use what you do on the Internet to predict what you might buy or look at. There was nothing simple or obvious about it in 2006. There was much information Morey’s model needed that simply was not available. The Rockets began to gather their own original data by measuring things on a basketball court that had previously gone unmeasured. Instead of knowing the number of rebounds a player had, for instance, they began to count the number of genuine opportunities for rebounds he’d had and, of those, how many he had snagged. They tracked the scoring in the game when a given player was on the court, compared to when he was on the bench. Points and rebounds and steals per game were not very useful; but points and rebounds and steals per minute had value. Scoring 15 points a game obviously meant less if you had played the entire game than if you had played half of it. It was also possible to back out from the box scores the pace at which various college teams played—how often they went up and down the court. Adjusting a college player’s stats for his team’s pace of play was telling. Points and rebounds meant one thing when the team took 150 shots a game and something different when it took just 75. Just adjusting for pace gave you a clearer picture of what any given player had accomplished than the conventional view did.
The Rockets collected data on basketball players that hadn’t ever been collected before, and not just basketball data. They gathered information on the players’ lives and looked for patterns in it. Did it help a player to have two parents in his life? Was it an advantage to be left-handed? Did players with strong college coaches tend to do better in the NBA? Did it help if a player had a former NBA player in his lineage? Did it matter if he had transferred from junior college? If his college coach played zone defense? If he had played multiple positions in college? Did it matter how much weight a player could bench-press? “Almost everything we looked at was nonpredictive,” says Morey. But not everything. Rebounds per minute were useful in predicting the future success of big guys. Steals per minute told you something about the small ones. It didn’t matter so much how tall a player was as how high he could reach with his hands—his length rather than his height.
The model’s first road test came in 2007. (The Rockets had traded their picks in 2006.) Here was the chance to test a dispassionate, unsentimental, evidence-based approach against the felt experience of an entire industry. That year, the Rockets held the 26th and the 31st picks in the NBA draft. According to Morey’s model, the odds of getting a good NBA player with those picks were, respectively, 8 percent and 5 percent. The chance of getting a starter was roughly one in a hundred. They selected Aaron Brooks and Carl Landry, both of whom became NBA starters. It was an incredibly rich haul.? “That lulled us to sleep,” said Morey. He knew that his model was, at best, only slightly less flawed than the human beings who had rendered the judgments about job applicants since time began. He knew that he suffered from a serious dearth of good data. “You have some information—but often from a single year in college. And even that has problems with it. Apart from it’s a different game, with different coaches, different levels of competition—the players are twenty years old. They don’t know who they are. So how are we supposed to?” He knew all this and yet he thought maybe they had figured something out. Then came 2008.
That year the Rockets had the 25th pick in the draft and used it to pick a big guy from the University of Memphis named Joey Dorsey. In his job interview, Dorsey had been funny and likable and charming—he’d said when he was done playing basketball he intended to explore a second career as a porn star. After he was drafted, Dorsey was sent to Santa Cruz to play in an exhibition game against other newly drafted players. Morey went to go see him. “The first game I watch he looks terrible,” said Morey. “And I’m like, ‘Fuck!!!!’” Joey Dorsey was so bad that Daryl Morey could not believe he was watching the guy he’d drafted. Perhaps, Morey thought, he wasn’t taking the exhibition seriously. “I meet with him. We have a two-hour lunch.” Morey gave Dorsey a long talk about the importance of playing with intensity, and making a good impression, and so on. “I think he’s going to come out the next game with his hair on fire. And he comes out and sucks the next game, too.” Fairly quickly, Morey saw he had a bigger problem than Joey Dorsey. The problem was his model. “Joey Dorsey was a model superstar. The model said that he was like a can’t-miss. His signal was super, super high.”
That same year, the model had dismissed as unworthy of serious consideration a freshman center at Texas A&M named DeAndre Jordan. Never mind that every other team in the NBA, using more conventional scouting tools, passed him over at least once, or that Jordan wasn’t taken until 35th pick of the draft, by the Los Angeles Clippers. As quickly as Joey Dorsey established himself as a bust, DeAndre Jordan established himself as a dominant NBA center and the second-best player in the entire draft class after Russell Westbrook.?
This sort of thing happened every year to some NBA team, and usually to all of them. Every year there were great players the scouts missed, and every year highly regarded players went bust. Morey didn’t think his model was perfect, but he also couldn’t believe that it could be so drastically wrong. Knowledge was prediction: If you couldn’t predict such a glaringly obvious thing as the failure of Joey Dorsey or the success of DeAndre Jordan, how much did you know? His entire life had been shaped by this single, tantalizing idea: He could use numbers to make better predictions. The plausibility of that idea was now in question. “I’d missed something,” said Morey. “What I missed were the limitations of the model.”
His first mistake, he decided, was to have paid insufficient attention to Joey Dorsey’s age. “He was insanely old,” says Morey. “He was twenty-four years old when we drafted him.” Dorsey’s college career was impressive because he was so much older than the people he played against. He’d been, in effect, beating up on little kids. Raising the weight the model placed on a player’s age flagged Dorsey as a weak NBA prospect; more tellingly, it improved the model’s judgments about nearly all of the players in the database. For that matter, Morey realized, there existed an entire class of college basketball player who played far better against weak opponents than against strong ones. Basketball bullies. The model could account for that, too, by assigning greater weight to games played against strong opponents than against weak ones. That also improved the model.