Last season, the NBA took its stats game up a level. The league began tracking the movements of every single player and the ball during every second of every game. Six cameras—three on each side of the court—were installed on the catwalk above every NBA arena. This video data is fed into a computer vision system that spits out the positions of all the players, which can be paired with traditional statistics to better understand what’s happening on the court.
The NBA itself is using this positioning data to figure out several new statistics, including the distance players travel throughout the game, how many rebounding opportunities they had, and the number of passes they made. The data can also be used to create fascinating visualizations of the action. Fathom Information Design created the diagram last year with the data from a single NBA basketball game: the February 11, 2011 matchup between the San Antonio Spurs and the Oklahoma City Thunder. Fathom was also able to chart Tim Duncan’s forward trails against Thabo Sefolosha’s shooting guard movements. Duncan's down-the-center plodding is in gray. Sefolosha's wider diagonals are in blue.
And now, with a year of complete data to play with, sports analysts have begun to use this new dataset to analyze basketball in unprecedented and newly rigorous ways.
That translates into new research, and a Harvard-student led team presented some of it at the MIT Sloan Sports Analytics Conference in Boston this past weekend. The researchers wanted to answer a notoriously difficult question: who is good at defense?
Offensive output is fairly easy to measure because the outcomes—made baskets, points, field goal percentage, assists—are naturally captured in the statistics of the game. More simply: it’s easy to count a basket. It’s much harder to quantify a defensive player forcing a miss. Who is the best? And how can that be put into numbers?
The first question has a surprisingly clear-cut answer, at least for guards. Los Angeles Clippers guard Chris Paul is the best defensive guard by almost every measure looked at by the researchers.
Using the new player tracking data, the researchers were able to create a model of who was defending whom at all moments in the game. And that let them create two new types of metrics: the volume score (or shots taken against a defender) and the disruption score (or amount a defender decreases the efficiency of the offender). Laying those out on a diagram of the court, they could put together what they called the “defensive shot chart” — or, a map of all the spots on the court where a given defender gives up shots and points.
Even assigning blame for points scored in slightly different ways, to try to account for the fluidity of possessions, Paul shows up at the top of the list, along with Norris Cole, Nick Calathes, C.J. Watson, and Greivis Vasquez. Among wing players, the journeyman Mike Dunleavy is atop the list.
These stats have always been easy to get, for players on offense—it's just the number of shots, and the percentage of shots made. But to generate them for defense, you need the precision of machine-vision data. It would take an obscene amount of time to hand-code all the game video into statistics for one game, let alone for all games. And machine-vision data is required to calculate some of the defensive stats, too. For example, the model in this paper takes into account how close the defender is to the shooter.
One might anticipate that NBA teams will continue along the path to full Moneyballization, the quantification trend that swept through Major League Baseball, led by the Oakland A’s Billy Beane, among others. The league’s experiment with tracking players began with 15 teams in 2010-2011 and since last season, has expended to all 30 teams. So it must be working—for teams, fans, or both.