Now That We Have Your Attention — Demonstrating Basketball Bonafides at Sloan 2015

Feb 20, 2015; Dallas, TX, USA; Houston Rockets guard James Harden (13) laughs while guarding Dallas Mavericks guard Rajon Rondo (9) during the first quarter at American Airlines Center. Mandatory Credit: Kevin Jairaj-USA TODAY Sports

This past weekend at the MIT Sloan Sports Analytics Conference, I went into the presentation of the “Counterpoints: Advanced Defensive Metrics for NBA Basketball” research paper ready to hate it. In my mind it had several strikes against it already: it involved a top-down one-number model; it possibly purported to “solve” defense; and most importantly, it did so in a conceptually problematic manner.[1. While understanding the desire to want to capture the value of the defensive contributions of individual players, I strongly feel that reducing to individual shot defense and prevention is too narrow a look at defense and by focusing so narrowly on “matchups” threatens to obscure as much if not more than it illuminates. But that’s a problem with the basketball side of the paper rather than the statistics. Further it’s an issue which the authors to some degree owned up to as they noted application of the model required much more refinement including more and better inclusion of basketball expertise.]

In short, it appeared like it might be an example of solid statistical methods reaching conclusions of questionable utility for the sport because of inappropriate or misguided assumptions about the underlying game of basketball as played on the floor. I was quickly convinced otherwise. The presenters, Alexander Franks and Andy Miller had me in basically one slide.

Backing up slightly to describe the paper, first discussed here by Kirk Goldsberry, the research produced a model which could comb through SportVU data and identify with high accuracy which player a given defender was guarding at any point during a possession. Or, more accurately, how much the defender was “guarding” each of the five opposing players at any given time. Regardless of anything else in the paper, this represents a potentially massive step.

No matter what the next step in valuing defense will be, identifying individual responsibilities will be a key component. Learning who is guarding whom is an invaluable first step, one that would be complicated enough if the NBA was primarily played in the style of most pickup games with an emphasis on one-on-one play. This would still mean around 250,000 individual possessions would need to be charted to get even one seasons worth of such data. Add in the complexity of NBA defense with all its designed rotations, switches, hedges, traps and the time needed to discover this information manually becomes prohibitive. So an algorithm which could make these determinations from player tracking data not only speeds the process but makes it possible to obtain.

Even without a holistic method of valuing defense, just this data could be tremendously valuable in terms of coaching evaluation and game prep. That is, if you trust the data. As Celtics Assistant GM Mike Zarren said during the basketball panel, “What people who say ‘I don’t believe those numbers’ really mean is ‘I don’t believe that guy telling me those things.'” Passing the “how full of crap is this guy?” is the most important hurdle to clear. And it’s clear it Franks and Miller did. Below is the slide shown as they discussed the basic output of their model in terms of where it most often found a defender relative to the player he was guarding:

Slide from “Counterpoints: Advanced Defensive Metrics for NBA Basketball” by permission of Andy Miller

For a “basketball type” the right side of the slide might as well be in the same Greek as many of those symbols. But the other half? Pure basketball. Defense at it’s most fundamental. Defined by triangles, ball-man-basket and ball-defender-man. On the line, up the line. If the model says this is what defense looks like on average, I’m listening to the rest of what you have to say. And if I don’t understand it, I’m might just call someone who does to explain it to me. Because you’re speaking my language, so now we can talk.