# Nylon Calculus: The NBA Finals and clustering team offensive styles

We’re rapidly approaching the three-play between the Golden State Warriors and the Cleveland Cavaliers, and the anticipation for this NBA Finals matchup has been building since (let’s face it) the beginning of the season. Despite the obviously stacked talent level on both teams, two divergent narratives have stood out throughout these playoffs.

For the Cavaliers, the focus is perhaps misleadingly on the singular generational greatness of LeBron James. For the Warriors, the focus is on the impossibly daunting task of trying to pick the poison on a team where most everyone’s poison densities range somewhere from poison dart frog to king cobra. Steph Curry’s true shooting percentage this postseason has been an absurd 67.5 percent and barely anyone has batted an eye.

Those narrative differences go hand in hand with the philosophical differences in the designs of both teams’ offenses. The NBA has become a pace-and-space league, with teams playing at record speeds and hoisting shots from deep at record rates. However, even in the midst of this paradigm shift, teams play with distinct styles and tendencies. Ian Levy has done some visualization for defining various offensive styles, and I’ve even previously looked at stylistic differences based on the speed of ball movement.

However, inspired by Ryan Stimpson’s recent work on clustering hockey teams, I decided to go a step further. Drawing on Synergy data from NBA Stats, I compiled 20 statistical features for all 30 teams, from 3-point attempt rate to offensive rebound percentage to frequency of possessions ending in pick-and-roll. It’s important to note that I only focused on style, not effectiveness, in essence looking at process over results. Using that data, I ran a K-Means clustering algorithm to try and define the true offensive styles from this past season. The resultant clusters are visualized in a 2D spatial plot below:

The algorithm returned optimally three distinct types of offensive attacks — teams that fit pretty neatly into the modern pace-and-space paradigm, teams that were throwbacks to a bygone ground-and-pound era, and teams that could generally be traced back to their primary offensive initiators (I’m using this term rather than the position restrictive ‘point guard’ so as to include players like LeBron).

Such a clustering has various applications, from analyzing how teams fare defensively versus certain styles to identifying players to target in transactions. It’s important to understand that as with any clustering, not all teams will fit neatly into the characteristics of their clusters. There are outliers and teams close to their decision boundaries. Not every team in a certain cluster behaves the same way across all statistical features.

The first category is defined by teams like Golden State and Boston who are typically egalitarian offenses, operating at a fast pace and swinging the ball around. The second category is defined by teams like San Antonio and Minnesota who hold the ball, trying to pound it into the post and take mid-range jumpers, an affront to current ideas of efficient shot selection. The third category, on first glance, is somewhere in between. However, in general, the teams in this cluster boast good to great offensive initiators, which explains why they do a good job of taking care of the ball.

Curiously enough though, despite the presence of these good point guards (and LeBron), the teams posted subpar assist percentages and points created off assists. It’s almost an inverted category, defined by what they (perhaps surprisingly) did not do rather than what they did. Below are (brace yourselves) radar charts1 comparing the average percentile ranks across these three categories:

One of the first things we notice about the clustering is the presence of four outliers — the Warriors, Rockets, Raptors, and Pistons. Each of them are outliers for various reasons. The Warriors were tops in the league in assist percentage, touches in the paint, transition frequency, cutting frequency, and points created off assists. The Rockets’ Moreyball offense was tops in the league in 3-point attempt rate (by a sizeable margin), and a close second in free throw attempt rate, as to be expected.

On the other side, the Raptors were tops in frequency of possession ending in a pick-and-roll play and average seconds per touch, while coming in second in drives per 100 possessions. They were also worst in the league in both assist percentage and points created off assists as well as in cutting frequency. The Pistons led the league in the percent of their points coming from midrange, while being the worst in the league in free throw attempt rate. As to be expected from a team in their clustering, Detroit was near the bottom of the league in 3-point attempt rate as well.

The other key takeaway is that styles do not determine effectiveness or team success. Among the four conference finalist teams this season, all three offensive styles were represented. The system alone doesn’t create success; it is dependent just as much on the players that are tasked with executing the system.

It is interesting to note though, that young teams that are often floated as the future of the league, such as the Bucks, Nuggets, and 76ers are all in the ‘Movers and Shakers’ cluster, no doubt a reflection of where the league is heading. The only generally accepted “future” team which is in a different cluster is Minnesota, full of poor shooters and attracted to pounding the ball in the post while staying inside the arc.

But back to the present — what about the upcoming Finals matchup?

While both teams sport tendencies central to modern NBA offenses such as their high 3-pointer attempt rates and transition frequency, the differences aren’t hard to spot. Golden State, with their egalitarian offense, plays a way faster game than Cleveland, where LeBron James is the primary offensive initiator. Golden State throws assists like no one else in the league, cuts, and gets into the paint far more often than Cleveland.

Next: Nylon Calculus -- Smart rotations are key to beating the Golden State Warriors

The Cavaliers, however, isolate at a league-leading rate and hold the ball far longer on each touch. Each of those factors is philosophically representative of the way both teams are constructed, one with a smorgasbord of stars and the other to revolve around one of the top five players in NBA history. Like the patterns displayed in last year’s Finals matchup, whoever wins the series will have done so staying true to their style. LeBron and his cavalry vs. the Bay Area villains — I can’t wait.

1My quick aside about radar charts: Obviously I’ve used radars with a fair amount of regularity in my previous work. Daryl Morey, Luke Bornn, Ted Knutson, Todd Whitehead, and other have all made important points in outlining the scope of usefulness of radar charts. As long as the variable ordering is made in a quasi-sensible way (such as grouping certain features into adjacent axes) and is held consistent, my belief is that it can be an effective way to quickly communicate a visual understanding of how individual observations stack up to each other across a wide range of features. The overarching geometry of a radar helps to provide a visual definition to certain observations in a way that can’t as easily be gleaned from simply looking at a table of data. Like Ted Knutson stated, as with anything, if the limitations and context are properly understood, radar charts can be a quite useful visualization.