Nylon Calculus: Defining 23 offensive roles using the NBA’s play-type data

At the beginning, there were just three basketball positions — the center, whose job it was to win possession of the ball during the tip-off; the forwards, who pushed the ball ahead to score; and the guards, who hung back to defend the basket.

During this primordial era of hoops, a jump-ball initiated every possession that followed a made basket and, as a result, the positional responsibilities were something akin to college lacrosse — with its faceoff specialists, attackers, and defenders. Eventually, rule changes phased out the non-stop tip-offs and allowed players to experiment with dribbling the ball. The result was further role specialization. For a long time thereafter, it seemed like building lineups to fit a five-position system was the way to go. There was a certain logic to it, after all. Five players on the court for each team, five positions. It made sense. But recently, the list of positions has begun to shrink again.

“I don’t have the five positions anymore,” Celtics coach Brad Stevens told the Associated Press. “It may be as simple as three positions now, where you’re either a ball-handler, a wing or a big.”

LeBron James is taking it a step further, promising that he and the Lakers will embrace “positionless basketball” this season.

With traditional positional designations falling by the wayside, some fear that the NBA is heading for a dystopian future. Michael Pina conjured the menacing threat of a league comprised entirely of “a homogeneous collection of positionless 3-and-D cogs”. Perhaps the league is tilting in that direction; but, for now, at least, there is still a rich diversity of individual offensive skill sets, approaches, and roles.

I used the NBA play-type data from last season to define 23 different offensive roles and I created a few Tableau dashboards to describe the groups that were formed. You can use these interactive charts to discover in which group your favorite player landed, how he’s similar to his groupmates, and how efficiently he converts his opportunities relative to other players who are tasked with the same role.

*****

The NBA’s individual play-type data are provided by Synergy using a proprietary, real-time, video-indexing, statistical engine that logs every play of every game. These stats show the number of possessions during which a player attempted to score as the pick-and-roll ball-handler, in isolation, with a spot-up shot, working around an off-ball screen, from a handoff, as the pick-and-roll roll man, off a cut, on a putback, on a post-up, in transition, or anything in between (the nebulous “miscellaneous” category). For more information about how specific types of plays are coded, check out this helpful guide.

I used hierarchical clustering (hclust in R) to form groups of players by offensive role as defined by their use of the eleven types of scoring. That is, I grouped players based on the ways they TRIED to score — not whether or not they actually COULD score. Using NBA.com jargon, I grouped individuals by play-type frequencies and ignored points-per-possession (PPP, and the corresponding percentile ranks). I included players who had a record of at least 250 attempted scoring possessions during the 2017-18 regular season (297 qualified players were used in the clustering analysis).

In the dashboard below you can toggle through the levels of the player clustering. At the top level, we find the three offensive position groupings mentioned by Coach Stevens: ball-handlers (26 percent of all players), wings (50 percent), and bigs (24 percent). You can see how the ball-handlers were defined by their heavy use of the pick-and-roll and isolation to generate scoring opportunities (shown in pink). In contrast, the wings got shots by spotting up or by working around off-ball screens and handoffs (green). The bigs found scoring chances as the roll man in pick-and-roll, off of cuts, on putbacks of offensive rebounds, or with post-ups (orange).

You can find an external link to this dashboard here.

Drill down further and you’ll see how the three broad position groups were split into eight player types by further dividing the wings (into four types: wings with handle, wings who work off the ball, spot-up wings, and tall wings) and the bigs (into three types: mobile bigs, skilled bigs, and bigs at the rim). Notice that if you click on the boxes in the legend you can change the view of the chart to show a single color (i.e., a single type of play) at a time. With this feature, you can examine how the eight player types differ within and between the three position groups. For example, the spot-up wings have larger spot-up frequencies than the other types of wings (i.e., they get a larger percentage of their shooting possessions via spot-ups).

Descend one more level down and you’ll find the complete list of 23 offensive roles: from the league’s domineering ball handlers, LeBron James, Russ Westbrook, and James Harden, to the bigs who subsist on a steady diet of rolls to the rim, like Clint Capela and Jarrett Allen.

These 23 offensive roles were defined exclusively by play-type frequencies (no other stats were incorporated in the grouping process). However, the next dashboard illustrates that the differences between the groups run a bit deeper than just the play-type stats. For example, you can see how the ball-handlers could also be characterized by their extensive time of possession; whereas certain wings (e.g., stationary spot-up wings) and bigs (e.g., bigs who cut to the rim) possess the ball for a very short amount of time each game. The highest 3-point rates (3PA per FGA) are concentrated among some of the wing roles (e.g., wings who work around screens). Not surprisingly, the highest paint rates are found among the bigs.

You can find an external link to the second dashboard here.

One big theme that emerges from the play-type data is the contrast between creators and their beneficiaries. The creative types of plays are pick-and-roll ball handling, isolation, and posting-up — these are the scoring opportunities which are initiated by the person with the ball and the ones that are most challenging to convert at an efficient rate. In contrast, the spot-ups, off-ball screens, handoffs, rolls, cuts, and putbacks are all dependent play types that tend to require somebody else to create an advantage and these are the types of plays that tend to be the most productive. To see proof of this dynamic, we can look at the percent of a player’s made field goals which were set up by an assist, something I’m calling the dependency rate, here.

To make a fair evaluation of a player’s offensive efficiency, you really need to account for the role they are being asked to play. As such, comparisons between players’ scoring rates are more meaningful if both players fill the same role for their teams.

*****

Now that we’ve defined our 23 offensive roles and learned something about the characteristics of each group, we can use the third dashboard to check out where each individual player is grouped and how his scoring rate compares to other players who have the same role.

The last dashboard is a little too big to embed, here, but you can follow this external link for a chance to hover, click, and filter the third chart and to interrogate the data at your leisure. For now, here’s a still image of the default view of the data with all 297 players shown:

And now, time for some trivia! Can you name the top two players in true shooting last season? (Listening…) Ok, you got Curry correct, but you definitely got the other guy WRONG.

The second-most-efficient scorer in the league last year was Anthony Tolliver (66.3 true shooting percentage). It’s a surprising tidbit, but you can use the dashboard to see how Tolliver did it. Click on that green dot floating above everybody else on the spot-up panel. That’s Tolliver with his 1.30 points per possession on spot-ups, a type of play which comprised a whopping 50 percent of his total chances. As a stationary spot-up wing, Tolliver wasn’t asked to initiate the offense in Detroit, but he did a great job of cashing in opportunities when his teammates created them for him.

Rudy Gobert also depended on his teammates to create opportunities for him and he likewise converted them at a very high rate (true shooting of 65.7 percent, third-best in the league). To find Gobert, you can hover over the group of orange dots at the right edge of the panel for cuts. His signature scoring play was the cut, for which he scored 1.30 points per possession and used 31 percent of his scoring opportunities. He rolled and boarded a lot, too.

Last season’s top-20 leaderboard for true shooting represents a wide variety of offensive roles. In addition to play finishers like Gobert and Tolliver, you’ll find the league MVP atop the list — demonstrating that players who initiate offense can be efficient, too (James Harden true shooting percentage of 61.9 was 18th in the NBA). Harden’s points-per-possession values are mostly unremarkable, except for in isolation, where he was super-human. He initiated 35 percent of his scoring possessions out of isos and notched 1.22 points per possession in these scenarios. Evidently, the Rockets strategy to force defensive switching and then attack any big defender who was unlucky enough to be stuck guarding Harden or Chris Paul has worked really well.

Despite being his team’s primary ball handler, Steph Curry has a pretty balanced mix of the creative types of plays like pick-and-roll (26 percent frequency) and isolation (7 percent) versus dependent types of plays like spots-ups (11 percent), off-ball screens (18 percent), handoffs (5 percent), and cuts (5 percent). As you probably guessed, he’s on the upper edge of the league’s points-per-possession range for a bunch of these different play-type categories.

Curry’s play-type distribution is a good reminder that the role labels given here — while not completely unrelated to individual skill sets — are highly dependent on the system in which the player is deployed. In other words, these labels are descriptive, but not necessarily predictive. In another system, a player might fill a different role.

For another example of a team’s system influencing individual stats — check out the bigs-who-cut-to-the-rim group, which was comprised almost entirely of Golden State centers. These guys were asked to fill a very limited offensive role and their play-type data reflects the coaching and personnel around them.

I hope you have fun exploring this dataset and that, like me, you come away believing that the NBA is still a vibrant collection of unique scoring talents and not some dreary factory churning out spot-up corner-3-shooting machines.

The above was an update of a similar clustering analysis that I did last year. If you want to see the clusters for the 2016-17 season, you can find them here.