Defining Types of Role Players and What We Can Learn From Them

facebooktwitterreddit

Credit: Jesse Blanchard | Rolling Stone

When I initially put together the “consistency” statistic hosted here at Nylon Calculus, and wrote my main post describing what it is and how it can be thought about[1. again, major hat tip to fellow writer Krishna Narsu for the idea and biggest push forward in methodology], I went through the work of it all with two major uses in mind: first, we spend so much time talking about consistency it would seem like an obvious thing to try and quantify, and second, it seemed like a really obvious tool to use to talk about and reconsider what we know about the roles players have on their various teams.

The first use has been taken care of: the data is public and can be used to talk about consistency at any time. This, though, is my best attempt to start using the data to do something new.

Consistency, as you may recall, is based on game-to-game variance. The idea is that when we talk about a player being consistent or inconsistent, we’re generally talking about how much their game-to-game performance varies; it’s a pretty obvious connection. The difference between consistency and variance, though, is basically that the consistency has been standardized, as a way of controlling for the fact that variance tends to get unilaterally larger as usage gets largerSimply, when we go from variance to consistency we’re controlling for expectations.

The upshot, then, is that variance actually works as a decent proxy for a player’s role, in a manner separate from raw usage, for a few reasons. First, as a player’s average in a stat (e.g. points, rebounds, etc) increases so does his variance for the most part; after all, the more opportunities a player gets to score, the more he’s going to vary in his scoring from game to game, it makes intuitive sense. Because variance increases as minutes, points, rebounds, etc all increase, but it also disproportionately increases with usage, it seems to reflect not just how many proportionate possessions their using, but how much of a role they play overall. Consistency, too, I thought when I started the analysis, could be a great way to differentiate players that are comfortable in their role, and players who are not. Players who are more consistent, or consistent within our expectations, are players who are doing what we expect from game to game, where players who are less so don’t necessarily do what we expect usually, which probably indicates a player who isn’t “fitting.”

To try and confirm this, I did a cluster analysis of the players and their relationship with variance and consistency[1. For the statistically inclined, I used a very roughshod K-means cluster for convenience and time, but I am well aware that that was probably not the best or neatest way to go about this. The results worked out pretty well, if someone would like to give it another go with a better clustering method, I would highly encourage it]. The results were interesting: the upshot is that variance turns out to be a great way to distinguish “effective role players,” though it isn’t great at separating out stars from non-stars (the Star players tend to get muddled in with hardcore gunners and quite a few low-minute outliers). In other words, we can use variance and consistency to tease out different types of role players in a way that hasn’t totally been looked at before.

The cluster analysis recognized three different groups of role players: Low-Variance-High-Consistency players (“Quiet Role Players”), Average-Variance-High-Consistency players (“Impact Role Players”), and Average-Variance-Low-Consistency players (“Uncomfortable Role Players”).

We can look at those clusters below, but keep in mind that variance and consistency are all about usage and role, and not about skill. You could even think about it as “coach usage;” variance seems to be as much about how the player is deployed on the court by the coach as it is about player performance. There’s an interesting relationship between skill and role, here, but I’ll talk about that later. At the moment, the important thing to remember about these roles is that the Quiet Role Players are simply players in a small, comfortable role. Impact Role Players are in a larger, more important role, and Uncomfortable Role Players are players in a larger role who are remarkably inconsistent and irregular in that role.




There’s a lot to try and unpack from this. Let’s get to it.

Are these players in their roles because they play with a given variance, or does the role cause them to vary as much as they do?

This is sort of a chicken and egg question, and it’s probably a little bit of both. For example, a lot of the league’s best rim protectors fall under “Quiet Role Players,” largely because they don’t do much on the ball and have very little usage. This is both a part of their role — they’re played as rim protectors and have most value doing little — and part what they’re comfortable with. Short of Bogut, most of those players probably don’t want to have to do anymore with the ball, or can’t do anything more with it if it were given to them. There’s both a limit set by the coach and organization, and one set by skill. The causation isn’t the most interesting part though: more interesting is what we can take out of the fact that different types role players can be delineated by consistency and variance. For one, it’s fascinating that we can delineate types of role players at all.

So, what exactly is it that we can learn about who’s in what role?

The short answer is: a lot. In all of the graphs demonstrating who has been in what kind of role this season, the dots are sized by a very simple SPM, or estimate of +/- based on boxscore metrics. The version above isn’t perfect — I was limited in what I could use to build it for technical reasons — but the message is clear: the players in the “Impact” Role Player cluster are far better on average than the players in the other clusters. There are plenty of outliers, to be sure, but that’s the case on average. In some ways that doesn’t matter: the players with the biggest role of the “role players” are generally better. Duh. But looking at the ways in which all these teams have been clustered tells us something about the value of consistency. In a previous analysis of consistency, there didn’t seem to be a relationship between consistency and performance overall, but when you reduce it to the role player clusters, that changes. After all, if the High-Consistency, Average-Variance group is the most effective, that tells us that the Role Players that are the most consistent tend to be much better than the role players that are not.

This is a relatively big deal. In that same prior study, I found that a player’s overall consistency (not variance) was only very loosely correlated with usage and with minutes consistency. In essence, for the whole league, a player’s role doesn’t determine his consistency, which doesn’t determine his overall effectiveness. But among role players, being able to be consistent is huge for that player’s ability to be effective. With the exception of rim protectors, most of the role players need enough of a role to make mistakes, it would seem, but once they’ve gotten that role, the key is to be able to perform consistently. The minutes from game-to-game, too, to that end, are less important than a player’s ability to play within expectations when called upon. This could be potentially huge for teams looking for effective role players.

Teams are on the market for a 3-and-D guy all the time, but how do you differentiate one who’ll perform well in a system from one who won’t perform well in a system? How do you anticipate finding the Lou Williamses of the world on the open market versus the Lance Stephensons? While it’s certainly not the final answer, it turns out that you can get a damn good start by looking at a player’s consistency[2. Some may be wondering why certain awesome, prototypical role players aren’t showing up on the list, i.e. Kyle Korver. It turns out that he ended up in the spectrum including stars and gunners, and hopefully, as I’ll discuss in a bit, this might get some of us to think a little differently about who’s really just “playing a role” and who’s more fundamental to the offense. As a hint: I think in a lot of these cases guys who we think of as “role players” are used in ways so structurally important to how a team plays that their role is actually far bigger than being simply secondary.].

Well that’s great…but why is Kevin Love just a role player? Or Klay Thompson? And Why is Zach Randolph an “Uncomfortable Role Player?”

I think, probably, the most interesting thing that this cluster analysis captures is the sense of the role in which the player is thrust by both the offense and the coach, as opposed to usage, which is largely “what does the player do when he has the ball.” In some ways, variance seems to capture, “how long does he have the ball,” and “how does he effect others who have the ball” since it increases linearly with more shooting possessions (more time of possession) and FG% (more gravity) as well as with usage. It’s not perfect, some players just have huge variance as a matter of how they personally implement their skill independent of role, but what seems to be clear is that the coach has to let a player vary a lot from game to game before they can get there.

What that might tell us about clear outliers like Love, Klay, and Randolph is that it’s possible that they’re in roles that we don’t typically ascribe to them. I think that this makes sense though. Everyone’s noted how odd it seems how uninvolved Kevin Love has been in Cleveland’s offense. His usage rate is still pretty darn high in Cleveland, but his time of possession is way down from what it was in Minnesota, and he’s spent far less time doing work from the elbow. In short, Kevin Love’s being used as a role player, and he’s been playing like a very, very good one, even with his struggles. Last season, he would not have been included in this cluster. In that sense, this seems to be revealing to an extent about players that are misused.

With someone like Klay, I think it’s revealing in a different way.  In the last few seasons, everyone commented how he was “just” a role player, who happened to also be an all-time great 3-point shooter. In this season, he’s expanded his toolbox massively and become a much more complete player. What the data seems to be indicating, though, is that he’s still being deployed similarly. He’s not being posted up, certainly, but he’s the second threat, the result of a play, less than the player itself. It just also happens that he might be the best player in the league in a position like that.

The case of Zach Randolph demonstrates more of the limits with an analysis like this, I think. It’s worth noting that the Uncomfortable Role Players cluster is awash with post players, and I really think that might be a function of the fact that strict post-up play is becoming more of an oddity in today’s NBA. But there’s no questioning that Randolph thrives there and is wildly successful, this might just be a case where he’s less consistent than most players, and that’s not a function of “discomfort” so much as it appears to be with Jonas or Lance Stephenson; it’s just how Randolph plays. Point being, as with a lot of other stats analysis, the apparent outliers might be the most interesting cases. What does it tell us that Rudy Gay is a high usage Role Player? Jrue Holiday, Eric Bledsoe, and Mike Conley, too? There’s a lot of fun analysis to be had here with what it is about the team structure that these players seem to be in the role they’re in.

Is there any other kind of clustering we can do?

I’m sure there is, but it’ll take someone a bit more creative than me to figure it out, I think. There are a few other obvious ones, though: Even without running a formal cluster, we can filter a player’s Points Per Game consistency versus a player’s efficiency consistency to figure out which players were most likely to “get theirs,” i.e. who made sure to get the same amount of points in every game versus irrespective of FG%, versus which players were just consistent shooters in both respect. The results were fairly entertaining. In particular, most of the players who are most consistent in both points per game and FG% are also bad at shooting…and then there’s Kyle Korver.



For all other kinds of clusters, though, it’ll take a subtler hand than mind to parse out what’s there and what isn’t.

In many ways, though, this is the kind of thing that can be done just by playing with our publicly available data. After all, that’s why we have it there in the first place.