Explaining Synergy’s Offensive Roles

Our Analytics and Insights squad has been working on a new feature to help coaches and scouts develop winning team strategies – learn all about how we are describing players’ offensive roles.

What position does Brandon Ingram play? If you’re trying to shoehorn him into the conventional five-category system you’d probably call him a small forward, I guess. That’s the position he’s assigned on Basketball-Reference, for example. On the official NBA site he’s just listed as a forward, same as his teammates Herb Jones and Trey Murphy. But neither of those labels tells us much about his role in the Pelicans offense or how it differs from the offensive roles played by Jones and Murphy.

The five traditional basketball positions tend to sort players roughly by their heights: from the point guard (traditionally the shortest) to the center (usually the tallest). But traditional position labels have become less relevant in today’s fluid, perimeter-oriented game where you might find a 6-foot-8 guy like Ingram playing the role of a ball handler for his team.

So, rather than rely on the old-school, height-based position labels, it would be helpful to have functional descriptions of how players play that decouple offensive role from defensive responsibility (and height). We suggest that offensive roles are defined by how a player tries to help his team score points, either through the shots that he takes himself or the ones he sets up for his teammates with his playmaking; whereas defensive responsibilities are defined by the amount of time a player spends guarding opponents of various offensive roles. So, for example, a player could be a playmaking ball handler (his offensive role) who mostly guards spot-up shooting wings (his defensive responsibility).

Sounds like a smart idea, right? But the devil – as they say – is in the details, so let’s talk about how exactly we are using Synergy data to come up with a description of each player’s offensive role.

We started with a thought experiment. In a perfect world – one where any conceivable datapoint could be within reach – what information would we use to define a player’s offensive role? We brainstormed within the Analytics and Insights squad, asked for input from other basketball folks from around the company and discussed the question with a wider audience before eventually deciding on a short list of role-defining features. Everything that followed was based on the idea that a player’s offensive role is defined by how often his team asks him to do each of the following tasks:

• Ball handling
• Creating advantages
• Exploiting defensive rotations, ie. “keeping the advantage”
• Finishing (at the rim or from three)
• Spacing, eg. via off-ball screen setting, cutting, and gravity, and
• Positioning, ie. standing and catching the ball in various places around the court.

These six role-defining features were the guideposts for our player sorting process helping us to decide, for example, how many groups we needed to create. Our goal was to represent the full complexity of the game without creating artificial distinctions between players that would not be meaningful to coaches or scouts in a practical way. We wanted to create offensive role groupings that separated players along these six dimensions without creating any extra groups. Potential sorting approaches (we tried lots of them!) that split up groups of players who had similar role-defining features just because they differed along other dimensions – such as height, playing time, or defensive responsibilities – were considered counterproductive.

Of course, the role-defining features outlined above are only abstract concepts, but many of those abstract concepts can be represented by simple stats that we can derive from familiar Synergy-logged data, like play types and shot types. We experimented with various ways of combining and aggregating play-type and shot-type rates and systematically evaluated how well each option represented our role-defining features (we can talk about how we used publicly available, center-of-mass, player-tracking data to validate our sorting process in a future post). This process helped us identify a set of role-defining rates (%TIME stats derived from play types and shot types) that describe how a player helps his team try to score.

Our sorting approach consists of a two-step process: first we create the framework for the offensive role groups, sorting players by the highest-priority role-defining stats and then we make final adjustments to this framework, smoothing out the rough edges of the initial sorting by incorporating additional information. In the first step, we use a rule-based decision tree to do the initial (rough) sorting of players based on the critical- and high-priority role-defining stats (the stats like ball handling rate and big-man-rate which are listed in the table above).

The players are first divided into three big classes – ball handlers, wings, and bigs – based on their ball handling- and big-man-rates. Each of those three classes is further divided into three or four offensive role groups based on its most relevant features. For example, a big man’s role could be defined as a playmaking-, stretch-, post-, or rim-finishing-big depending on his or her rates of playmaking, shooting, posting up, and rim-finishing. Any big that doesn’t qualify for any of the subgroups would slip through the branches of the sorting tree and land in the misfit bucket at the bottom (meaning that no offensive role label would initially be assigned to them). We find the best group for each of these tough-to-categorize “misfits” and smooth out some of the other rough edges of the initial sorting in the next step.

In the second step of our role sorting process, we incorporate all of the role-defining stats (including the moderate-priority ones from the table above) to see how closely each misfit resembles the average player from each offensive role group and then we determine which of these “centroids” the player is nearest (across the 14-dimensional space) and put them in that group.

Whew boy, that paragraph got a little technical!

Let’s take another crack at that explanation. I have a silly visual analogy that might help clarify the way each misfit is sorted into the most appropriate group for him. Imagine you could see the 14 role-defining features represented as a basketball-playing cartoon (in reality, the sorting process is *NOT* influenced by how a player looks; that is one of the ways we are making the groups more objective and the sorting process less prone to bias…but the cartoons help to illustrate the point, so please just play along!). In the analogy, ball handling rate is represented by basketball size, post-up rate is waistline width, movement rate is shoe size, etc., etc.

And in the second step of the sorting process, we look at all 14 role-defining stats and compare each misfit to the average player from each offensive role group. Whichever group offers the best look-a-like (taking into consideration everything from his metaphorical “socks” to his “hairline”) is the group where our misfit will end up. So, in the example above, during the first step of the sorting process, our misfit’s basketball is too small for him to join the ball handlers, his eyes are too little for him to be a playmaking wing, and his shoes aren’t big enough for him to run with the dynamic shooting wings which leaves him…unlabeled (at first!). However, in the second step of the sorting process, we see that – when we take into account all of his features – he looks more like the average spot-up shooting wing than anyone else, so we lump him into that group.

And, actually, we do the same process – of finding the closest look-a-like among the average player from each of the 11 offensive role groups – for each non-misfit player as well. If a player was successfully sorted into a group during the initial sorting done by the decision tree rules, we will tend not to change that player’s label in the second step of the sorting process. However, if the player has obviously landed in the wrong group, we will rescue him and move him to a more appropriate group. Typically, more than 80% of the players remain in their initial group as defined by the sorting tree and less than 20% are rescued to a different group during the second step of the sorting process. In this way, we are able to keep the framework of the roles consistent across all leagues and levels while smoothing out the rough edges of the initial sorting and dealing with misfits in a holistic way based on all of the role-defining stats.

Each league and level has a different standard for what defines a player’s offensive role. For example, most high school teams don’t run as much pick-and-roll as NBA teams do, does that mean there aren’t any playmaking ball handlers at the high school level?

Of course not!

And, because we want our offensive role sorting process to scale to all leagues and levels on the Synergy platform, we are choosing not to define the thresholds of the decision tree in absolute terms (by saying, for example, “ball handlers should have ball handling rates above 50%”). Instead, we are defining thresholds based on “z-scores” so that each player’s characteristics are compared to the distribution of characteristics in his own league (for example, “each ball handler should have a ball handling rate that is at least 1 standard deviation above the league average ball handling rate”). Using thresholds that are defined relative to league average role-defining stats gives us flexibility to use the same sorting approach in leagues (and historic seasons) with very different distributions of role-defining stats. The result is a more robust offensive role framework that will look somewhat consistent from league to league, ie. most leagues will have at least a few examples of players from each offensive role (although that will not always be the case for less ubiquitous archetypes like the playmaking bigs).

With all that being said, our hope is that the offensive role descriptions that we are producing will resonate with coaches and scouts on an intuitive level without much additional explanation needed. In the end, the proof will be in the pudding – so here’s a look at some examples of NBA players who are filling each offensive role this season.

Returning to Ingram, Jones, and Murphy, we see that – while all three are nominally small forwards who are basically the same height – they are playing different roles in the Pelicans offense this season. Ingram is a scoring ball handler, Jones is a slashing wing, and Murphy is a spot-up shooting wing. This gives us a much better idea of how each player has tried to help New Orleans score this season and provides us with a new way to describe these three players in addition to what we know about their position and height.

There are lots of other stats that we could have included in our sorting process but which we opted to *LEAVE OUT*. Here are some of the omissions which we considered most seriously for inclusion:

• Performance measures, like 3P%
• Offensive role(s) from previous seasons
• Height
• Age
• Minutes per game
• Usage rate (ie. total possessions involved/total team possession while on the court)
• Defensive responsibilities

We acknowledge that the above omissions (and other factors!) could be useful in sorting players, but we consider them supplemental to our definition of offensive roles (ie. they are factors that coaches and scouts should consider separately, in addition to role). For example, we are going to create complementary metrics to characterize a player’s defensive responsibilities by quantifying how often the player guards each offensive role. Our offensive roles describe how a player has tried to help his team score in his current context (ie. for his current team during the current season). They do not describe how well a player performs his role (player quality/player impact) nor are they meant to predict what a player’s role could or should be in the future (ie. his “potential”).

We’re hoping our new offensive role descriptions will help coaches prepare for and game plan against their opponents and help scouts to better understand the contexts in which prospects and recruits have played their basketball in the past and to imagine how those prospects and recruits might play differently in a new context. And we’ll have more to say about how, specifically, coaches and scouts might use these new metrics to develop winning team strategies in future posts, so check back here soon!



Todd is building tools to help coaches, scouts, and players find winning team strategies as part of Synergy’s Analytics and Insights Team. He creates inviting infographics, engaging charts and interactive displays that make data compelling and accessible. Follow him on Twitter @crumpledjumper


The Ups and Downs of Rookie Shot Making


Characteristics of each Synergy offensive role