The DRAY: Assessing Defensive Versatility in the NBA

Jan 28, 2015; Atlanta, GA, USA; Brooklyn Nets forward Joe Johnson (7) passes out of the defense by Atlanta Hawks forward Paul Millsap (4) in the third quarter of their game at Philips Arena. The Hawks won 113-102. Mandatory Credit: Jason Getz-USA TODAY Sports
Jan 28, 2015; Atlanta, GA, USA; Brooklyn Nets forward Joe Johnson (7) passes out of the defense by Atlanta Hawks forward Paul Millsap (4) in the third quarter of their game at Philips Arena. The Hawks won 113-102. Mandatory Credit: Jason Getz-USA TODAY Sports /
facebooktwitterreddit

Last week, Senthil Natarajanand Chris Pickard both had the opportunity to compete in the inaugural NBA Hackathon event hosted by the NBA at Terminal 23 in New York City. For the Nylon Calculus group, this event also marked the debut of Team Nylon who competed against over 200 students representing over 50 different universities across the country. Team Nylon walked away with a second place finish, and they wanted to share their project with the general public. Here’s what they worked on during those eight hours…

Think about the hallmarks of Golden State’s defense, especially in the Death Lineup. The manic switching, the aggressiveness, and ability to apply pressure from any area of the floor. It’s all keyed by the presence of Draymond Green, one of the most unique players in the NBA. In an era where offensive personalities are becoming more specialized and commonly described for their versatility such as stretch- and point-forwards, a defender’s ability to guard a variety of offensive skills has become a premium in the NBA. While rim-protectors in Hassan Whiteside and pesky guards in Avery Bradley can dominate certain aspects of defense, versatile defenders can have a more holistic impact on a team’s defense that can often help overcome lineup deficiencies.

More from Nylon Calculus

So when one of the three prompts at the NBA Hackathon asked to develop a new method or tool for the evaluation of defensive performance in the NBA, Team Nylon set out to measure just that – defensive versatility. Let’s have a little fun and, for the sake of its inspiration, call it a player’s Defensive Range AdaptabilitY score or DRAY. While offensive styles are typically categorized by traditional positions, the reality on the court is much different. A defender’s responsibility guarding Channing Frye is going to be much different than guarding Tyson Chandler despite both players being categorized as centers. So when we tackled this problem, we divided the framework into two underlying components:

  1. Offensive players should be categorized by their true offensive positions based on scoring habits.
  2. Defensive players should be evaluated based on their ability to minimize the impact of scoring plays.

While players earn scores across each offensive playing style, their complete DRAY score is the weighted total across each of these styles. The basic premise is that players that consistently defend multiple playing styles will be rewarded while separation amongst similar players is based on their relative performance against these styles.

Now, it is important to clarify before we jump into the rest of our hack that defense is one of the more difficult aspects of basketball to quantify. A more advanced, but similar approach to our method (not known at the time of the hackathon) by Luke Borne, Kirk Goldsberry, Alexander Franks and Andrew Miller has been worked on. Given our constraints at the event, our definition for defensive performance is specific to shot events and, as a result, other important elements of good defense such as steals, tight coverage, etc. slip through our score. Given our framework, a player with great face-guarding skills like Bruce Bowen who forces ball movement away from his man will not be rewarded despite playing good defense. However, with an 8-hour time limit, this was the framework we developed.

Phase 1: Identifying Offensive Scoring Styles

Positional flexibility means that NBA players are no longer constricted to archaic definitions of the five traditional positions. Guarding a player like Channing Frye, a center who shoots like a guard, is a much different challenge than guarding a player like DeAndre Jordan, a center that stays close to the basket. For this reason, it is crucial to first establish “true” scoring styles, and classify each player according to their scoring style rather than position. The question we are trying to answer is what positions does each player actually score like, rather than what position appears next to them on the depth chart?

While in our report we described in more technical detail the methodology of this determination, the Reader’s Digest version of our methodology was divided into two segments:

  • Each player was categorized by our pre-selected list of offensive features that we felt described a player’s offensive tendencies based on the data that was provided. There were 19 features, standardized to a per minute basis that included:
    • Total shot attempts from 0-5, 5-10, 10-15, 15-20, 20+ distances (in feet).
    • Catch-and-shoot attempts from 10-15, 15-20, 20+ and 3PT distances (in feet).
    • Shot attempts after 0, 1-2, 3-4 and 5+ dribbles.
    • Total drives, free-throw attempts from drives, field-goal attempts from drives
    • Shot attempts from the post, paint and elbow.
  • Based on a player’s feature set, a Gaussian Mixture Model clustering algorithm was used to classify the players. Using each player’s list of features, the model simply tries to maximize the likelihood that a player would be in a certain category without any other bias. As a result, we returned five different types of offensive players shown below.
scorerstyles
scorerstyles /

Examples of Low-Usage Perimeter players include: Pablo Prigioni and Mike Dunleavy. Examples of Versatile Wings include: LeBron James and Gordon Hayward. Examples of Perimeter Specialists include: Kyle Korver and Channing Frye. Examples of Close to the Basket players include: DeAndre Jordan and Bismack Biyombo. Examples of Skilled Offensive Bigs include: Anthony Davis and Nikola Vucevic.

Phase 2: Evaluating Defensive Performance

After determining the scoring styles of each player, the next step is to evaluate how well a player defends against those various scoring styles. In order to accomplish this, we used a differential point value method stratified by the distance of the shot attempt from the basket.

    1. We first found a defender’s expected point value against each offensive playing style across 6 shot location ranges (0-5, 6-10, 11-15, 16-20, 20+ and 3PT). EPV was simply defined as the average points result of the shot minus the expected points of the shot; the more negative, the better the defender performed. The total offensive score was compared to the actual average expected points within each shot range and then weighted by the frequency that the player was marked as the nearest defender for that offensive style in each range.
    2. The raw score is just a number and although it based on points, it doesn’t have a lot of obvious translation to the scoreboard during a game. As Dean Oliver makes a point of in Basketball on Paper, understanding the average is important if we want to understand what is truly great and bad. As a result, the final defensive score for a given offensive style was the difference between the raw and the average scores.
    3. Total versatility was dependent on the number of opportunities with the total score measured as the accumulation of each defended offensive style score scaled by the total number of times that the player was the nearest defender. We are trying to measure, in some sense, how effectively a player defends multiple styles. The sophistication of this last step was certainly impacted by time restraints.

Results and Analysis

The first question you ask yourself after a model is do the results make sense? Fortunately, our results produced some very expected outcomes. While you can explore the full extent of the results in our R-Shiny web-application here, the top-5 by offensive playing style and overall defensive performance are listed below.  

final-results-table
final-results-table /

Our most versatile defender is Draymond Green, which is really a sanity check result to make sure that the model is functioning as intended. Green is widely perceived as the most versatile defender in the NBA, and that versatility is the key to Golden State’s defensive excellence. What’s important is that general trends exist and they tell a story that we would expect from watching any basketball game. Big men are best at guarding other big men while wings and guards are best at guarding other wings and guards, aligning with expected defensive assignments. One issue that is apparent in our results is that the bottom defenders are mostly all low-usage players, which makes sense given the framework we implemented, but more importantly points to an area our approach needs to better address.

One prominent storyline from the past season’s playoffs was the noticeable effect that Oklahoma City’s length had on the Warriors offense. Interestingly enough, three of the top 5 DRAY scores last year are from Oklahoma City — Durant, Ibaka and Adams. This begins to explain why the Thunder were so effective playing big last season. With frontcourt players that could guard multiple positions giving them to flexibility to hedge screens that the Warriors often rely on to create confusion, mismatches and open shots.

One of the main advantages of the smaller lineup is the use of speed to overcome size, but for Billy Donovan, he could sacrifice little in speed and maintain size on the court that not only made it more difficult for the Warriors on offense, but also created favorable matchups on the offense side of the ball. In the WCF, OKC’s switch ability showed up time and time again with players like Ibaka and Adams playing huge roles in their ability to switch and slow down the Warriors’ offense.

However, there were also a few surprises which provide excellent case studies in understanding how a player is used. James Harden showed up as the best defender against perimeter specialists, which runs counter to popular theory about his defensive ability. In terms of defensive schemes, the Rockets often moved Harden to guard off-the-ball perimeter specialists (saving him for offense) and, as a result, will start most defensive possessions in the corners, so it’s not too surprising to see him show up in this category. It should also be noted that many of Harden’s primary defensive matchups result with the offensive player blowing by Harden, alternating the identity of the nearest defender away from Harden, which would reduce the number of times he “guarded” a certain style of player.

Further Work and Limitations

Of the finalists, we were one of three other teams that looked at defense. The purpose of our hack to was to set up a framework that looked at defensive versatility based on the type of offensive skill tendencies the defender is having to defend. While this data used in this model classifies the nearest defenders and their distance to shooters for each shot attempt, it does not take into account whether the defender was actually the primary defender on the shooter.

For example, image a pick and pop between a screener and a three-point shooter that leaves the three-point shooter separated from his man. The nearest defender closing out on the shot is most likely not the actual primary defender, therefore the model’s results don’t necessarily hold each defender responsible for their man and measure true on-ball defense and, so, naturally James Harden would benefit from this shortcoming of the model. This issues could be addressed in a further expansion of this work by using raw SportVU data that marks the true defender for the given shot attempt. This expansion would provide more accurate defensive skill values produced by the model with respect to made shots.

Of course, contesting shot attempts are only one facet of defense. One thing the model fails to account for is anything that could prevent a shot attempt from occurring. Strong face-guarding skills or steals to name a couple are all consider good defensive qualities, but also do not result in a shot event. So a player like Avery Bradley doesn’t look as good in this model as maybe he really is while on the court.

Next: Yao Ming and Hall of Fame comparisons

Other smaller things that open the door to further improvement of our framework is using shot coordinates (x, y) to provide greater depth to the offensive playing style characteristics. For example, within perimeter specialists there could be two groups of shooters because we didn’t consider the difference between a corner and a wing three-point attempt. In addition, our method for calculating versatility wasn’t sophisticated and was volume based. Improving this area along could provide the biggest benefit to improving the model.

At the end of the day, there are certainly some holes within our framework that could be better addressed with more time and data, but the model does lay some groundwork for the ability to measure one of the key characteristics of defense in the modern NBA game – versatility.

The herein basketball analytics tool is the exclusive intellectual property of NBA Properties, Inc. and may not be used, in whole or in part, without the prior written consent of NBA Properties, Inc. © 2016 NBA Properties, Inc. All Rights Reserved.” 

Data provided courtesy of STATS LLC and NBA Properties, Inc.  All Rights Reserved.