Quantifying NBA Player Chemistry

Championship level NBA teams have been built in a multitude of ways. There have been Big 3’s like the Miami Heat, “Small Ball” lineups like the Golden State Warriors, and well-rounded teams like the San Antonio Spurs. Do some players have better chemistry than others? Personally, I thought that chemistry was not as valuable as people said, and for the most part, the optimal strategy was trying to get the best overall players on your team. As per usual, I dove into the numbers to see what I could find.

My approach was to group all similar players in the NBA. For example, I would have a group of 3-point shooters, a group of defensive big men, and a group of pass-first guards, etc. I then looked at all 5-player lineup combinations over the past 20 years to see which groups made up each lineup. By determining the performance level of a 5-group combination in a lineup, I could quantify the chemistry of that group mixture. And finally, I fit today’s NBA rosters into the best lineup combinations to see if teams were playing the optimal player personnel.

To group players, I used a machine learning technique called K-Means Clustering. In simple terms, once I decided on the number of groups/clusters, the algorithm classified each player into a group based on similar statistics. I chose to classify NBA players from 2000 to 2020 into 20 groups. A player was classified into a group based on a single season. For example, Lebron James would have 17 data points; one for each season. Here are the statistics I chose to use in the analysis.

Per 100 Possession Box Score Statistics (100 possessions is about 48 minutes or a full game):

3PA, 3P%, 2PA, 2P%, FTA, FT%, PF

Advanced Statistics:

FTr, ORB%, DRB%, AST%, STL%, BLK%, TOV%

Once I completed this process, I created group descriptions and displayed the mean statistics for each player type, as seen below. Five of the groups were full of outliers, so we are left with 15 clusters. I only included the basic box score statistics per 100 possessions in this chart to get an easier understanding of the type of player within each group.

I then collected all 5-player lineups since 2000 and each lineup per 100 possession statistics. There were about 50,000 lineups that played over 10 minutes since 2000 and the primary metric I used to judge the performance of a lineup was Net Rating. Net Rating is the estimated point difference per 100 possessions for a given lineup vs all opponents.

Now, this part may get confusing, but stay with me. Next, I corresponded each players’ group to the lineups they played on. So along with each lineup combination since 2000, there were also the corresponding five groups of players. For example, if a lineup were denoted by 20, 20, 15, 15, 15, it would mean this lineup had two Star Guards/Wings, and three Efficient Shooters.

I then used a weighted average to combine all lineup combinations that had the same five player types. An important consideration I made was how many total minutes were enough to accept the sample size of a lineup combination. Using trial and error and looking at the total minutes of NBA teams starting lineups, I removed all lineup combinations with under a total of 500 minutes played. I viewed any lineup with less than 500-minutes, as coaches acknowledging that the lineup was not an optimal grouping. Here are the resulting top 15 lineups.

An initial finding of the analysis was that there were 112 different lineup combinations with a Net Rating higher than five. There is a considerable amount of possible player type combinations that can yield success in the NBA, meaning there are several ways to assemble a great roster. A Net Rating of five is also substantial, as this season only six teams had a Net Rating greater than five.

The lineup that displayed the best chemistry consisted of a Star Guard/Wing, a Well-Rounded Forward, a Defensive First Center, a Starting Caliber Point Guard, and an Efficient Shooter. The majority of this lineups’ minutes belong to the 2008–09 Cleveland Cavaliers including:

Lebron James | Zydrunas Ilgauskas | Ben Wallace | Mo Williams | Delonte West.

Despite there being a common NBA narrative that LeBron played with “bad” teammates in Cleveland, they did build a team of player types with tremendous synergy.

This logic may be reverse causation because the lineup might just have been successful because of Lebron James. But there are other examples of weak teams that found success with this lineup. The 2016–17 Knicks who finished 31–51 found a lot of success in the few minutes they played this lineup.

Derrick Rose | Kristaps Porzingis | Joakim Noah | Brandon Jennings | Justin Holliday

This lineup had a Net Rating of 47.7 in 26 minutes of action. Of course, this production is not sustainable, but it’s interesting to see the lineup be successful for a weak team.

Another measure to consider is which player type had the most lineup versatility. The graph below displays how many times a player appeared in one of the top 100 lineups.

A Group 15 player (Efficient Shooter) appeared 82 times among the top 100 lineups, which was significantly more than any other group. This made sense, considering it is generally easy to play with this type of player. Many of the players in this group, like Jae Crowder, Klay Thomson, and Kyle Korver, act as floor spacers and don’t demand the ball on offense. This makes them a natural fit in any offensive scheme or with any player grouping.

I then fitted each teams’ roster to the best possible lineup. Overall most teams’ starting lineup similarly matched their best chemistry lineup. For example, based on the Houston Rockets’ current roster, the best lineup they can field is two, Star Guards/Wings, and three Efficient Shooters (20, 20, 15, 15, 15). This lineup combination ranked 23rd in terms of overall Net Rating and played in 172 games for 792 minutes. Houston can play 35 lineups with this five-player grouping including their likely best lineup:

James Harden | Russell Westbrook | Robert Covington | PJ Tucker | Eric Gordon

One team that stood out was the Orlando Magic, whose projected lineups were noticeably different than their starting lineup. There were 545 lineups that Orlando could field with a positive net rating, but none of them included Nikola Vucevic, Jonathan Isaac, and Aaron Gordon. These three players were classified in group 19 as Well-Rounded Starting Forwards.

Here are all the lineup combinations Orlando can field along with an example five players. Highlighted are the lineups that had two of Vucevic, Isaac, and Gordon.

(Net Rating is cumulative for all lineups with that cluster grouping)

As seen from the player examples and clusters, in the lineups with two of the three noted forwards, the other players are guards and wing shooters. Additionally, three of the highlighted lineups have a negative net rating. It likely isn’t optimal for Orlando to build their team around these three forwards whose player types haven’t had successful chemistry on the floor together in the past. Vucevic, Isaac, and Gordan are core pieces of the Orlando Magic and had a three-player Net Rating of -0.1 this season. Without elite guard play and three-point shooting, the Magic will continue to struggle with these three players in their lineup.

Do some players have better chemistry than others? I would argue, yes. With a substantial sample size, there is a clear difference in performance based on the type of players within each lineup.

Is chemistry important for a championship level team? Although there were 112 different lineup combinations with a Net Rating greater than 5, only 25 different lineup combinations had a Net Rating higher than 10. For a team to reach a championship caliber, I think chemistry plays a definite role.