Finding NBA Players Deserving of More Minutes

Are players outplaying their role?

Every basketball loving fan at some point in their life will have the discussion of which player is better. As a child, my friends and I would often argue over Jason Kidd or Steve Nash, D-Wade or Lebron, or Kobe or Allen Iverson (AI). To this day, I have friends who say the 2001-2002 season AI was a better scorer than Kobe and I continue to respond with it wasn’t as close as you think. AI scored 31.4 points per game in the 01-02 season and Kobe scored 25.2 points per a game. Six points is a significant difference in points per game but once you take into account that AI played 43.7 minutes per game and Kobe played 38.3 minutes, simple math tells you that Kobe scored .66 points per minute and AI scored .71 points per minute. That means if both played forty minute games at their season scoring rate Kobe would have 26.4 points and AI would have 28.4 points. A measly two-point difference. Add in the fact that Kobe shot 46.9 percent from the floor and AI shot 39.8 percent from the floor, the statement becomes said with a little less confidence. While we could argue over who was a better or more efficient scorer they are both all time greats, any team would’ve loved to have them on their team.

Instead I focused on determining players who were underutilized in the 2018-19 season with clustering. Once identified, I analyzed how certain underutilized players performed in the 2019-2020 season compared to the previous season. In this experiment I solely looked at offensive ability as defensive statistics are not very good at completely capturing defensive ability.

The Method

Typically, for offensive evaluation I would look at efficiency and scoring ability using effective field goal percentage, true shooting percentage, or more advanced metrics but kept it simple for visualizations by only using three dimensions: points, assists, and rebounds; the classic triple double. I kept the statistics simple in hopes that someone without significant statistical background would have less difficulty interpreting the results.

The data set came from basketballreference, I used player’s per game and per 36 minute tables found here [https://www.basketball-reference.com/leagues/NBA_2020_per_minute.html]. Now that we know the setup, here’s how my clustering unsupervised learning works.

NOTE: Analysis up until break of 2020 NBA Break, with update to data analysis will not match numbers for 2020 exactly.

1) I began by looking at the per-36-minute data to search for outliers. Players with a small sample of minutes could have a large volume of scoring. For example, scoring a two-point basket in a minute would skew the per-36-minute numbers to score seventy-two points per-36-minutes. To deal with this I filtered for players whose minutes per game were in the top ninety percentile or players who played in more than ten games and played for more than two standard deviations below the mean number of minutes played per game. This controls for players who may have played large minutes in a few games but were injured after or players who were just barely on rosters.

2) I then ran a cluster analysis of offensive ability using both 2018-19 and 2019-2020 seasons to create the clusters. More than three variables can be used in creating clusters but, in this situation I only used three to allow for 2-D and 3-D interpretable graphs that coaches can use to take action. Again, I kept it simple and clustered using PTS per game, AST, and TRB. I standardized AST, PTS, and TRB per game with rescaling (min-max normalization) by subtracting the minimum and then dividing each value by the range. After I fit a hierarchical agglomerative cluster analysis (HCA) model to determine the number of clusters recommended by critical values on per_game data basis. Note however, there was no wrong answer for numbers of cluster as this was an exploratory analysis. Once I determined the number of clusters I assigned the players to a cluster. I used HCA over k-means because k-means requires me to know the number of clusters I desire beforehand and HCA makes less assumptions about data shape.

3) I applied the player’s assigned per-game cluster to the players per-36-minute data. By doing this I was able to analyze which players when remaining constant with their per game pace would be considered much better players compared to their counterparts when given the same minute’s opportunity.

4) Lastly, I analyzed if a player changed clusters year over year.

I started by determining the number of cluster to use. HCA had seven indexes suggest two clusters and six suggest four clusters, all other number of cluster only received three or less index votes. If you would like the code to see the output, email me! I fit the model with four clusters as it was the most practical when it came to the NBA. With four clusters I hoped to see an elite or superstar player, starters, a bench player cluster, and reserves.

The graph above shows a trend I hoped for. The blue represents elite offensive players, the red and green represents potentially starter or role players, the deep green represents bench players who come in for a few points a game, and the purple are deep reserves. The guard and forward groups do not necessarily have to be guards or forwards but the clusters are splitting by rebounds which are commonly higher in forwards and centers. Immediately, I saw a major flaw in the clusters, assists are barely taken into account when splitting the clusters. It mostly looked like the clusters were being split by points. However, notice the overlap? The overlap was due to rebounds, some players with less points were in the “higher” group due to their rebound value.

However, what I was more interested in was if I should be classifying players differently based on their per-36-minute statistics. The below plot assigns the per game statistic cluster to a player on their per-36-minute numbers.

A simple glance shows there are some players who on a per minute basis were better than their per game statistics show. I do acknowledge the large variation of reserve players; with small minutes, usually when games have already been won or lost, they score quickly as defense lacks. If the graph is too crowded, it is interactive, simply click on the reserve group (or any group for that matter) to remove them. Double click on one cluster if you would like only that cluster to be shown. Hover over a mark to see the details of that player such as the season, PTS, AST, and TRB.

Overall, I did find some players who were more similar than many think; their statistics can be seen in the tables below. In 2018-19 Victor Oladipo only averaged two more points, half an assist and 3 rebounds more than 2018-19 Spencer Dinwiddie. But, give Dinwiddie the three extra minutes Oladipo is playing they would be almost identical in points and assists, yet Oladipo was an all star reserve that year. Looking at 2020 and given the minutes Dinwiddie improved on his efficiency jumping from 21.5 to 23.8 points per 36 minutes.

Also shown below is Dante Exum in 2019, a candidate to deserve more minutes. He is classified as a bench player per game but given more minutes could be a starter/role. Lastly, a player I believe is deserving of more minutes is Michael Porter Jr. Playing just 14 minutes per game in his rookie year in the 2019-2020 season he is averaging more than .5 points per minute could score close to twenty a game if given the opportunity.

These players are just a few of the examples I picked out, because I like those player’s game, but there are many other players that have similar situations that can be seen from the graphs.

Per Game data

Per 36 Minute

Now, would Dante Exum actually score 15.6 points if he played 36 minutes. Most likely no, it is hard to sustain scoring ability over time. But he could possibly contribute more than a coach might initially realize if he was given more time.

My visualizations made it difficult to see the breaks for rebounds, but made the clusters easily digestible in 2-D. Here’s an interactive 3-D visualization which encapsulates all variables used for clustering.

Lastly, I wanted to see which players changed clusters year over year. By doing this I was able to easily find a list of players who have improved year over year and could use this list to see what the reasons were.

Players who went from a Starter/Role player to Elite player in 2020.

Bench or Reserve to Elite Players:

Only Devonte’ Grahm of the Hornets and Norman Powell went from a reserve or bench player to an Elite player. Jumping 14 and 8 points respectively a game does help!

Bench to Role/Starter Players:

Reserve players to Starter/Role players: