Years ago Muthu Alagappan opened up the debate over how to categorize the true positions played by NBA players, arguing persuasively that our postional labels may not be as accurate or descriptive as we think they are.
Since then, many have furthered this work by making use primarily of unsupervised learning techniques. Some examples include:
Redefining Basketball Positions with Unsupervised Learning.
Redefining Positions and Players in Todays NBA
In this project I take a swing at the very same question. Data for this analysis will span the 2019/2020 season and all data from the current season through March 1st, 2021. Only those players with a minimum 23 games played during the 2019/2020 season will be included and only those with at least 15 games played in the current season will be included. This gives us 1,159 individual player seasons to work with.
All player data will come from basketball-reference.com. Because advanced stats such as PER take so many different aspects of play into their calculation, they tell us more about overall player value than about player type. For our purposes here, basic descriptive statistics such as points, steals, blocks, three-point shot attempts, two-point shot attempts, turnovers, assists, offensive rebounds and defensive rebounds are all that will be used. To contextualize these stats, each will be converted to a per minute played format. Additionally, we will include minutes per game played to provide player usage context. Next, Principal Component Analysis will be performed, allowing us to simplify our data inputs even further by reducing all this information to just four components, which will make the task of identifying player types much easier.
Full Project Coming Soon!
Data Augery 