The best era of basketball analytics

Rui Qiu

Updated on 2021-12-04.

For this part of the portfolio, the goal is to take a look at the portfolio at a higher level and try to conclude the progress with what has been learned and what needs to be changed.

First start with some bads:

The portfolio began with raising ten questions out of a topic of interest that we plan to take a deep dive into. The issue is that there was no information about how the upcoming chapters would pan out. For example, the original questions included analyzing how to quantify a team’s offensive system, while no actual part of this portfolio left the space to conduct, such an analysis. Most parts contained several restrictions such as fitting in three different models and visualizing the results. So this brings out the second point: what has been learned.

Without a doubt, the portfolio serves well as a collection of showcasing various basic, both supervised and unsupervised, machine learning algorithms. Additionally, some experimental trials with visualization packages make the portfolio more vivid and interpretable. But even so, the whole procedure of how each model is built does not imitate the real-world machine learning or data science workflow. Each chapter is more or less like a showcase of what I can do with this sort of model, rather than sophisticatedly tuning the model towards a correct direction.

Still, there are some highlights which are worth mentioning:

1. The ambiguity of modern basketball players' positions was successfully captured in the clustering. Players tend to be more versatile than their predecessors 20 years ago. (Although a comparison would make this statement more sound.) Specifically, big players shoot more from perimeter, and tall guards with defensive ability guarding multiple positions are favored. The versatility, for sure, elevates the tactical possibilities.
2. The play-by-play data indeed works well to train a predictive model. It's a pity the in-game tracking now largely prefers result-based in-game events. That is to say, people care about if the shot leads to scoring or not. That also means lots of details like pick-and-roll, covering, box-out, hustle, etc. are omitted because they probably do not affect the team score directly. However, if the in-game events can be imagined as a series of consequential but more general events, these omitted details will play their parts. Back to the topic, the prediction based on a static in-game scenario, which is the exact moment when the shot is taken, accurately reflects the fact if the shot is made. Therefore, a team's analyst could provide some informative insights of their opponent teams ahead of every game.
3. The stats vary team by team, season by season. Even a player could evolve or recede in the span of his career. All the probability calculation are conditional. That's also the beauty of modern sports, everything is evolving all the time. By looking at a single player or a team, one can easily tell the rises and falls through his playing stats. And what's always fascinating is that there are always stories behind those numbers.

So what’s next? The completion of such a portfolio is never the end, as most of the questions remained unanswered. Modern basketball has never ceased to evolve, and the data scientists should always being craving a better approach to apprehend today’s games. As the in-game tracking data becomes more and more detailed and accessible at the same time, the recap of tactics won't be limited to a two-dimensional tactic board any more. Moreover, data will be utilized to more aspects of running a professional basketball team such as team development, player scouting, even the ticket price setting and other operational decision making. In short, this is the best era of basketball analytics.