Scouting NBA Three-Point Shooting

by Christopher Cheng

(Photo by Michael Reaves/Getty Images)

NBA teams have begun to rely more heavily on the 3-point shot, especially in the clutch — when efficiency matters most. As a result, the importance of drafting capable 3-point shooters has skyrocketed. However, scouts have disagreed on how to predict 3-point shooting success in the NBA. Some think a college player’s 3-point skills are more important, while others place more of an emphasis on free throw shooting ability. Information on the relative importance of these factors would be extremely helpful in making draft decisions. In particular, this could help unearth players like Kawhi Leonard, who were dismal 3-point shooters in college (25%), but still went on to be quality shooters in the NBA. With the NBA Draft fast approaching, teams must identify which prospects have the potential to excel at the next level.

To create our model predicting NBA 3-point shooting, we use college and pro shooting data for NBA players over the last 20 years. To isolate the predictive ability of college 3-point shooting and free throw shooting, we focused on shooting percentage and attempts per game, controlling for the number of college games played. For both 3-point and free throw shooting, we included a delta parameter, tracking the change in a player’s shooting percentage from his first to last college season. If a player was only in college for one year, they would have a delta of zero. Since much pre-draft debate concerns whether a player will be a serviceable 3-point shooter, we set out to predict what quartile a player’s 3-point shooting percentage would be in using only college 3-point shooting and free throwing shooting data.

We can use a decision tree model to predict the shooting success of college players. Decision trees are constructed by recursively evaluating different features to predict an outcome variable, which in this case is a quartile of the 3-point shooting distribution. We split our dataset into one dataset for training our model and one dataset to verify the accuracy. Our decision tree model with a max tree depth of 5 resulted in a 66% prediction accuracy on the training set and a 65% accuracy on our test set. The closeness in prediction accuracy indicates that our model is relatively generalizable for predicting the 3-point shooting success of college players that the model has not encountered before.

Now knowing that our model is able to predict the quartile of a player’s 3-point shooting percentage with about 65% accuracy, we can examine what variables were the most important indicators. Calculating the relative importance of each of our variables, we find that college free throw percentage was most important (34%), college 3-point percentage was second (22%), college 3-point percentage delta (12%) was third, and college free throw percentage delta (10%) was fourth. Interestingly, free throw percentage was more than 1.5 times as important as college 3-point percentage in predicting NBA 3-point success. The deltas (3-point and free throw) had about the same relative importance, perhaps signaling that shooting improvement in general was valuable instead of the specific shooting area.

Several times, the model predicted a player’s future 3-point success in spite of negative public perception. Consider Kawhi Leonard — his 25% 3-point shooting percentage in college would signal poor results in the NBA, but he has excelled in the pro game, shooting 38% over his career. While scouts labeled Kawhi’s shot as a concern, the model used his good free throw shooting (74%) to accurately predict that his 3-point shooting would be in the top quartile in the NBA.

Below is a sample of a few other players who had questionable college 3-point shooting for whom the model accurately predicted NBA 3-point success.

The model predicts with about the same accuracy for all types of players, but some of the biggest surprises came among the centers. In college, many taller players are discouraged from straying outside the paint. Therefore, some of the players in the above table, such as Karl-Anthony Towns and Channing Frye, rarely took threes in college. Their coaches wanted them to play down low, so scouts never got the opportunity to see their shooting potential. Nevertheless, the model didn’t need much information on these players’ long-range skills to determine their future success. The model emphasizes success at the line, so it knew that players who knocked down their free throws, like Towns and Frye, would be solid long-range shooters once they reached the NBA.

A potential explanation on the greater importance of free throw percentage is that free throws are standardized across players while there are varying types of 3-pointers (open, catch and shoot, off the dribble, step back). Therefore, free throw shooting could be more accurate in displaying a player’s true shooting ability while the 3-point shooting percentage may be more skewed by a player’s shot selection tendencies. Another potential explanation is that the NBA 3-point line is a longer distance than the college one. Previous HSAC analyses suggest that when the NBA moved the 3-point line back, certain teams and players got an advantage. Therefore, it may take different skills to shoot from different three-point distances. Free throws may be a better gauge of a player’s shooting touch and mechanics, which can be translated across all areas of the court.

In conclusion, we found that in predicting the quality of a 3-point shooter, the player’s college free throw percentage was more important than college 3-point percentage or attempts. For future work, one could compare the model to scouting projections to see where the model adds value to the existing perceptions.

Editor’s Note: If you have any questions for Christopher about this article, please feel free to reach out to him at christophercheng@college.harvard.edu

About the author

harvardsports

View all posts

2 Comments

  • Were those players listed in the table in the training set? If so, I’d be concerned about overfitting. It would be interesting to see the predictions on a test set.

Leave a Reply

Your email address will not be published. Required fields are marked *