What’s the difficulty of this shot by Klay Thompson? Learn more with the new NBA statistic.
There are great shot takers, and there are great shot makers. How does a turnaround fadeaway jump-shot compare to a contested dunk? Is it more likely that the average NBA player will make an uncontested three-pointer or a contested floater? Automated tracking systems and probabilistic machine learning classification have created a more advanced measure of player agnostic Expected Field Goal Percentage than ever before.
The expected field goal percentage model learns the impact of defensive contest posture, shooter orientation and balance, and court location to determine the odds a shot is going in based on the exact situation of the shot. Keeping this player agnostic allows insight into which players are beating their expected field goal percentage, along with the features around decision making of every shot.
The foundation of any model is the data that goes into it. That’s where the NBA’s optical player tracking systems come in. Using advanced 3D pose detection models, these systems track 29 points on each player’s body, generating 3D coordinates for every movement. All of this is captured 60 times per second.
Expected Field Goal Percentage takes in the pose data and creates features relevant to a shooting situation to classify if a shot will go in or not. For each shot, the data looks from one second before the shot (60 frames) up to the release of the ball to calculate various features. The model does not account for situational features like shot clock, player identity, or score differential.
A machine learning model is trained on these feature sets and the true outcomes of the shots to predict the correct label – made or missed shot. There are two separate models: one for jump shots and one for finishes around the rim. Each model learns key features and relationships between features to estimate the probability of a made shot based on similar shots.
Feature engineering is integral for the interpretability and understanding of the Expected Field Goal Percentage metric. Let’s delve into what drives these shot probabilities. All features are calculated around time of release unless otherwise stated.
Definitions – Shooter Features
- Shot Distance: Distance from player’s center of mass to the hoop
- Shooter’s Vertical Tilt: Degrees of the shooter’s tilt with respect to the Z axis
- Head Angle: Angle of the shooter’s head relative to the basket
- Head Facing Percentage: Percentage of time the shooter’s head has potential visibility of the basket leading up to the shot
- Speed: Shooter’s speed at time of shot
- Velocity to Basket: Shooter’s velocity with respect to the direction of the basket. This is measured by how shot distance is changing over time
- Rotational Speed/Velocity: Direction and speed the shooter’s hips are rotating with respect to the basket
- Full Rotation: Total degrees of rotation the shooter rotates in the second before and after the shot
- Hang Time: The time spent during the air from the last time the shooter leaves the floor before the release up to release time
- Air Time Distance: Two-dimensional distance the shooter travels after last takeoff to release
Definition – Defender Features
- Closest Defender Joint Distance: Distance from the closest defender’s nearest hand and the ball
- Relative Height of Contest: The difference between the closest defender’s nearest hand height and the height of the ball
- Defensive Hand Contest Angle: Angle between the closest defender’s hand and the shooter’s potential apex
- Vision Interference: Quantity to which the defender is hindering the shooter’s view of the basket
- Elbow Contact Prior To Shoot: Distance between closest defender’s nearest hand to the shooter’s elbow in the 1 second leading up to the shot
- Defender Hip Distance: Distance between defender’s and shooter’s hips
- Defender Angle Relative to Basket: Angle between the closest defender and the shooter, and the shooter and the hoop
- Defender Speed/Velocity: Speed and direction which the defender is moving with respect to the shooter
- Pressure Score: Pressure exerted on shooter by matchup defender
- Average Pressure Score: Average exerted on shooter by all defenders on court
