Free-Throw Prediction from Motion Capture

1st place, SPLxUTSPAN 2026 Data Challenge — predicting where a free throw lands (angle, depth, left/right) from 69 joints of motion-capture data.

The problem

Over 1–2 seconds, a basketball player coordinates 69 tracked body joints to propel a ball toward the hoop. Given 240 frames of motion capture at 60fps — 207 features per frame — predict three scaled outcomes describing where the ball lands: its angle, depth, and left/right offset relative to the hoop. 345 training shots, 113 test shots, 5 players; scored on mean MSE across the three targets.

Approach

The core insight: every player shoots differently — not just in skill, but in mechanism. A global model averages over player-specific biomechanics and captures none of them. The final pipeline treats each player independently at every stage.

• Per-player locally weighted regression. Each prediction is built from the player's own data, weighted by biomechanical similarity via a Gaussian kernel — each test shot is explained primarily by the most similar training shots from that same player. (−13% error)
• Temporal commitment points. Different targets are decided at different moments: depth commits ~930ms before release (forward momentum fixed mid-jump), angle ~550ms before (elbow geometry locked), left/right only after release (final wrist snap). Features are extracted at the frame each outcome is actually decided.
• Kinetic chain + hand physics. Features track energy transfer from hips through trunk and shoulder to the fingertips — fingertip velocities, finger spread, wrist flexion and curl across all five fingers.
• Velocity CNN + Position CNN. Two 1D CNNs over the raw sequences — one on joint velocities, one on positions. They fail on different shots (predictions correlate at only r=0.63 for depth), which made the ensemble worth more than either alone.

Final leaderboard score: 0.006148 MSE, first place.

What I learned

Feature engineering isn't dead. The end-to-end models alone were good; the model with biomechanically-motivated features was better. Knowing something about the domain — how a shooting motion actually chains from legs to wrist — was worth more than extra layers.

Full writeup and code in the repo; there's also a presentation walkthrough.