Bayesian Spatial Modeling for Evaluating Hockey Goaltending Performance

This applied modeling workshop provides a practical walkthrough of Bayesian spatial modeling for NHL goaltending performance using PyMC. We’ll examine how to model shot outcomes while accounting for shot location, shot difficulty, and shot type within a principled Bayesian workflow designed for real-world sports analytics.

The session progresses through a sequence of increasingly expressive models, beginning with a baseline specification and extending to a full spatial model. We introduce a Gaussian Process (GP) to capture spatial structure in scoring probability across the rink, implemented with a scalable Hilbert Space Gaussian Process (HSGP) approximation. Along the way, we explore key modeling decisions, uncertainty quantification, and interpretation using prior and posterior simulation and predictive checks.

The workshop recreates and extends the approach from Chris Fonnesbeck’s PyMC Labs blog post, “Bayesian Spatial Modeling for Evaluating Hockey Goaltending Performance,” highlighting workflow, model construction, and translating spatial inference into actionable insight. The original post is available at: