Research paper and software

Least-Squares Shapley Performance Attribution

An efficient method and open-source implementation for attributing the out-of-sample performance of a least-squares regression model to its input features using Shapley values.

Current state

The research was published in Statistics and Computing in 2024. The public LS-SPA package implements the method, includes a worked notebook and tests, and reached version 2.0.0 in January 2026. A separate repository preserves the paper’s benchmark implementation and experiments.

Details

Phase
Published paper · software v2.0.0
Domain
Statistics · interpretable machine learning
My role
Coauthor and developer
With
Nikhil Devanathan, Stephen Boyd