Quantco/glum

What it solves

glum is a fast, Python-first library for Generalized Linear Models (GLM), providing a unified interface for common statistical methods like least-squares, Poisson, and logistic regression. It addresses the need for a broad range of distribution support, flexible regularization techniques, and high performance, particularly when the number of observations significantly exceeds the number of predictors (N >> K).

How it works

The library provides a scikit-learn-like API for fitting models. It supports various distributions (Normal, Poisson, Binomial, Gamma, Inverse Gaussian, Negative Binomial, and Tweedie) and customizable link functions. It integrates with formulaic for formula-based model specification and narwhals for compatibility with multiple dataframe backends like pandas and polars.

Who it’s for

Data scientists and statisticians who need a high-performance GLM implementation with support for L1, L2, and Elastic Net regularization, as well as classical statistical inference for unregularized models.

Highlights

• Broad Distribution Support: Includes Normal, Poisson, Binomial, Gamma, Inverse Gaussian, Negative Binomial, and Tweedie distributions.
• High Performance: Consistently faster than other modern libraries when N >> K.
• Flexible Regularization: Supports L1, L2 (including Tikhonov penalties), and Elastic Net regularization, with built-in cross-validation for optimal regularization paths.
• Formula-based Specification: Supports formulaic-based models with monotonic constraints.
• Backend Agnostic: Works with pandas, polars, and other dataframes via narwhals.