Regularized Online GAMLSS

Florian Ziel, University of Duisburg-Essen

Co-authors: Jonathan Berrisch, University of Duisburg-Essen; Simon Hirsch, University of Duisburg Essen, Statkraft Trading GmbH

Abstract: Probabilistic online learning algorithms have become increasingly important for modeling data streams, particularly time series.

Distributional regression models estimated using the generalized additive models for location, scale and shape (GAMLSS) framework are a suitable tool for those tasks as the estimation is based on iteratively reweighted least squares. This enables the application of online learning algorithms to distributional regression models. In particular, online coordinate descent algorithms have been shown to be effective scalable solutions.

In this paper, we focus on regularization of those on-line estimation with regularization methods like lasso, ridge and elastic net.

We discuss the methodology for online estimation of GAMLSS models and its implementation in a python package.