Enhancing Multivariate Time Series Forecasting via Multi-Task Learning and Random Matrix Theory

Romain Ilbert, Malik Tiomoko, Cosme Louart, Vasilii Feofanov, Themis Palpanas, Ievgen Redko

Keywords: Multivariate Time Series Forecasting, Regularization, Multi-Task Learning, Random Matrix Theory

TL;DR: We propose a multi-task learning approach for multivariate time series forecasting that improves single-channel predictions by leveraging cross-channel information. Using random matrix theory, we provide theoretical insights and show empirical gains.

Abstract: We present a novel approach to multivariate time series forecasting by framing it as a multi-task learning problem. We propose an optimization strategy that enhances single-channel predictions by leveraging information across multiple channels. Our framework offers a closed-form solution for linear models and connects forecasting performance to key statistical properties using advanced analytical tools. Empirical results on both synthetic and real-world datasets demonstrate that integrating our method into training loss functions significantly improves univariate models by effectively utilizing multivariate data within a multi-task learning framework.

Submission Number: 20