Multivariate probabilistic time series forecasting: transformer-based copulas and the limitations of proper scoring rules

Abstract

Accurate estimation of time-varying quantities is crucial for effective decision-making across various domains, including healthcare and finance. The value of these estimates largely depends on their ability to quantify predictive uncertainty and capture interdependencies between relevant variables. In this talk, we will explore two recent contributions to the field of multivariate probabilistic forecasting. Firstly, we will introduce TACTiS, a flexible transformer-based method that effectively handles high-dimensional, unaligned, and non-uniformly sampled time series, delivering state-of-the-art forecasts. We will delve into the model’s architecture, particularly its attention-based decoder, which provably learns valid non-parametric copulas. Secondly, we will dive into the evaluation of forecasting models, focusing on the finite-sample performance of proper scoring rules for detecting errors in forecast distributions. We will present critical findings that challenge the current methods of assessing progress in the field, underscoring the need for more reliable evaluation procedures.

Alexandre Drouin

Staff Research Scientist

My research interests include machine learning, causal inference, and computational biology.