Causal Structure Learning in Heavy-Tailed Models

• Activity Number: 237 - Graphical Models and Causality in Extremes
• Type: Topic Contributed
• Date/Time: Tuesday, August 9, 2022 : 8:30 AM to 10:20 AM
• Sponsor: Section on Risk Analysis
• Author(s): Nicola Gnecco* and Sebastian Engelke and Simon Chatelain and Stanislav Volgushev
• Companies: University of Geneva and Université de Genève and University of Geneva and University of Toronto
• Keywords: Causal discovery; Extreme value theory; Heavy-tailed distribution; Nonparametric estimation

Abstract

Learning causal structures from observational data is challenging. However, in applications such as hydrology, meteorology, or finance, one can exploit the heavy-tailed distribution of the data to recover causal information. This talk discusses a recent methodology that identifies and learns causal information from a linear structural causal model (SCM) with heavy-tailed noise (Gnecco et al., 2021). The first part will introduce the causal tail coefficient and the related EASE algorithm, which recovers a causal order of the underlying SCM. The second part will discuss recent results on the asymptotic distribution of the causal tail coefficient estimator for a broad spectrum of tail dependence settings.