Threshold Selection for Multivariate Heavy-Tailed Data

TL;DR

This paper introduces a statistically rigorous method for selecting thresholds in multivariate heavy-tailed data analysis by testing the independence between radial and angular components using distance covariance, improving model accuracy.

Contribution

It proposes a novel threshold selection algorithm based on independence testing with distance covariance, including a subsampling scheme for dependent data and reduced computational complexity.

Findings

Effective threshold selection demonstrated on simulated data.

Method successfully applied to real-world heavy-tailed datasets.

Improved accuracy in modeling tail dependence.

Abstract

Regular variation is often used as the starting point for modeling multivariate heavy-tailed data. A random vector is regularly varying if and only if its radial part $R$ is regularly varying and is asymptotically independent of the angular part $\Theta$ as $R$ goes to infinity. The conditional limiting distribution of $\Theta$ given $R$ is large characterizes the tail dependence of the random vector and hence its estimation is the primary goal of applications. A typical strategy is to look at the angular components of the data for which the radial parts exceed some threshold. While a large class of methods has been proposed to model the angular distribution from these exceedances, the choice of threshold has been scarcely discussed in the literature. In this paper, we describe a procedure for choosing the threshold by formally testing the independence of $R$ and $\Theta$ using a measure of dependence called distance covariance. We generalize the limit theorem for distance covariance to our unique setting and propose an algorithm which selects the threshold for $R$. This algorithm incorporates a subsampling scheme that is also applicable to weakly dependent data. Moreover, it avoids the heavy computation in the calculation of the distance covariance, a typical limitation for this measure. The performance of our method is illustrated on both simulated and real data.