[Paper Review] Estimation of extreme risk regions under multivariate regular variation

Juan‐Juan Cai, J.H.J. Einmahl+1 more|Research portal (Tilburg University)|Nov 22, 2012

Financial Risk and Volatility Modeling3 citations

TL;DR

This paper proposes a novel estimator for extreme risk regions in multivariate regularly varying distributions, using extreme value theory to extrapolate beyond observed data. It establishes consistency of the estimator under weak regular variation assumptions and demonstrates strong finite-sample performance via simulations and financial data applications, enabling reliable detection of rare, high-impact events in multivariate settings.

ABSTRACT

When considering d possibly dependent random variables, one is often interested in extreme risk regions, with very small probability p. We consider risk regions of the form ${\mathbf{z}\in\mathbb{R}^d:f(\mathbf{z})\leqβ}$, where f is the joint density and $β$ a small number. Estimation of such an extreme risk region is difficult since it contains hardly any or no data. Using extreme value theory, we construct a natural estimator of an extreme risk region and prove a refined form of consistency, given a random sample of multivariate regularly varying random vectors. In a detailed simulation and comparison study, the good performance of the procedure is demonstrated. We also apply our estimator to financial data.

Motivation & Objective

• To estimate extreme risk regions defined by low-density sets in multivariate regularly varying distributions where data are scarce or absent.
• To develop a nonparametric estimator that leverages the asymptotic behavior of the spectral measure and tail dependence structure.
• To ensure consistency of the estimator under weak regular variation assumptions, even when the number of observations in the extreme region is zero.
• To provide a practical tool for risk management and stress testing in finance and insurance by identifying rare, high-impact events.
• To extend univariate extreme quantile estimation to the multivariate setting by estimating entire level sets rather than single values.

Proposed method

• The method uses a nonparametric kernel-based estimator of the spectral measure derived from exceedances over a high threshold.
• It constructs an estimator of the extreme risk region $ Q = \{ \mathbf{z} \in \mathbb{R}^d : f(\mathbf{z}) \leq \beta \} $ by transforming the estimated spectral measure into a density-level set.
• The estimator relies on the generalized Pareto approximation of the tail behavior, using the peaks-over-threshold approach in the multivariate setting.
• It applies a normalization based on the $ L_2 $-norm and uses a bandwidth-dependent kernel to smooth the empirical spectral measure.
• The consistency of the estimator is proven via convergence in probability of the estimated spectral measure and the associated level set.
• The method accounts for the asymptotic scaling of the density via a regularly varying function $ V(t) = \mathbb{P}(\|\mathbf{X}\| > t) $, ensuring proper normalization of the density estimate.

Experimental results

Research questions

• RQ1 How can extreme risk regions be consistently estimated in multivariate regularly varying distributions when no data fall in the region of interest?
• RQ2 What is the asymptotic behavior of the estimated extreme level set under weak regular variation assumptions?
• RQ3 How does the proposed estimator compare to existing nonparametric density level set estimators in terms of consistency and finite-sample performance?
• RQ4 Can the method reliably detect rare events in high-dimensional financial data, even when the number of observations in the extreme region is zero?
• RQ5 What is the impact of bandwidth selection and kernel choice on the accuracy of the estimated extreme risk region?

Key findings

• The proposed estimator of the extreme risk region is consistent in probability, with the symmetric difference between the true and estimated region shrinking relative to the true probability mass $ p $, i.e., $ \frac{P(\widetilde{Q}_n \triangle \widehat{Q}_n)}{p} \stackrel{\mathbb{P}}{\rightarrow} 0 $.
• The estimator achieves consistency even when no data points fall in the extreme region, due to the use of extreme value theory for extrapolation.
• Simulation studies show strong finite-sample performance, with the estimated region closely approximating the true extreme set under various dependence structures.
• The method successfully identifies extreme risk regions in real financial data, demonstrating utility for stress testing and risk monitoring.
• The convergence of the estimated spectral measure and the associated level set is established under weak regular variation, with explicit control over the error via bandwidth and kernel parameters.
• The estimator’s performance is robust to moderate misspecification of the tail index and bandwidth, as shown by simulation robustness checks.