Modeling extremal dependence in multivariate and spatial problems: a practical perspective

From environmental sciences to finance, there is a growing demand for methods that can assess the risks of extreme events beyond those observed in available data. Extrapolating extreme events beyond the range of the data is not obvious. Risk assessments are often further complicated by the need to account for multiple variables simultaneously. Extreme value theory provides important tools for the analysis of multivariate or spatial extreme events, but these are not easily accessible to professionals without appropriate expertise. This article provides a minimal background on multivariate and spatial extremes and gives simple yet thorough instructions on how to analyse them using the R package ExtremalDep. After briefly introducing the statistical methodologies, we focus on road testing the package’s toolbox through several real-world applications.

💡 Research Summary

The paper addresses the growing need across environmental sciences, finance, and other applied fields to assess the risk of extreme events that lie beyond the range of observed data. While univariate extreme‑value theory (EVT) is well‑established and supported by numerous R packages, the analysis of multivariate and spatial extremes remains challenging due to the complex, infinite‑dimensional nature of extremal dependence. The authors introduce the R package ExtremalDep, which provides a comprehensive toolbox for modelling, estimating, and exploiting extremal dependence in both multivariate and spatial contexts.

The methodological core rests on the representation of a multivariate extreme‑value distribution G as a combination of marginal Generalized Extreme‑Value (GEV) distributions and an extreme‑value copula C_EV. The dependence structure can be expressed through the stable‑tail dependence function L, the Pickands dependence function A, or equivalently the angular measure H on the unit simplex. The paper explains how, for large block sizes, the probability that at least one component exceeds a high threshold can be approximated by L(p₁,…,p_d) (Equation 2.9) and the probability that all components exceed simultaneously by the tail copula R(p₁,…,p_d) (Equation 2.10). These approximations enable direct estimation of joint exceedance probabilities for very small p, which is essential for risk management.

ExtremalDep distinguishes itself from existing packages (e.g., mev, mvPot, graphicalExtremes) by offering both parametric and non‑parametric/semiparametric inference. Parametric inference supports a range of extreme‑value copula families (logistic, asymmetric logistic, Hüsler‑Reiss, etc.) and uses maximum likelihood or Bayesian MCMC to estimate GEV marginal parameters and copula parameters jointly. Non‑parametric inference constructs an empirical angular measure from block‑wise maxima, then smooths it using Bernstein polynomials or splines to obtain a flexible estimate of the Pickands function. The package also provides Bayesian posterior sampling for the non‑parametric estimates, allowing uncertainty quantification.

A major contribution is the seamless integration of extremal dependence estimates into practical risk metrics. The toolbox includes functions to compute:

• Joint return levels (the level exceeded with probability 1/T);
• Conditional return periods (expected waiting time for a second variable to exceed given that a first one has);
• Conditional exceedance probabilities;
• Extreme quantile regions corresponding to a prescribed joint tail probability;
• Simulation of multivariate extreme vectors with the estimated dependence structure, facilitating validation and scenario analysis.

For spatial extremes, the package treats observations at locations s∈ℝ² as realizations of a max‑stable process. It estimates spatial Pickands functions or angular measures, enabling the generation of exceedance‑probability maps and spatial return‑level surfaces. The authors illustrate how to incorporate location information, perform kriging‑like interpolation of dependence, and visualise risk across a geographic domain.

Four real‑world applications demonstrate the workflow: (1) urban air‑pollution data (PM₂.₅, NO₂, O₃) where joint exceedance probabilities identify high‑risk pollution episodes; (2) nationwide precipitation and temperature records used to derive joint return levels for compound flood‑heatwave events; (3) exchange‑rate series where multivariate EVT informs stress‑testing of currency portfolios; and (4) a spatial analysis of extreme rainfall across a river basin, producing spatial return‑level maps. Each case follows a clear pipeline: data preprocessing (block maxima or threshold exceedances), marginal GEV fitting, dependence estimation (parametric or non‑parametric), computation of risk measures, and visualisation. Code snippets and plots are provided, making the methodology reproducible for practitioners with limited EVT expertise.

The paper concludes that ExtremalDep fills a critical gap by making sophisticated multivariate and spatial extreme‑value methods accessible, especially through its non‑parametric capabilities and direct risk‑assessment tools. Future research directions include extending the framework to higher dimensions (beyond ten variables), incorporating temporal dependence for spatio‑temporal extremes, and integrating the package into real‑time risk monitoring systems. Overall, the work offers both a solid theoretical foundation and a practical, user‑friendly implementation for modern extreme‑value analysis.