Multivariate generalized Pareto distributions along extreme directions

Anas Mourahib, Anna Kiriliouk, Johan Segers

TL;DR

This paper expands the toolkit for modeling multivariate risks by focusing on "extreme directions"—scenarios where only specific subsets of variables peak simultaneously while others remain at normal levels. The authors provide a rigorous mathematical framework for these distributions, a flexible "mixture model" construction that can represent any such extreme behavior, and a practical algorithm for simulating these complex risk scenarios.

What Problem Does This Paper Address?

In risk management (finance, hydrology, climate science), we often use the "peaks-over-threshold" method to model extreme events. Mathematically, these extremes are described by Multivariate Generalized Pareto Distributions (MGPDs).

However, most traditional MGPD models rely on a simplifying assumption: when things go wrong, they go wrong everywhere at once. In technical terms, they assume all variables are "asymptotically dependent." In the real world—especially in high-dimensional systems—this is rarely true. For example, a massive flood might hit three specific gauging stations in a river network while leaving others unaffected.

Existing literature lacks a comprehensive way to handle these "sparse" extremes where only certain groups of variables (called extreme directions) exhibit joint tail behavior. This paper fills that gap by characterizing these directions and providing the tools to build and simulate models that respect them.

Methods

The authors approach the problem through several key technical contributions:

  1. Characterization: They define extreme directions using the "angular measure" of max-stable distributions. They prove that an extreme direction exists if there is a non-zero probability of a specific subset of variables being large while the rest are small.
  2. Density Derivation: Calculating the probability density of an MGPD is tricky because the "support" of the distribution is spread across different lower-dimensional subspaces. The authors derive explicit density formulas with respect to a "dominating measure" (essentially a sum of Lebesgue measures across all possible subsets of variables).
  3. The Mixture Model: They introduce a stochastic construction called a smoothed max-linear model. While standard max-linear models assume perfect dependence within factors, this "smoothed" version allows for more flexible, realistic dependency structures within each group of variables.
  4. Simulation Algorithm: They provide a generic rejection sampling algorithm (Algorithm 1) to generate random vectors from these distributions. This is crucial for "stress testing" systems where you need to simulate thousands of "what-if" extreme scenarios.

Datasets

As a theoretical statistics paper, this work does not use real-world empirical datasets. Instead, it uses synthetic datasets generated via their proposed simulation algorithm to validate the theory.

  • The authors specifically implement and test their model using extensions of the Logistic and Hüsler–Reiss families.
  • They provide a simulation example with a 3-dimensional system ($d=3$) to demonstrate how points concentrate on specific "sub-faces" of the risk space (see Figure 2 and Table 1).

Key Insights and Ideas

  • Extreme Directions as Signatures: The paper treats the set of variables that can be large together as a "signature" of the model. By choosing which signatures to include in their mixture model, researchers can "program" exactly which groups of assets or sensors are allowed to fail together.
  • Generality: A major "aha moment" is Theorem 4.8, which proves that their proposed mixture model construction is universal—it can represent any possible MGPD, no matter how complex its extreme directions are.
  • The Generator Link: The authors bridge the gap between different mathematical representations of extremes (spectral vectors $S$, $T$-generators, and $U$-generators), showing how to move between them to make density calculations and simulations easier.

Key Results

  • Mathematical Equivalence: Theorem 3.6 provides eight equivalent ways to define an extreme direction, providing a robust theoretical foundation for the field.
  • Simulation Accuracy: In numerical tests (Table 1), the empirical proportions of simulated extremes perfectly matched the theoretical probabilities for both the mixture logistic and mixture Hüsler–Reiss models. For instance, in a test case, the model predicted a specific group would be extreme 55.5% of the time, and the simulation produced 55.9%.
  • Density Formulas: The paper provides the first explicit density expressions for MGPDs that account for multiple, arbitrary extreme directions (Propositions 4.15 and 4.18).

Limitations and Future Work

  • Statistical Inference: The paper focuses on construction (how to build the model) and simulation (how to generate data from it). It does not yet provide a full framework for inference—the process of estimating the model's parameters from real-world data.
  • Model Selection: While the paper shows how to build a model with specific extreme directions, it doesn't solve the problem of how to discover which directions exist in a raw dataset (though it cites other works that attempt this).
  • High-Dimensional Sparsity: While the framework handles high dimensions better than previous models, the number of possible extreme directions grows exponentially with the number of variables. Incorporating "sparsity constraints" to keep the models manageable in very high dimensions remains an open challenge.

Summary updated

Abstract

When modeling a vector of risk variables, extreme scenarios are often of special interest. The peaks-over-thresholds method hinges on the notion that, asymptotically, the excesses over a vector of high thresholds follow a multivariate generalized Pareto distribution. However, existing literature has primarily concentrated on the setting when all risk variables are always large simultaneously. In reality, this assumption is often not met, especially in high dimensions. In response to this limitation, we study scenarios where distinct groups of risk variables may exhibit joint extremes while others do not. These discernible groups are derived from the angular measure inherent in the corresponding max-stable distribution, whence the term extreme direction. We explore such extreme directions within the framework of multivariate generalized Pareto distributions, with a focus on their probability density functions in relation to an appropriate dominating measure. Furthermore, we provide a stochastic construction that allows any prespecified set of risk groups to constitute the distribution's extreme directions. This construction takes the form of a smoothed max-linear model and accommodates the full spectrum of conceivable max-stable dependence structures. Additionally, we introduce a generic simulation algorithm tailored for multivariate generalized Pareto distributions, offering specific implementations for extensions of the logistic and Hüsler-Reiss families capable of carrying arbitrary extreme directions.
BibTeX
@inproceedings{mourahib2024multivariate,
  author = {Anas Mourahib and Anna Kiriliouk and Johan Segers},
  title = {{Multivariate generalized Pareto distributions along extreme directions}},
  booktitle = {Extremes},
  year = {2024},
  doi = {10.1007/s10687-024-00501-4},
}

Loading related content...

Figure