Neural Simulation-Based Inference for Multivariate and Spatial Extremes: When Model Flexibility Meets Computational Feasibility

This dissertation develops flexible statistical models for multivariate extremes together with efficient neural simulation-based inference methods that make such models usable in practice for reliable risk assessment for environmental hazards.

Overview

Reliable risk assessment for environmental hazards requires statistical models that accurately describe the joint behavior of several variables, from ordinary observations to rare extremes, and across space. In applications such as precipitation or sea-surface temperature, both unusually low and unusually high values may have important societal and economic consequences, so that the dependence structure in each joint tail must be represented. Although recent models provide substantially greater flexibility than classical extreme-value models, this flexibility comes at a substantial computational cost, since the mechanisms that provide it also render the likelihood analytically unavailable or prohibitively expensive to evaluate. This dissertation addresses these two complementary challenges by developing flexible statistical models for multivariate extremes together with efficient neural simulation-based inference methods that make such models usable in practice.

The dissertation contains two main methodological contributions. First, we develop a multivariate extension of the extended generalized Pareto distribution (eGPD), providing a unified framework for modeling low, moderate, and high values without threshold selection. The proposed model has eGPD-like margins, satisfies the limiting properties prescribed by multivariate extreme-value theory in both lower and upper tails, allows their dependence structures to be modeled separately, and remains straightforward to simulate from. Its likelihood is intractable, and we accordingly fit it using neural Bayes estimators (NBEs) and neural posterior estimators (NPEs). Second, we develop an amortized simulation-based Bayesian inference framework for a complex multivariate spatial extremes model by adapting the BayesFlow framework to replicated multivariate spatial data. The proposed summary network combines convolutional neural networks with permutation-invariant aggregation to accommodate variable numbers of replicated spatial fields, allowing a single trained neural network to provide fast posterior inference for arbitrarily many datasets generated from the same model.

The proposed methodologies are assessed through simulation studies and applied to two environmental datasets. The proposed multivariate eGPD model is illustrated using daily rainfall measurements from the Netherlands, where it provides satisfactory marginal and dependence fits across the full range of the distribution. The amortized Bayesian framework is applied to a flexible multivariate spatial extremes model for sea-surface temperature anomalies over the Arabian Sea, providing efficient posterior inference for the extremal dependence structure across many spatial subregions. Together, these developments illustrate how flexible extremal dependence modeling and amortized simulation-based inference jointly enable complex environmental risk analysis to be performed efficiently.

Presenters

Brief Biography

Noura Alotaibi is a Ph.D. student in Statistics at King Abdullah University of Science and Technology (KAUST), under the supervision of Prof. Raphaël G. Huser. Before joining KAUST, Noura obtained an M.S. degree in Statistics from Ohio University, United States.