A Bayesian multivariate extreme value mixture model

1. Problem Statement

Accurate risk assessment in multivariate systems requires models that capture both typical (bulk) behavior and rare extreme events. Existing approaches suffer from two primary limitations:

• Theoretical vs. Practical Trade-off: Classical Extreme Value Theory (EVT) provides asymptotically justified models for tails but fails to describe the bulk of the data. Conversely, standard mixture models often "entangle" bulk and tail inference, where the bulk copula influences the tail region and vice versa, leading to instability, especially in small samples.
• Threshold Selection: In Peaks-Over-Threshold (POT) methods, the threshold is typically fixed a priori. In multivariate settings, selecting a threshold is complex (often multimodal), and fixing it arbitrarily ignores uncertainty, leading to biased inference.
• Computational Challenges: Jointly modeling bulk and tail parameters creates strong posterior dependencies (particularly between the threshold and tail scale parameters), causing standard Markov Chain Monte Carlo (MCMC) samplers to mix poorly.

2. Methodology

The authors propose a Bayesian Multivariate Extreme Mixture Model (BEMM) that decouples bulk and tail modeling on disjoint supports while maintaining a unified probabilistic framework.

A. Model Structure

The model partitions the sample space $\mathbb{R}^d$ based on a multivariate threshold vector $\mathbf{u}$:

1. Bulk Region ($\mathbf{x} \le \mathbf{u}$): Modeled using a flexible parametric distribution. In the paper, a Multivariate Normal distribution is used for illustration, defined by mean $\boldsymbol{\mu}$ and covariance $\boldsymbol{\Sigma}$.
2. Tail Region ($\mathbf{x} \not\le \mathbf{u}$): Modeled using a Multivariate Generalized Pareto Distribution (mGPD). The authors utilize the U-representation with an independent reverse exponential generator. This ensures that the tail behavior is governed solely by the mGPD, independent of the bulk specification.

The joint density is:
$$f(\mathbf{x}) = \begin{cases} f_{bulk}(\mathbf{x}|\boldsymbol{\mu}, \boldsymbol{\Sigma}) & \text{if } \mathbf{x} \le \mathbf{u} \\ [1 - F_{bulk}(\mathbf{u}|\boldsymbol{\mu}, \boldsymbol{\Sigma})] \cdot h_U(\mathbf{x} - \mathbf{u}|\mathbf{a}, \boldsymbol{\sigma}, \boldsymbol{\gamma}) & \text{otherwise} \end{cases}$$
where $h_U$ is the mGPD density and $\mathbf{a}, \boldsymbol{\sigma}, \boldsymbol{\gamma}$ are tail dependence, scale, and shape parameters.

B. Bayesian Inference & Priors

• Learnable Threshold: Unlike traditional POT, the threshold $\mathbf{u}$ is treated as an unknown parameter. It is parametrized as $u_j = q_{\tau, j} + o_j$, where $q_{\tau, j}$ is a data-driven reference quantile and $o_j$ is a positive offset with a Half-Normal prior.
• Priors:
  • Bulk: Half-Normal priors for standard deviations, LKJ prior for the correlation matrix, and Normal priors for the mean.
  • Tail: Weakly informative priors with hard constraints (e.g., $\gamma_j < 1$) to ensure finite expectations.
• Posterior Sampling Strategy:
  • Reparametrisation: To address the strong dependence between the threshold $\mathbf{u}$ and the tail scale $\boldsymbol{\sigma}$, the authors reparametrize $\tilde{\boldsymbol{\sigma}} = \boldsymbol{\sigma} - \boldsymbol{\gamma}\mathbf{u}$. This transformation exploits the threshold stability property of the mGPD.
  • Automated Factor Slice Sampler (AFSS): Standard Metropolis-Hastings struggles with the posterior geometry. The authors employ AFSS, which performs univariate slice sampling along the eigenvectors of the covariance matrix of the target distribution. This significantly improves mixing efficiency.

3. Key Contributions

1. Decoupled Inference: The model strictly separates bulk and tail supports, ensuring that tail inference relies exclusively on the mGPD, eliminating the "entanglement" issues found in copula-based mixture models.
2. Learnable Multivariate Threshold: By treating the threshold as a parameter, the model provides coherent uncertainty quantification for threshold selection, which is critical in multivariate settings where thresholds may be multimodal.
3. Computational Efficiency: The combination of reparametrisation and AFSS enables robust posterior exploration even under strong parameter dependencies, overcoming a major bottleneck in multivariate EVT.
4. Theoretical Guarantees: The model is proven to lie in the max-domain of attraction of a multivariate Generalized Extreme Value Distribution (mGEVD), preserving asymptotic properties.

4. Results

Simulation Studies

• Parameter Recovery: In well-specified scenarios (varying tail heaviness), the model accurately recovered marginal and dependence parameters with coverage rates close to the nominal 95%.
• Robustness to Misspecification: When applied to asymptotically independent data (bivariate normal), the model (which assumes asymptotic dependence) still provided reasonable estimates for dependence measures (Kendall's $\tau$, $\bar{\chi}$) and outperformed or matched a bivariate mixture copula model (BMCM) in moderate sample sizes.
• Threshold Estimation: The model successfully identified the true threshold values and quantified their uncertainty.

Application: UK Temperature Data

• Data: Daily maximum temperatures from two UK stations (2016–2021).
• Preprocessing: Seasonality and autocorrelation were removed via regression, leaving residuals with heavy right tails.
• Findings:
  • Multimodal Thresholds: The posterior for the thresholds was multimodal, reflecting genuine ambiguity in the data rather than sampling artifacts.
  • Predictive Accuracy: The BEMM significantly outperformed a standard bivariate Gaussian model in estimating high quantiles and joint tail risk.
  • Scoring: The BEMM achieved lower Energy Scores (ES) and Threshold-Weighted Energy Scores (twES), indicating superior predictive performance, particularly in the tail regions.

5. Significance and Implications

• Environmental Risk Assessment: The model offers a rigorous framework for assessing joint risks (e.g., heatwaves, flooding) where both moderate and extreme events matter. It improves the estimation of high quantiles, which is crucial for infrastructure design and disaster planning.
• Methodological Advancement: It resolves the long-standing tension between theoretical EVT guarantees and practical bulk-tail modeling. By making the threshold learnable, it removes a major source of subjectivity in extreme value analysis.
• Limitations & Future Work: The model assumes asymptotic dependence (though it performs well even when this is weakly violated). It is currently best suited for low-dimensional settings ($d \le 3$) due to the complexity of high-dimensional dependence structures and the computational cost of MCMC. Future work could explore integrating generative models (e.g., normalizing flows) for the mGPD to handle higher dimensions, though this would require shifting from Bayesian to frequentist inference.

In summary, this paper presents a robust, theoretically grounded, and computationally efficient Bayesian framework for multivariate extreme value analysis, offering a superior alternative to existing mixture and copula-based approaches for joint risk characterization.