Non-Parametric Bayesian Inference on Bivariate Extremes

Simon Guillotte, François Perron, Johan Segers

TL;DR

This paper introduces a novel Bayesian framework for analyzing extreme events in pairs of variables. It tackles the challenge of modeling complex dependence structures in extreme value theory by proposing a flexible, non-parametric approach to the "spectral measure" that governs this dependence. This allows for more accurate predictions of rare events and extreme conditional quantiles, crucial for risk management.

What Problem Does This Paper Address?

In fields like finance and engineering, understanding extreme outcomes of multiple variables simultaneously is critical. For instance, predicting the combined impact of two simultaneous river floods or assessing financial risk across correlated assets requires modeling the "tail" of a bivariate distribution – the region where both variables take on very large values. Traditional methods often rely on parametric assumptions for the dependence structure, which can be too restrictive. This paper addresses the need for a more flexible, non-parametric approach to model this dependence, particularly in the context of Bayesian inference, which naturally handles uncertainty.

Methods

The core of the paper lies in developing a Bayesian framework for non-parametric inference on bivariate extreme value distributions. Key methods include:

  • Censored Likelihood: Since extreme value data is often sparse, a censored likelihood approach is used. This means that data points falling below a certain threshold are treated as "censored," allowing them to contribute to the inference without requiring them to be extreme themselves.
  • Infinite-Dimensional Model for Spectral Measure: The dependence structure of bivariate extreme values is characterized by a "spectral measure." Instead of assuming a specific parametric form for this measure, the authors propose an infinite-dimensional model that is dense within the space of all possible spectral measures. This offers greater flexibility.
  • Trans-dimensional Markov Chain Monte Carlo (MCMC): To perform Bayesian inference with this complex, infinite-dimensional model, a specialized MCMC algorithm is developed. This algorithm allows the model to "jump" between different dimensional representations of the spectral measure, enabling exploration of the entire parameter space.
  • Prior Construction: A crucial contribution is the construction of a prior distribution on the space of spectral measures. This prior is designed to be computationally manageable while still allowing for dense approximation of the true spectral measure.

Datasets

  • Artificial Data: The methodology is validated using simulated bivariate data generated from a specific bivariate cumulative distribution function with varying dependence parameters.
  • Danish Fire Insurance Claims: The method is applied to a real-world dataset of industrial fire losses, previously analyzed in univariate contexts. This dataset comprises 2167 observations across three categories of damage.

Key Insights and Ideas

  • Flexibility in Dependence Modeling: The non-parametric approach to the spectral measure is a significant advancement, allowing for the capture of complex and potentially unusual dependence structures in extreme events that parametric models might miss.
  • Bayesian Approach for Uncertainty: The Bayesian framework naturally incorporates statistical uncertainty, which is particularly important when dealing with rare extreme events where data is scarce. This leads to more robust predictions.
  • Computational Tractability: Despite the infinite-dimensional nature of the spectral measure, the authors develop a computationally feasible MCMC algorithm, making the approach practical for real-world applications.
  • Predictive Power: The developed methodology enables the computation of bivariate predictive densities in tail regions, which is directly applicable to predicting rare event probabilities and extreme conditional quantiles.

Key Results

  • The proposed Bayesian method, implemented via a trans-dimensional MCMC algorithm, successfully estimates the spectral measure of bivariate extreme value distributions.
  • Simulations show that the Bayes estimator for the spectral measure performs favorably compared to frequentist non-parametric estimators.
  • The application to Danish fire insurance data reveals non-trivial dependence structures between different categories of damage in extreme loss scenarios, which is crucial for risk assessment.
  • The method provides credible intervals for the spectral measure, quantifying the uncertainty in the dependence structure.

Limitations and Future Work

  • Dimensionality: The authors acknowledge that generalizing the approach to higher dimensions (d > 2) presents significant computational challenges due to the exponential increase in the number of censored likelihood branches and the complexity of the spectral measure space.
  • Asymptotic Independence: While the paper focuses on extreme value attractors, the authors note that for cases of asymptotic independence (e.g., highly correlated bivariate Gaussian distributions), the tail approximation might be less satisfactory. They point to existing frameworks that handle asymptotic independence and dependence more broadly.
  • Future Work: The paper suggests that extending the methodology to arbitrary dimensions is a promising avenue for future research, despite the inherent difficulties. Further exploration of models that encompass both asymptotic independence and dependence is also highlighted.

Summary updated

Abstract

Summary The tail of a bivariate distribution function in the domain of attraction of a bivariate extreme value distribution may be approximated by that of its extreme value attractor. The extreme value attractor has margins that belong to a three-parameter family and a dependence structure which is characterized by a probability measure on the unit interval with mean equal to 12, which is called the spectral measure. Inference is done in a Bayesian framework using a censored likelihood approach. A prior distribution is constructed on an infinite dimensional model for this measure, the model being at the same time dense and computationally manageable. A trans-dimensional Markov chain Monte Carlo algorithm is developed and convergence to the posterior distribution is established. In simulations, the Bayes estimator for the spectral measure is shown to compare favourably with frequentist non-parametric estimators. An application to a data set of Danish fire insurance claims is provided.
BibTeX
@article{guillotte2011non,
  author = {Simon Guillotte and François Perron and Johan Segers},
  title = {{Non-Parametric Bayesian Inference on Bivariate Extremes}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2011},
  doi = {10.1111/j.1467-9868.2010.00770.x},
}
Figure