ExtremeRisks: Extreme Risk Measures

Simone A. Padoan, Gilles Stupfler, Carlotta PacificiOpenalex Unknown 2020 · DOI · OpenAlex

Abstract

The ExtremeRisks package provides a robust suite of procedures for estimating risks associated with extreme events using measures such as Expectiles and Value-at-Risk. It supports univariate independent observations, temporal dependent observations, and independent multidimensional datasets. By leveraging both parametric and non-parametric estimators, the package facilitates rigorous statistical inference, including the calculation of confidence intervals, regions, and hypothesis testing based on asymptotic theory.

Introduction

Standard statistical models often struggle to account for extreme events—such as market crashes, 100-year floods, or catastrophic insurance claims—which reside in the sparse "tails" of probability distributions. Traditional risk measures like Value-at-Risk (VaR) are frequently criticized for lacking mathematical "coherency." Expectiles have emerged as a compelling alternative, offering greater sensitivity to the magnitude of extreme values and better properties for risk management. However, estimating these measures at extreme levels requires specialized Extreme Value Theory (EVT). The ExtremeRisks package bridges this gap, providing practitioners with tools to handle temporal dependence and multivariate relationships in extreme risk estimation.

Methodology

The package implements a variety of frequentist and Bayesian techniques to quantify extreme risks:

  • Expectile Estimation: Utilizes Least Asymmetrically Weighted Squares (LAWS) and indirect Quantile-Based (QB) estimators to project risk levels far beyond observed data ranges.
  • Tail Index Estimation: Provides multiple methods to calculate distribution "heaviness," including Hill, Maximum Likelihood, and Expectile-Based (EB) approaches.
  • Temporal Dependence Handling: To accommodate time-series data, the package employs "big-block/small-block" techniques to adjust asymptotic variance, ensuring valid confidence intervals even in the presence of autocorrelation.
Parameters for the estExpectiles function, illustrating support for various estimation methods and temporal dependence adjustments.

Data and Simulation

To support benchmarking and demonstration, ExtremeRisks includes two primary real-world datasets: dowjones and sp500, both containing over 8,700 negative log-returns from 1985 to 2019. Additionally, the package features robust simulation functions (rtimeseries, rmdata, rbtimeseries) that allow users to generate synthetic heavy-tailed data with specific dependence structures using Gumbel, Clayton, or Frank copulas.

Key Insights and Results

Beyond basic estimation, the package offers advanced analytical capabilities:

  • Conditional Extremes: Through scedasis function estimation, users can model how the intensity of extreme events fluctuates based on external concomitant covariates.
  • Bayesian Integration: Bayesian Peaks-Over-Threshold (POT) modeling allows for the incorporation of prior beliefs, providing full posterior distributions for risk measures.
  • Systemic Risk Modeling: The package facilitates the calculation of Marginal Expected Shortfall (MES), a critical metric for identifying how individual components contribute to systemic risk during crises.
Parameters for modeling conditional extremes with concomitant covariates, used in scedasis function estimation.

Conclusion

ExtremeRisks provides a comprehensive toolkit for point and interval estimation of extreme risk measures. While currently focused on independent multidimensional observations and univariate time series, future updates are expected to expand multivariate tail indexing options and optimize performance for high-frequency datasets. The package remains a vital resource for researchers and practitioners needing to quantify the impact of rare, high-consequence events.

Summary updated

Abstract

A set of procedures for estimating risks related to extreme events via risk measures such as Expectile, Value-at-Risk, etc. is provided. Estimation methods for univariate independent observations and temporal dependent observations are available. The methodology is extended to the case of independent multidimensional observations. The statistical inference is performed through parametric and non-parametric estimators. Inferential procedures such as confidence intervals, confidence regions and hypothesis testing are obtained by exploiting the asymptotic theory. Adapts the methodologies derived in Padoan and Stupfler (2022) &lt;<a href="https://doi.org/10.3150%2F21-BEJ1375" target="_top">doi:10.3150/21-BEJ1375</a>&gt;, Davison et al. (2023) &lt;<a href="https://doi.org/10.1080%2F07350015.2022.2078332" target="_top">doi:10.1080/07350015.2022.2078332</a>&gt;, Daouia et al. (2018) &lt;<a href="https://doi.org/10.1111%2Frssb.12254" target="_top">doi:10.1111/rssb.12254</a>&gt;, Drees (2000) &lt;<a href="https://doi.org/10.1214%2Faoap%2F1019487617" target="_top">doi:10.1214/aoap/1019487617</a>&gt;, Drees (2003) &lt;<a href="https://doi.org/10.3150%2Fbj%2F1066223272" target="_top">doi:10.3150/bj/1066223272</a>&gt;, de Haan and Ferreira (2006) &lt;<a href="https://doi.org/10.1007%2F0-387-34471-3" target="_top">doi:10.1007/0-387-34471-3</a>&gt;, de Haan et al. (2016) &lt;<a href="https://doi.org/10.1007%2Fs00780-015-0287-6" target="_top">doi:10.1007/s00780-015-0287-6</a>&gt;, Padoan and Rizzelli (2024) &lt;<a href="https://doi.org/10.3150%2F23-BEJ1668" target="_top">doi:10.3150/23-BEJ1668</a>&gt;, Daouia et al. (2024) &lt;<a href="https://doi.org/10.3150%2F23-BEJ1632" target="_top">doi:10.3150/23-BEJ1632</a>&gt;.
BibTeX
@inproceedings{padoan2020extremerisks,
  author = {Simone A. Padoan and Gilles Stupfler and Carlotta Pacifici},
  title = {{ExtremeRisks: Extreme Risk Measures}},
  booktitle = {Openalex Unknown},
  year = {2020},
  doi = {10.32614/cran.package.extremerisks},
}
Figure