Maximum empirical likelihood estimation of the spectral measure of an extreme-value distribution

J.h.j. Einmahl, Johan SegersResearch 2008 · DOI · arXiv

Abstract: Consider a random sample from a bivariate distribution function $F$ in the max-domain of attraction of an extreme-value distribution function $G$. This $G$ is characterized by two extreme-value indices and a spectral measure, the latter determining the tail dependence structure of $F$. A major issue in multivariate extreme-value theory is the estimation of the spectral measure $Φ_p$ with respect to the $L_p$ norm. For every $p\in[1,\infty]$, a nonparametric maximum empirical likelihood estimator is proposed for $Φ_p$. The main novelty is that these estimators are guaranteed to satisfy the moment constraints by which spectral measures are characterized. Asymptotic normality of the estimators is proved under conditions that allow for tail independence. Moreover, the conditions are easily verifiable as we demonstrate through a number of theoretical examples. A simulation study shows a substantially improved performance of the new estimators. Two case studies illustrate how to implement the methods in practice.

Introduction

When a "once-in-a-century" flood happens, does a "once-in-a-century" storm surge happen at the same time? Understanding how extreme events are linked—known as tail dependence—is critical for insurance, finance, and disaster management.

In their 2009 paper, John H. J. Einmahl and Johan Segers tackle a persistent problem in Multivariate Extreme Value Theory: how to create a nonparametric estimator for this dependence that is both mathematically valid and statistically efficient.

The Problem: Invalidity and Tail Independence

In extreme value theory, the dependence between tails of a bivariate distribution is captured by a spectral measure. However, existing nonparametric estimators face two major hurdles:

  1. Invalidity: A true spectral measure must satisfy specific "moment constraints." Standard empirical estimators often violate these rules, resulting in an estimate that is not a valid spectral measure.
  2. Tail Independence: Many models struggle when variables are "tail independent"—meaning an extreme in one doesn't make an extreme in the other more likely. Older estimators often failed to prove mathematical consistency in these scenarios.

Methods: Maximum Empirical Likelihood

The authors bridge the gap between raw data and theoretical requirements using Nonparametric Maximum Empirical Likelihood Estimation (MELE).

  1. Rank-Based Transformation: Data is transformed into a margin-free format using ranks, focusing purely on the relationship between variables.
  2. Enforcing Constraints: Instead of simple counting, weights are assigned to observations to maximize the likelihood subject to theoretical moment constraints.
  3. $L_p$ Norm Generalization: The framework supports any $L_p$ norm ($p \in [1, \infty]$), allowing researchers to choose the representation that best fits their field.
Comparison of the geometric regions $C_p$ for $p=1$ and $p=2$.
  1. Asymptotic Theory: The authors prove that the estimator converges to a normal distribution (asymptotic normality), even in cases of tail independence.

Datasets and Validation

The method was validated using three types of data:

  • Simulated Models: Logistic, Cauchy, and Mixture models to compare MELE against standard empirical estimators.
  • Loss-ALAE Insurance Data: 1,500 insurance claims (indemnity payment vs. expense).
  • NHANES Health Data: Body measurement data (height vs. weight) for 2,237 females.

Key Insights

  • Constraints as a Feature: Forcing the estimator to satisfy moment constraints increases efficiency. The mathematical structure provides more information than raw data alone.
  • The Independence "Aha": NHANES data revealed that while height and weight are generally correlated, they exhibit tail independence. MELE successfully captured this "flattening" of dependence at the edges where other models might over-predict risk.
  • Verifiable Conditions: The paper provides "easily verifiable" conditions for its proofs, making the theory accessible for practical application.

Results and Conclusion

In simulations, MELE consistently showed a lower Mean Integrated Squared Error (MISE) than standard empirical methods. It proved robust across different $L_p$ norms and was particularly effective at identifying tail independence. Case studies demonstrated that MELE provides a smoother, more theoretically sound fit for real-world data like insurance claims.

While future work is needed for higher dimensions and automated parameter selection ($k$), this research provides a significant step forward in making extreme value modeling more reliable and mathematically consistent.

Summary updated

Abstract

Consider a random sample from a bivariate distribution function $F$ in the max-domain of attraction of an extreme-value distribution function $G$. This $G$ is characterized by two extreme-value indices and a spectral measure, the latter determining the tail dependence structure of $F$. A major issue in multivariate extreme-value theory is the estimation of the spectral measure $Φ_p$ with respect to the $L_p$ norm. For every $p\in[1,\infty]$, a nonparametric maximum empirical likelihood estimator is proposed for $Φ_p$. The main novelty is that these estimators are guaranteed to satisfy the moment constraints by which spectral measures are characterized. Asymptotic normality of the estimators is proved under conditions that allow for tail independence. Moreover, the conditions are easily verifiable as we demonstrate through a number of theoretical examples. A simulation study shows a substantially improved performance of the new estimators. Two case studies illustrate how to implement the methods in practice.
BibTeX
@inproceedings{einmahl2008maximum,
  author = {J.H.J. Einmahl and Johan Segers},
  title = {{Maximum empirical likelihood estimation of the spectral measure of an extreme-value distribution}},
  booktitle = {Research},
  year = {2008},
  doi = {10.48550/arxiv.0812.3485},
}
Figure