A Method of Moments Estimator of Tail Dependence

J.h.j. Einmahl, Andrea Krajina, Johan SegersSSRN 2007 · DOI

TL;DR

This paper introduces a novel "method of moments" estimator for the tail dependence structure of bivariate random variables. This approach offers a robust and flexible way to model extreme dependencies, even in cases where traditional likelihood-based methods fail, and is shown to be consistent and asymptotically normal.

What Problem Does This Paper Address?

In multivariate statistics, understanding how extreme values of different variables relate to each other (tail dependence) is crucial for risk management, finance, and other fields. While methods exist for modeling this dependence, they often struggle with:

  • High dimensionality: As the number of variables increases, estimation becomes significantly harder.
  • Model misspecification: Assuming a specific parametric form for the dependence can lead to inaccurate results if the true form is different.
  • Data limitations: Some extreme value distributions, particularly those arising from factor models, have discrete spectral measures, making standard likelihood estimation impossible.

This paper tackles the challenge of estimating tail dependence in a bivariate setting, aiming for a method that is robust to model misspecification and can handle complex dependence structures, paving the way for multivariate extensions.

Methods

The core contribution is a semi-parametric method of moments estimator. Here's how it works:

  1. Nonparametric Estimation of Tail Dependence: The authors first use a rank-based, nonparametric estimator for the stable tail dependence function. This estimator is derived from empirical quantiles and is invariant to the marginal distributions of the variables.
  2. Parametric Modeling: They then assume that the true tail dependence function belongs to a specific parametric family, denoted by $l(x, y; \theta)$, where $\theta$ is the parameter to be estimated.
  3. Method of Moments: The key idea is to match an integral of the nonparametric estimator with the corresponding integral of the parametric model. Specifically, they define a function $\phi(\theta)$ that involves integrating a function $g(x,y)$ multiplied by the parametric tail dependence function $l(x, y; \theta)$. The method of moments estimator $\hat{\theta}_n$ is then found by solving $\phi(\hat{\theta}_n) = \int\int g(x,y) \hat{l}(x,y) dx dy$, where $\hat{l}(x,y)$ is the nonparametric estimator.
  4. Goodness-of-Fit Test: A byproduct of this approach is a goodness-of-fit test. By comparing the nonparametric estimator to the estimated parametric model, they can assess how well the chosen parametric family fits the data.

The method is designed to work under very weak conditions, notably not requiring smoothness of the tail dependence function, which is a significant advantage over some other methods.

Datasets

The paper does not introduce or use specific real-world datasets. Instead, it relies on:

  • Simulated data: The performance of the proposed estimator is evaluated through extensive simulations using various models, including a two-point spectral measure arising from a factor model and families of meta-elliptical distributions. These simulations are crucial for demonstrating the estimator's properties like bias and root mean squared error (RMSE).

Key Insights and Ideas

  • Bridging Nonparametric and Parametric: The method elegantly combines the robustness of nonparametric estimators with the efficiency and interpretability of parametric models.
  • Handling Discrete Spectral Measures: A major breakthrough is the ability to estimate tail dependence even when the underlying spectral measure is discrete. This is a scenario where likelihood-based methods fail, making this estimator particularly valuable for certain factor models.
  • Weak Assumptions: The estimator's consistency and asymptotic normality hold under remarkably weak conditions, making it broadly applicable.
  • Flexibility for Multivariate Extensions: The semi-parametric nature of the approach is explicitly designed to facilitate extensions to higher dimensions, a significant challenge in extreme value theory.

Key Results

  • Consistency and Asymptotic Normality: The paper rigorously proves that the method of moments estimator $\hat{\theta}_n$ is consistent (converges to the true parameter $\theta_0$) and asymptotically normal (its distribution approaches a normal distribution as sample size increases).
  • Performance in Simulations: Simulations demonstrate that the estimator performs well, exhibiting reasonable bias and RMSE across various scenarios, including those with discrete spectral measures and different types of elliptical distributions.
  • Goodness-of-Fit Test: The proposed goodness-of-fit test is shown to be effective in detecting model misspecification.

Limitations and Future Work

  • Bivariate Focus: The current work is primarily focused on the bivariate case. While the authors state the approach opens the road to multivariate settings, the theoretical guarantees and practical implementation for higher dimensions are left for future work.
  • Choice of Auxiliary Function $g$: The performance of the estimator can depend on the choice of the auxiliary function $g(x,y)$. While the paper mentions that the optimal $g$ would minimize asymptotic variance, finding this optimal function is often difficult, and practical choices are made to simplify calculations.
  • Computational Complexity: For very high dimensions, the computational cost of the nonparametric estimation and the integral matching might become substantial.

The authors clearly indicate that extending this methodology to multivariate settings is a key direction for future research. Investigating the optimal choice of the function $g$ and its impact on efficiency would also be a valuable area of study.

Summary updated

Abstract

In the world of multivariate extremes, estimation of the dependence structure still presents a challenge and an interesting problem. A procedure for the bivariate case is presented that opens the road to a similar way of handling the problem in a truly multivariate setting. We consider a semi-parametric model in which the stable tail dependence function is parametrically modeled. Given a random sample from a bivariate distribution function, the problem is to estimate the unknown parameter. A method of moments estimator is proposed where a certain integral of a nonparametric, rank-based estimator of the stable tail dependence function is matched with the corresponding parametric version. Under very weak conditions, the estimator is shown to be consistent and asymptotically normal. Moreover, a comparison between the parametric and nonparametric estimators leads to a goodness-of-fit test for the semiparametric model. The performance of the estimator is illustrated for a discrete spectral measure that arises in a factor-type model and for which likelihood-based methods break down. A second example is that of a family of stable tail dependence functions of certain meta-elliptical distributions.
BibTeX
@misc{einmahl2007method,
  author = {J.H.J. Einmahl and Andrea Krajina and Johan Segers},
  title = {{A Method of Moments Estimator of Tail Dependence}},
  year = {2007},
  doi = {10.2139/ssrn.1025397},
  howpublished = {SSRN},
}
Figure