Nonparametric estimation of extremal dependence

Anna Kiriliouk, Johan Segers, Michał WarchołarXiv 2014 · DOI

TL;DR

This paper introduces and explores nonparametric methods for understanding how extreme events in multiple financial assets are related. It develops techniques to distinguish between situations where extreme events tend to happen together (asymptotic dependence) and where they occur independently, using real-world stock market data.

What Problem Does This Paper Address?

When analyzing financial markets, it's crucial to understand not just how asset prices move on average, but also how extreme events (like sudden crashes or booms) in different assets are linked. This is particularly important for risk management and portfolio diversification. Traditional statistical methods often focus on the "center" of data distributions, but this paper dives into the "tails" – the rare, extreme events. The challenge is to model this "extremal dependence" in a flexible, nonparametric way, meaning without assuming a specific underlying distribution. A key difficulty is distinguishing between "asymptotic dependence" (where extreme events in different assets are likely to co-occur) and "asymptotic independence" (where they are not).

Methods

The paper builds upon extreme-value theory, focusing on two key concepts: the stable tail dependence function and the spectral measure.

  • Stable Tail Dependence Function (l): This function mathematically describes how the joint tail of a multivariate distribution behaves. It's a way to quantify the dependence between extreme values.
  • Spectral Measure (H): This provides an alternative perspective, describing the distribution of the "relative magnitudes" of components of a random vector, given that the vector itself is extreme.

The authors propose nonparametric estimators for these concepts. For the spectral measure, they introduce three estimators:

  1. Empirical spectral measure: A direct empirical version of the theoretical definition.
  2. Maximum empirical likelihood estimator: An estimator that enforces certain moment constraints, aiming for better efficiency.
  3. Maximum Euclidean likelihood estimator: Another efficient estimator that provides explicit formulas for weights.

For the stable tail dependence function, they propose estimators based on both direct estimation and by transforming the estimated spectral measure. They also introduce coefficients and tests to quantify and distinguish between asymptotic dependence and independence.

Datasets

The primary dataset used for illustration and evaluation consists of weekly negative log-returns of stock prices for three companies: JP Morgan, Citibank, and IBM, collected from January 1, 2000, to December 31, 2013.

Key Insights and Ideas

  • Visualizing Extremal Dependence: The scatterplots of transformed log-returns (Figure 1 and 2) provide an intuitive visual cue for the strength of dependence. JP Morgan and Citibank show a tighter clustering of extreme events compared to JP Morgan and IBM.
  • Distinguishing Dependence Types: The paper emphasizes the importance of differentiating between asymptotic dependence and independence. This is not a trivial distinction, and the proposed coefficients and tests are designed to address it.
  • Spectral Measure as a Diagnostic Tool: The spectral measure, particularly its distribution of "angles" (ratios of components), offers a powerful way to diagnose the nature of extremal dependence. For instance, a uniform spread of angles suggests strong dependence, while peaks at the boundaries indicate weak dependence.
  • Financial Intuition: The observed stronger dependence between the two financial institutions (JP Morgan and Citibank) compared to a financial institution and an IT company (IBM) aligns with economic intuition about sector-specific risks.

Key Results

  • Empirical Evidence: Analysis of the JP Morgan, Citibank, and IBM stock data reveals stronger tail dependence between the two banking stocks than between a bank and an IT company.
  • Estimator Performance: The paper shows that the maximum empirical likelihood and maximum Euclidean likelihood estimators for the spectral measure are more efficient than the basic empirical spectral measure.
  • Testing Framework: The proposed tests for asymptotic independence are applied to the financial data, with results suggesting asymptotic dependence for JP Morgan vs. Citibank, and asymptotic independence for JP Morgan vs. IBM. The Citibank vs. IBM case shows mixed results depending on the chosen parameters.

Limitations and Future Work

  • Asymptotic Independence Challenges: The paper notes that in cases of asymptotic independence, the stable tail dependence function alone might not be sufficient to quantify dependence above high, finite thresholds. This motivates the development of refined coefficients and tests.
  • Choice of Threshold: The performance of nonparametric estimators often depends on the choice of thresholds (e.g., for selecting extreme observations). The paper uses specific thresholds (like the 95% quantile) but acknowledges that optimal threshold selection can be a complex issue.
  • Generalization: While the paper focuses on bivariate cases for illustration, it mentions that the generalization to higher dimensions is straightforward for some concepts.
  • Semiparametric Inference: The paper suggests that nonparametric estimators of the stable tail dependence function can serve as a basis for semiparametric inference, where a parametric family is assumed for the dependence structure.
  • Further Testing: The authors point to the need for formal statistical tests to draw definitive conclusions about asymptotic independence, especially when graphical procedures are ambiguous.

Summary updated

Abstract

There is an increasing interest to understand the dependence structure of a random vector not only in the center of its distribution but also in the tails. Extreme-value theory tackles the problem of modelling the joint tail of a multivariate distribution by modelling the marginal distributions and the dependence structure separately. For estimating dependence at high levels, the stable tail dependence function and the spectral measure are particularly convenient. These objects also lie at the basis of nonparametric techniques for modelling the dependence among extremes in the max-domain of attraction setting. In case of asymptotic independence, this setting is inadequate, and more refined tail dependence coefficients exist, serving, among others, to discriminate between asymptotic dependence and independence. Throughout, the methods are illustrated on financial data.
BibTeX
@misc{kiriliouk2014nonparametric,
  author = {Anna Kiriliouk and Johan Segers and Michał Warchoł},
  title = {{Nonparametric estimation of extremal dependence}},
  year = {2014},
  eprint = {1411.0414},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1411.0414},
}
Figure