Measuring Association between Random Vectors

Oliver Grothe, Friedrich Schmid, Julius Schnieders, Johan SegersarXiv 2011 · DOI · arXiv

Abstract

This paper suggests five measures of association between two random vectors $X = (X_1, ..., X_p)$ and $Y = (Y_1, ..., Y_q)$. They are copula based and therefore invariant with respect to the marginal distributions of the components $X_i$ and $Y_j$. The measures capture positive as well as negative association of $X$ and $Y$. In case $p = q = 1$ they reduce to Spearman's rho. Various properties of these new measures are investigated. Nonparametric estimators, based on ranks, for the measures are derived and their small sample behaviour is investigated by simulation. The measures are applied to characterise strength and direction of association of bond and stock indices of five countries over time.

Introduction

In multivariate statistics, quantifying the relationship between two sets of variables—random vectors—is a fundamental challenge. Traditional metrics like the RV coefficient or canonical correlation often rely on assumptions of normality or are limited to capturing only the magnitude of an association without indicating its direction. In financial contexts, where data is frequently non-normal and relationships can flip from positive to negative, these limitations are significant.

This paper introduces five new measures of association based on copulas. By focusing on the copula—the "glue" that binds variables together—these measures remain invariant to the marginal distributions of individual variables. Crucially, they extend the concept of Spearman’s rho to a multivariate setting, allowing researchers to detect whether two groups of variables are moving in tandem or in opposition.

The Need for Directional, Marginal-Free Measures

Existing tools for measuring vector association, such as distance correlation, have gained popularity because they can detect non-linear dependencies. However, distance correlation is always non-negative. In many applications, such as the relationship between stock and bond markets, knowing the sign of the association is as important as knowing its magnitude.

Furthermore, financial returns often exhibit heavy tails and skewness. A robust measure should be "marginal-free," meaning its value depends only on the dependence structure (the copula) and not on whether the individual variables follow a Gaussian, Student-t, or any other distribution.

Proposed Copula-Based Measures

The authors propose five measures ($\bar{ ho}, \rho_1, \rho_2, \rho_3, \rho_4$) designed to quantify the association between a $p$-dimensional vector $X$ and a $q$-dimensional vector $Y$.

  • Mean Pairwise Association ($\bar{ ho}$): This is the average of all possible Spearman’s rho correlations between the components of $X$ and $Y$. It provides a straightforward, intuitive baseline.
  • Pearson-based Transformations ($\rho_1, \rho_2$): These measures apply Pearson correlation to specific transformations of the variables. $\rho_1$ utilizes a product-based transformation of the ranks, while $\rho_2$ is based on the joint distribution functions.
  • Rank-based Transformations ($\rho_3, \rho_4$): These are Spearman-type extensions applied to the same transformations used in $\rho_1$ and $\rho_2$.

A defining feature of all five measures is that they reduce to the standard Spearman’s rho when $p = q = 1$. This ensures that the new framework is a consistent extension of classical bivariate rank correlation.

Methodology and Estimation

To make these measures practical for real-world data, the authors developed nonparametric estimators based on ranks. These estimators do not require knowledge of the underlying distributions, making them highly robust.

Simulation Studies

The performance of these estimators was tested using Gaussian and Clayton copulas across various sample sizes ($n=50, 100, 500$). The simulations demonstrated that the estimators are consistent; the bias is minimal even in small samples and vanishes as $n$ increases. The study also validated the use of bootstrap and jackknife methods for estimating the standard deviation of these measures, with the bootstrap showing slightly higher precision.

Empirical Application: Stocks vs. Bonds

The measures were applied to daily returns of stock and bond indices from five major economies (Australia, France, Germany, Japan, and the USA) between 1996 and 2010. Using a 250-day rolling window, the authors tracked how the association between these two asset classes evolved over time.

The results highlighted a significant shift in market dynamics:

  • Late 1990s: The association was generally positive.
  • 2002–2006: The association hovered near zero, suggesting a period of decoupling.
  • Post-2006: The association became sharply negative, particularly during the 2008 financial crisis.

This "flight to quality"—where stocks plummet and bonds rise—was clearly captured by the sign of the new measures. In contrast, traditional measures like the RV coefficient would only show a high magnitude of association, failing to distinguish this negative relationship from a positive one.

Conclusion

The proposed copula-based measures provide a robust, flexible framework for analyzing the dependence between random vectors. By being invariant to marginal distributions and sensitive to the direction of association, they offer a more nuanced view of complex systems than traditional methods. While future work remains to derive the asymptotic normality for all five measures and to extend the framework to more than two vectors, these tools represent a significant step forward in multivariate dependence analysis.

Summary updated

Abstract

This paper suggests five measures of association between two random vectors X = (X_1, ..., X_p) and Y = (Y_1, ..., Y_q). They are copula based and therefore invariant with respect to the marginal distributions of the components X_i and Y_j. The measures capture positive as well as negative association of X and Y. In case p = q = 1 they reduce to Spearman's rho. Various properties of these new measures are investigated. Nonparametric estimators, based on ranks, for the measures are derived and their small sample behaviour is investigated by simulation. The measures are applied to characterise strength and direction of association of bond and stock indices of five countries over time.
BibTeX
@misc{grothe2011measuring,
  author = {Oliver Grothe and Friedrich Schmid and Julius Schnieders and Johan Segers},
  title = {{Measuring Association between Random Vectors}},
  year = {2011},
  eprint = {1107.4381},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1107.4381},
}
Figure