A Parsimonious Test of Constancy of a Positive Definite Correlation Matrix in a Multivariate Time-Varying GARCH Model

Jian Kang, Johan Stax Jakobsen, Annastiina Silvennoinen, Timo Teräsvirta, G. WadeEconometrics 2022 · DOI

TL;DR

This paper introduces a new, more efficient statistical test to determine if the correlation patterns between multiple financial assets remain stable over time. This is crucial for risk management and portfolio optimization, especially in complex markets where correlations can shift unpredictably. The proposed test is more "parsimonious," meaning it requires fewer computational resources and is more practical for large numbers of assets compared to existing methods.

What Problem Does This Paper Address?

In finance, understanding how the prices of different assets move together (their correlation) is fundamental for managing risk and building investment portfolios. While simple models assume these correlations are constant, real-world markets are dynamic, and correlations can change significantly, especially during turbulent periods like financial crises. Existing models, like the Constant Conditional Correlation (CCC) GARCH, assume constant correlations, but more advanced models like DCC-GARCH and VC-GARCH allow for time-varying correlations. However, a significant challenge has been the lack of robust statistical tests to definitively check if these correlations are indeed constant or if they are changing deterministically over time. This gap makes it difficult to choose the right model and can lead to misinformed investment decisions. The authors aim to fill this gap by developing a parsimonious and practical test for correlation constancy in multivariate time-varying GARCH models.

Methods

The core of the paper lies in developing a new statistical test. The authors leverage the spectral decomposition of the correlation matrix. Instead of looking at each individual correlation coefficient, they focus on the eigenvalues of the correlation matrix. This approach is more "parsimonious" because it reduces the number of parameters to test.

The test is derived using a Lagrange multiplier (LM) principle. This involves estimating the model under the null hypothesis (constant correlations) and then using the score and information matrix derived from the model's likelihood function to construct the test statistic. A key innovation is the use of a Taylor series expansion to approximate the time-varying components, allowing for the derivation of a testable hypothesis. The test is designed to be flexible, allowing for correlations to change as a function of time, and can be generalized to other transition variables.

Datasets

The authors use daily returns data for 26 stocks included in the Dow Jones Industrial Average for the period from January 2, 2001, to December 31, 2020. This period encompasses significant market events like the dot-com bubble and the Global Financial Crisis. The data is split into two halves for analysis.

Key Insights and Ideas

  • Eigenvalues as a Shortcut: Focusing on the eigenvalues of the correlation matrix, rather than individual correlation coefficients, is a more efficient way to test for overall correlation stability, especially in high-dimensional systems.
  • Parsimony is Practical: The proposed test is computationally less demanding than existing alternatives (like the HST-test), making it more applicable to large portfolios and real-time analysis.
  • GARCH Specification Matters: The simulations strongly suggest that correctly specifying the underlying GARCH equations (i.e., accounting for volatility clustering, asymmetry, etc.) is crucial before testing for correlation constancy. Mis-specifying the GARCH model can lead to spurious rejections of correlation constancy.
  • Time-Varying Correlations are Common: The empirical application to Dow Jones stocks provides evidence that correlations between many pairs of stocks do indeed change over time, particularly during periods of market stress.

Key Results

  • Size Properties: Simulation studies show that the proposed test has good size properties, meaning it correctly rejects the null hypothesis of constant correlations at the specified significance level in most scenarios. The test is generally well-behaved, with only minor over-sizing observed in specific cases (e.g., small sample sizes or higher-order polynomials).
  • Power: The test demonstrates good power in detecting time-varying correlations, especially when the underlying GARCH models are correctly specified.
  • Empirical Evidence: The application to 26 Dow Jones stocks reveals that the null hypothesis of constant correlations is strongly rejected for most pairs of stocks, particularly during the first period (2001-2010) which included market turbulence. The second period (2011-2020) also shows evidence of time-varying correlations, though sometimes less pronounced or non-monotonic.
  • Comparison to HST-test: The paper shows that their parsimonious test is often more practical and yields similar conclusions to the more computationally intensive HST-test, especially when the number of assets is large.

Limitations and Future Work

  • Transition Variable Choice: The proposed test, like the HST-test, is conditioned on the choice of a transition variable (in this paper, a linear time trend). While flexible, practitioners need to select an appropriate transition variable that reflects the underlying economic drivers of correlation changes. The theoretical foundations are unaffected by this choice, but practical application requires careful consideration.
  • GARCH Model Specification: The simulations highlight the critical importance of correctly specifying the GARCH equations. Future work could explore integrated approaches that jointly test GARCH specification and correlation constancy.
  • Exogenous Variables: The paper does not explore the impact of ignored exogenous variables in GARCH models on the test's performance. This is a potential area for future research.
  • Computational Intensity: While more parsimonious than some alternatives, the simulations still required significant computational resources, especially for higher dimensions and longer time series. The authors mention using MS Azure Virtual Machines for processing.

Summary updated

Abstract

We construct a parsimonious test of constancy of the correlation matrix in the multivariate conditional correlation GARCH model, where the GARCH equations are time-varying. The alternative to constancy is that the correlations change deterministically as a function of time. The alternative is a covariance matrix, not a correlation matrix, so the test may be viewed as a general test of stability of a constant correlation matrix. The size of the test in finite samples is studied by simulation. An empirical example involving daily returns of 26 stocks included in the Dow Jones stock index is given.
BibTeX
@inproceedings{kang2022parsimonious,
  author = {Jian Kang and Johan Stax Jakobsen and Annastiina Silvennoinen and Timo Teräsvirta and G. Wade},
  title = {{A Parsimonious Test of Constancy of a Positive Definite Correlation Matrix in a Multivariate Time-Varying GARCH Model}},
  booktitle = {Econometrics},
  year = {2022},
  doi = {10.3390/econometrics10030030},
}
Figure