Multivariate tests of independence and their application in correlation analysis between financial markets (Q2196137)

The authors propose multivariate sign tests for the independence of two random vectors which may offer superior power if the distributions are fat-tailed and/or the number of observations is less than the dimensions of the vectors. The asymptotic normality of the tests is obtained for elliptically symmetric distributions. Let \(X\) denote a \(p\)-dimensional random vector with elliptically symmetric density \[ \operatorname{det}(\Omega_{X})^{-1/2}g_{X} \left(\|\Omega_{X}^{-1/2}\cdot (x-\theta_{X})\|\right), \] where \(\theta_{X}\) is the center of symmetry and \(\Omega_{X}\) is symmetric and positive semidefinite, and let \(X_{1},\dots,X_{n}\) be independent, identically distributed realizations of \(X\). Similarly, let \(Y_{1},\dots,Y_{n}\) be independent, identically distributed realizations of the \(q\)-dimensional \(Y\) given by the density \[ \operatorname{det}(\Omega_{Y})^{-1/2}g_{Y} \left(\|\Omega_{Y}^{-1/2}\cdot (y-\theta_{Y}) \|\right). \] The proposed high-dimensional multivariate sign test for the independence of \(X\) and \(Y\) is \[ T= \frac{n\cdot \displaystyle\sum_{1\leq i<j\leq n}\hat U_{X,i}^{\top} \hat U_{X,j} \hat U_{Y,i}^{\top} \hat U_{Y,j}}{\sqrt{2\cdot \displaystyle\sum_{1\leq i<j\leq n}[\hat U_{X,i}^{\top} \hat U_{X,j}]^{2} \cdot \displaystyle\sum_{1\leq i<j\leq n}[\hat U_{Y,i}^{\top} v\hat U_{Y,j}]^{2}}} \] with \(\hat U_{X,i} = U(X_{i}-\hat \theta_{X})\), where \(U\) is the spatial sign function defined by \(U(z) = z/\|z\|^{-1}\) \((z\neq 0)\) and \(\hat \theta_{X}\) is the spatial median defined as the minimizer of \(L(\theta) = \sum_{i=1}^{n}\|X_{i}-\theta\|\); the definition of \(\hat U_{Y,j}\) is analogous. The authors also define the high-dimensional Spearman's rho test: with \(\sum^{\ast}\) denoting summation over distinct indices, \[ T_{S} = \frac{\sqrt{2n}\sum^{\ast}U(X_{i}-X_{j})^{\top}U(X_{k}-X_{l})U(Y_{i}-Y_{l})^{\top}U(Y_{k}-Y_{j})}{\sqrt{\sum^{\ast}[U(X_{i}-X_{j})^{\top}U(X_{k}-X_{l})]^{2}\sum^{\ast}[U(Y_{i}-Y_{j})^{\top}U(Y_{k}-Y_{l})]^{2}}}; \] and the high-dimensional Kendall's tau test: \[ T_{K} = \frac{n\sum^{\ast}U(X_{i}-X_{j})^{\top}U(X_{k}-X_{l})U(Y_{i}-Y_{j})^{\top}U(Y_{k}-Y_{l})}{\sqrt{2\sum^{\ast}[U(X_{i}-X_{j})^{\top}U(X_{k}-X_{l})]^{2}\sum^{\ast}[U(Y_{i}-Y_{j})^{\top}U(Y_{k}-Y_{l})]^{2}}}. \] The authors show that under the null hypothesis that \(X\) and \(Y\) are independent, \(T\), \(T_{S}\), and \(T_{K}\) are asymptotically normally distributed if \(p,q = O(n^{2})\) and the covariance matrices of \(X\) and \(Y\) satisfy some smallness conditions. In a simulation-based analysis, it is shown that for multivariate \(t\)-distributions and mixture normal distributions, the tests above have better power than some alternative tests proposed in the literature. The tests are applied to weekly returns of S\&P500 and CSI300 stocks to suggests that in most years, the stock returns in these indices are independent, but in certain periods their independence can be rejected at the 99\% confidence level. Economic interpretations are given for the high correlation in 2010--2012 and 2016--2017. A similar analysis rejects the independence of the Shanghai and Shenzhen stock exchanges.