Detection of block-exchangeable structure in large-scale correlation matrices (Q1755136)

The authors propose a technique for learning the block structure within the matrix \(\mathbf{T}\) of pairwise Kendall rank correlations for underlying data \(\mathbf{X}=(X_1,\ldots,X_d)\) with continuous marginals. This is done under an assumption of partial exchangeability which is weaker than full exchangeability. Information about the correlation structure based on this data then allows improved estimation of the matrix \(\mathbf{T}\) as a by-product, with the improved estimators shown to be consistent, asymptotically Gaussian, and to have an asymptotic variance smaller than that of the empirical Kendall coefficients. In the case where \(\mathbf{X}\) is elliptical, this procedure may also be used to provide improved estimates of the linear correlation matrix with entries \(\text{cor}(X_i,X_j)\). The proposed algorithms are illustrated using a simulation study and an application to financial stock return data.