Cross-Frequency Dependence Overlap in the Chinese A-Share Dependence Network.

Abstract

Multiresolution analysis is widely applied to equity markets on the assumption that different frequency bands capture distinct trading behaviors and information flows. Whether those bands reveal structurally distinct equity communities, or mostly re-express a shared dependence backbone, remains unresolved. We address this question in the Chinese A-share market using the maximal overlap discrete wavelet transform (MODWT) and a dual-channel affinity construction that processes positive and negative similarities separately before fusing them into a nonnegative matrix for normalized spectral clustering, over a grid of 15,360 configurations spanning wavelet bases, decomposition levels, and time windows. In the reference configuration (K=25), the adjusted Rand index (ARI) against the Shenwan first-level industry classification is 0.381, comparable to the strongest correlation-based baseline and with no singleton communities. Best-over-grid ARI declines monotonically from 0.381 at decomposition level 2 to 0.139 at level 6, in every time window and across the full configuration ensemble. Band-wise stock-pair similarity vectors show substantial rank agreement (mean Spearman ρ = 0.769), yet cross-band normalized mutual information averages 0.123 over 64 equal-frequency bins, and independently clustered bands give distinct partitions (mean pairwise ARI = 0.328). Rank agreement falls to 0.677 after market removal and to 0.486 after sequential market-and-industry removal. These results document measurable but incomplete cross-frequency dependence overlap, alongside declining industry alignment with increasing decomposition depth within the tested pipeline.