A1361
Title: Network-filtered copula segmentation for correlation break detection
Authors: Ning Zhang - Macquarie University (Australia) [presenting]
Abstract: Abrupt changes in cross-sectional dependence are common in financial, economic, and Network-indexed time series, especially during market stress, policy shifts, and macroeconomic instability. A joint Generalised Network Autoregressive (GNAR) copula framework is proposed for detecting structural changes in contemporaneous dependence while accounting for cross-variable temporal dynamics. The GNAR component parsimoniously models each variable as depending on its own lagged values and the lagged values of neighbouring variables defined by a prespecified Network. Conditional on these dynamics, a copula model characterizes the joint distribution of innovations, allowing flexible dependence, pairwise correlations, and tail dependence. The framework is formulated through a joint conditional likelihood for the GNAR and copula components. Within each candidate segment, mean dynamics and copula parameters are estimated jointly, enabling genuine dependence shifts to be distinguished from apparent correlation changes caused by serial dependence, Network spillovers, or marginal dynamics. Structural breaks are detected using a penalized likelihood segmentation criterion in the spirit of PELT. Simulations under Gaussian and heavy-tailed innovations show accurate breakpoint localization and reliable detection of dependence shifts. An empirical application to macroeconomic variables further demonstrates the model's ability to identify interpretable dependence regimes.
