Patterns of dependence in financial networks, such as global bank connectedness, evolve over time and across frequencies. Analysing these systems requires statistical tools that jointly capture temporal dynamics and the underlying network topology. This work develops a novel spectral analysis framework for Generalized Network Autoregressive (GNAR) processes, modeling dependencies beyond direct neighbours by incorporating r-stage neighbourhood effects, unlike existing methods that at best rely solely on adjacency-based interactions. We define the GNAR spectral density and related quantities, such as coherence and partial coherence, for which we propose both parametric and network-penalized nonparametric estimators. Extensive simulations demonstrate the strong performance of the parametric spectral estimator, as also backed up by theoretical arguments. The proposed framework has wide applications, and here we focus on the analysis of global bank network connectedness. The findings illustrate how the GNAR spectral quantities effectively capture the frequency-specific cross-nodal dependencies, thus yielding estimates consistent with established measures, while also uncovering richer temporal and structural patterns of volatility transmission.
翻译:暂无翻译