A1824
Title: Forecasting performance of dynamic approaches for spillovers in networks
Authors: Rebekka Buse - Karlsruhe Institute of Technology (Germany) [presenting]
Abstract: The purpose is to study the out-of-sample forecasting performance of smooth versus rolling window approaches for evaluating spillovers in dynamic networks. Each approach has particular advantages: The frequentist approach with rolling window vector autoregression (VAR), on the one hand, is very precise with respect to the dynamics within each window, while the Bayesian approach with time-varying parameter VAR, on the other hand, yields dynamics that are smooth and independent of window size. We compare the out-of-sample forecast performance of both approaches in order to distinguish the most precise and up-to-date measure for real-time interconnected systems. In a comprehensive simulation study, we investigate both level data as well as resulting network measures for different types of dynamics, including not only times of steady evolution, but also inflection points of different shapes and succession. We put a particular emphasis on evaluating the performance for these different types of settings and provide guidance on the choice of the method depending on the setting type taking tuning parameter options into account. Our empirical application to U.S. market sectors reveals that not only financial sectors, but also the real economy and, in particular, the commodity industries play an equally important role in understanding interaction effects.