A1477
Title: Group network multivariate GARCH
Authors: Jian Chen - University of Sussex (United Kingdom) [presenting]
Ganggang Xu - University of Miami (United States)
Abstract: Traditional multivariate generalised autoregressive conditional heteroskedasticity (GARCH) models (e.g., BEKK, DCC model) often suffer from the curse of dimensionality. A group network multivariate GARCH model is proposed in which the transitions of past variance and return shocks among assets are subject to an adjacency matrix and a latent group structure. This approach significantly reduces the number of parameters in high dimensions, thus facilitating estimation and forecasting. The theoretical properties of an estimator are developed that uses an optimisation algorithm estimating parameters and group memberships simultaneously. Simulation results confirm our theoretical findings. An empirical analysis is conducted on the S&P 100 constituents from 2015 to 2022 and is shown that the model improves portfolio selection in out-of-sample forecasts compared to other models.
