High-dimensional Multivariate Realized Volatility Forecasting with Community Network Structure

We introduce the Community Network Heterogeneous Autoregressive model with Quarticity (CNHARQ), a novel framework that integrates network-based information to enhance the forecasting of multivariate realized volatilities. To address the curse of dimensionality, we propose a Correlation-Based Stochastic Block Model (CBSBM) to uncover latent community structures from the correlation network of realized volatilities of N assets. This approach reduces the number of unknown parameters in the model from O(N2) to O(NK), where K≪N denotes the number of communities. Empirical analysis demonstrates that the CBSBM captures dynamic community structures, revealing shifts in the co-movement of asset volatilities over time. Furthermore, the CBSBM-based community structure outperforms the conventional Global Industry Classification Standard (GICS) in out-of-sample volatility forecasting, highlighting the superior forecasting power of network-based correlations over traditional industry classification schemes.