Cross-market volatility forecasting with attention-based spatial–temporal graph convolutional networks

We propose a cross-market volatility forecasting framework by applying attention-based spatial–temporal graph convolutional network model (ASTGCN) to forecast future volatility of stock indices in 18 financial markets. In our work, we construct cross-market volatility networks to integrate interrelations among financial markets and the corresponding features of each market. ASTGCN combines the spatial–temporal attention mechanisms with the spatial–temporal convolutions to simultaneously capture the dynamic spatial–temporal characteristics of global volatility data. Compared with competitive models, ASTGCN exhibits superiority in multivariate predictive accuracies under multiple forecasting horizons. Our proposed framework demonstrates outstanding stability through several robustness checks. We also inspect the training process of ASTGCN by extracting spatial attention matrices and find that interrelations among global financial markets perform differently in tranquil and turmoil periods. Our study levitates empirical findings in financial networks to practical application with a novel forecasting method in the deep learning community.