Machine learning models to predict stock market spillovers

We propose a set of machine learning models, based on recurrent neural networks, for the prediction of stock market spillovers. While classic spillover models, such as the Diebold–Yilmaz approach, can explain which are the spillover effects, we aim to predict them, and provide an early warning system. To assess the effectiveness of our proposal, we compare our predictions to the actual return and volatility spillovers across fourteen major equity market indices, spanning the period from January 2000 through January 2024, and considering two hundred rolling window test samples. Our empirical findings show that the predictions are quite accurate, and that the Gate Recurrent Unit network consistently outperforms the other models, primarily due to its ability to capture complex and non-linear dependencies.