VaR and ES forecasting via recurrent neural network-based stateful models

Abstract

Due to the widespread and quickly escalating effects of large negative returns, as well as due to the increase in the importance of regulatory framework for financial institutions, the accurate measurement of financial risks has become a relevant question in the academia and industry. This paper proposes three novel models based on stateful Recurrent Neural Networks (RNN) and Feed-Forward Neural Networks (FNN) to build forecasts for Value-at-Risk (VaR) and Expected Shortfall (ES). We apply the models to six asset return time series spanning over more than 20 years. Our results reveal that the RNN-based stateful models generally outperform the non-stateful RNN models and econometric benchmark models including rolling window models, Generalized AutoRegressive Conditional Heteroskedasticity (GARCH)-type models, and Generalized Autoregressive Score (GAS) models, in terms of VaR and ES forecasting.