Deep Learning and High-Frequency Data Enhanced Bayesian Financial Risk Forecasting

Abstract

This thesis, titled "Deep Learning and High-Frequency Data Enhanced Bayesian Financial Risk Forecasting," presents an innovative approach in forecasting financial tail risks, particularly focusing on Value-at-Risk (VaR) and Expected Shortfall (ES) through a Bayesian framework. The first part of this thesis introduces a semi-parametric framework for joint VaR and ES forecasting, incorporating multiple realized measures such as realized variance, bi-power variation, and realized kernel. Extending the realized exponential GARCH model, this framework employs a quasi-likelihood function based on the asymmetric Laplace distribution, enabling Bayesian estimation through adaptive Markov Chain Monte Carlo methods. The second part advances risk forecasting further by proposing a long-memory and non-linear model, termed RNN-HAR, for direct VaR prediction. This model integrates a Recurrent Neural Network (RNN) with the heterogeneous autoregressive (HAR) model to capture long-memory effects and non-linear dependencies within realized volatility measures. Using loss-based Bayesian Sequential Monte Carlo (SMC) for inference, the RNN-HAR model demonstrates superior predictive performance outperforming conventional HAR models in VaR forecasting. The final part of the thesis proposes a novel RNN-based framework, RNN-ReESCAViaR, for the joint estimation of VaR and ES. By embedding an RNN within the Realized-ES-CAViaR (Conditional Autoregressive VaR) model and employing an asymmetric Laplace quasi-likelihood function, this framework captures the interdependence between VaR and ES in a highly flexible manner. Bayesian SMC is again utilized, enhancing computational efficiency and allowing for real-time sequential prediction. Empirical results show that RNN-ReESCAViaR achieves significant improvements over both parametric and semi-parametric models, illustrating its capacity to adapt to complex volatility patterns in financial data.