Machine learning methods for financial risk measurement

This thesis explores the enhancement of forecasting methodologies for market risk measures, specifically Value-at-Risk (VaR) and Expected Shortfall (ES), with a particular focus on machine learning techniques. The first contribution is that it proposes three novel models based on stateful Recurrent Neural Networks (RNN) and Feed-Forward Neural Networks (FNN) for forecasting VaR and ES. Applied to several asset returns, the findings show that stateful RNN models generally outperform non-stateful RNN models and traditional econometric approaches, such as rolling window models, Generalized AutoRegressive Conditional Heteroskedasticity (GARCH)-type models, and Generalized Autoregressive Score (GAS) models, in terms of forecasting accuracy. This contribution highlights the potential of neural network architectures in improving financial risk measurement and prediction. Secondly, this thesis introduces an innovative method for improving VaR and ES predictions by combining the bagging technique with Deep Feed-forward Neural Networks (DFNN). By enhancing predictive accuracy and reducing overfitting, this approach addresses financial risk forecasting challenges in volatile markets. Simulations and empirical analysis show that bagged DFNN models consistently outperform standalone DFNN models by incorporating diverse sources of uncertainty. Applied to fifteen asset returns, the proposed method surpasses individual forecasting models and other combination techniques, demonstrating its effectiveness in enhancing financial risk prediction. Thirdly, this thesis contributes to financial risk forecasting by incorporating public and institutional attention to climate change into VaR and ES predictions using machine learning models. Climate attention is measured through two proxies: the Google Search Volume Index (SVI), which captures public interest, and Bloomberg’s news trends (NT) function, which reflects institutional investors’ focus on climate change. By integrating these variables into neural network models, we aim to improve the accuracy and robustness of financial risk forecasts, particularly in the face of growing climate related uncertainties. The findings indicate that incorporating climate attention proxies improves VaR and ES forecasts.