Volatilitetsprognoser med grafinlärning : Ett heterogent grafneuronnätsbaserat angreppssätt för prognostisering av realiserad volatilitet

This thesis explores the use of Heterogeneous Graph Attention Networks (HGATs) for predicting realized volatility (RV), presenting a novel approach that leverages the relational learning capabilities inherent in graph-based models. Specifically, the study investigates how structuring data as graphs can enhance both predictive performance and interpretability. Two graph-based models are proposed: the Dynamic Heterogeneous Graph Attention Network (DHGAT) and the Interpretable Heterogeneous Graph Attention Network (IHGAT). These models are evaluated against state-of-the-art time series forecasting methods, including Long Short-Term Memory (LSTM), Random Forest (RF), and the Heterogeneous Autoregressive Model (HAR-RV). Experimental results across 16 datasets demonstrate that the proposed DHGAT model consistently outperforms all non-graph models in predictive accuracy when considering overall performance across datasets, as confirmed by statistical testing. This suggests that incorporating relational dependencies through graph structures improves RV forecasting. The interpretable IHGAT model achieves predictive performance on par with LSTM and HAR-RV, while significantly outperforming RF. Moreover, no statistically significant reduction in accuracy is observed between DHGAT and IHGAT, indicating that enhancing model interpretability does not come at the expense of forecast quality.