A novel HAR-type realized volatility forecasting model using graph neural network

This study introduces a novel model that uses the convolutional neural network (CNN) architecture to fully utilize information from the heterogeneous autoregressive (HAR) family components across different windows in a two-dimensional image format, to forecast the direction of stock market volatility. The proposed model, Convolutional Neural Network-based Heterogeneous Autoregressive-Kitchen Sink (CNN-HAR-KS) model, leverages the CNN model's automated signal generation capabilities while incorporating the HAR components to ensure interpretability and comparability with traditional volatility forecasting models. Empirical analysis in China's stock market reveals that in terms of forecast accuracy, the CNN-HAR-KS model outperforms alternative models, such as logistic or machine learning models, using the same input variables. We construct investment portfolios based on realized volatility forecasts to further explore the CNN-HAR-KS model's economic value. The CNN-HAR-KS model yields the highest daily Sharpe ratio of 0.043, indicating its ability to generate better risk- adjusted returns than alternative models. Further analysis suggests that the proposed model can be applied to other realized volatility-related classification problems.