A0772
Title: Multivariate volatility forecasting through extended DCC-GARCH and relational graph convolutional network
Authors: Yasumasa Matsuda - Tohoku University (Japan)
Youjia Liu - Tohoku University (Japan) [presenting]
Abstract: Accurate multivariate volatility forecasting is vital in finance for portfolio optimization, risk management, and asset pricing. Traditional multivariate GARCH models are widely used but often struggle with capturing stock interrelationships and scaling to high-dimensional data. To address this, a new model is proposed combining DCC-GARCH with a modified relational graph convolutional network (RGCN) that includes edge-level normalization to better capture complex dependencies. Two versions are tested: One using return data only and another combining return and range data. Results show the model consistently outperforms traditional methods, especially during periods of high market volatility.
