Does multi-scale GARCH information enhance volatility prediction?

Abstract

The GARCH-LSTM model is widely used for volatility prediction. However, it fails to account for feature dependencies across different frequencies, and the lack of an attention mechanism limits its predictive accuracy. To address these issues, this paper introduces the GENSHIN framework, which integrates GARCH information with a multi-scale graph neural network to model volatility across multiple frequency bands with MixHop and multi-head attention mechanism. Empirical studies on the volatility of major Chinese stock indices demonstrate that GENSHIN outperforms other deep learning models, such as TimesNet and LSTM. These results highlight the effectiveness of multi-scale information in improving volatility forecasting.