A Multi-Scale Prediction Model for Stock Volatility Based on a Hybrid Attention Mechanism

This study proposes a multi-scale attention CNN–BiLSTM model to predict stock market volatility. The model combines CNN layers to capture short-term movements, BiLSTM layers to learn long-term patterns, and a multi-scale attention unit to adjust feature importance across different time spans. Daily data from the Chinese A-share and NASDAQ 100 markets from 2015 to 2024 were used for testing. The results show that the model reduced RMSE by 9.4% compared with the LightGBM baseline and performed better during high-volatility periods. These results suggest that using information from multiple time scales improves forecasting accuracy and model stability. The method can support risk management and investment decisions. Future work will apply the model to higher-frequency data and improve attention interpretation for real-time market use.