Abstract: As the global economy has faced numerous challenges and increased uncertainty in recent years, effectively predicting financial market risk has become crucial for understanding global financial trends and managing risks. This paper introduces an Att-LSTM model, based on Self-Attention mechanisms and Long Short-Term Memory neural networks (LSTM), which employs panic indexes from 15 international financial markets as input features to deeply investigate its performance in forecasting the volatility of five major stock indexes. Furthermore, a global investment strategy based on volatility forecasting was constructed. The research findings demonstrate that the Att-LSTM model outperforms ten other benchmark methods, including mainstream machine learning algorithms and OLS methods, in volatility forecasting. Simulated trading results indicate that the model achieves significant investment returns under low-risk conditions and is sensitive to extreme events. Interpretability analysis further reveals the predominance of the US market's panic sentiment in risk prediction across countries, as well as the importance of other key predictors, such as panic indexes from China, China (Hong Kong), and the international oil market in volatility forecasting. Replicate experiments validated the robustness of the interpretability analysis findings, emphasizing the potential applications of machine learning in financial forecasting and risk management. This provides substantial insights and guidance for policymakers and financial practitioners.

Key words: financial market risk, self-attention mechanisms, long short-term memory neural networks, volatility forecasting