Against the backdrop of increasing uncertainty in global financial markets and frequent extreme financial events, traditional risk prediction models face significant accuracy bottlenecks due to the scarcity of extreme event samples. This study centers on generative artificial intelligence (AI) technology to construct a dual-driven extreme risk prediction framework integrating “real data + generated data”, exploring the application value of the improved WGAN-GP model in generating extreme scenario data. Using daily data of China’s CSI 300 Index (January 2015–June 2025), regulatory penalty texts, and macroeconomic data as research samples, this paper compares the predictive performance of three models—EVT-GARCH, LSTM with single real data, and LSTM with fused data—through a three-stage “generator-filter-predictor” model architecture. Empirical results demonstrate that: (1) Extreme scenario data generated by the improved WGAN-GP model can accurately reproduce the statistical characteristics and dynamic structure of financial markets; (2) The fused-data LSTM model outperforms traditional models significantly, with a VaR prediction error of 9.4%, an ES prediction error of 12.7%, and an extreme event capture rate of 91.2%; (3) Incorporating regulatory text information extends the early warning period for extreme risks to 3–5 trading days. This study breaks the traditional paradigm of “relying solely on real data” and provides a new methodology for financial institutions’ risk prevention and regulatory authorities’ systemic risk monitoring.