Research on a Quantitative Financial Risk Prediction Model Based on Machine Learning

Abstract

To address the challenges faced by traditional financial risk prediction models in balancing volatility characterization and nonlinear dependency capture, this study proposes an Att-LSTM-GARCH hybrid model, integrating the volatility extraction capabilities of GARCH, temporal learning of LSTM, and key feature enhancement capabilities of the attention mechanism. Using daily data (2,428 samples) from the CSI 300 and CSI 500 indices from 2015 to 2024, prediction experiments were conducted focusing on volatility and 95% confidence level VaR. Results show that the MAE for one-day volatility prediction reaches 0.012, which is 32.2%, 22.5%, and 14.3% lower than those of GARCH, LSTM, and LSTM-GARCH, respectively. The one-day VaR prediction accuracy reaches 92.3%, outperforming the comparison model by up to 4.1%. Ablation experiments demonstrate that the GARCH and attention mechanisms can reduce the MAE by 25% and 33.3%, respectively. In volatile bull and bear markets, the MAE fluctuates by only 16.7%, demonstrating optimal stability under noise. This model provides financial institutions with a high-precision risk prediction tool, enhancing risk management.