金融理论与教学

针对传统GARCH-VaR模型尾部风险低估及分布僵化问题，提出基于Hurst指数分类的机器学习优化框架。通过Hurst指数将序列分为趋势反转、中等持续与强趋势三类，分别构建XGBoostGARCH、Prophet-SVR及混合模型。在模型优化成功的案例中，优化模型在极端压力下的VaR违反率平均降低10.16%,MSE的最大降幅达到93.6%。XGBoost-GARCH借助波动率跨期传导机制，显著提升尾部风险捕捉能力；Prophet-SVR因趋势分解弱化波动特征，导致压力测试违反率从6.67%达到38.78%。SHAP分析揭示条件波动率特征是模型优化的核心驱动因素，其重要性占比超75%，为市场风险的动态预测和模型透明化决策提供方法论支持。

In response to the underestimation of tail risk and the rigidity of distribution in traditional GARCH-VaR model, this paper proposes a machine learning optimization framework on the basis of Hurst index classification. The series are categorized into three types-trend-reversal, moderate persistence, and strong trend by virtue of the Hurst index. Corresponding models, namely XGBoost-GARCH, Prophet-SVR,and a hybrid model, are constructed accordingly. Among the successfully optimized cases, the optimized models achieve an average reduction of 10.16% in VaR violation rates under extreme stress, with a maximum reduction in MSE of 93.6%. The XGBoost-GARCH model significantly enhances tail risk capture by leveraging the intertemporal transmission mechanism of volatility. In contrast, the Prophet-SVR model, due to its trend decomposition that weakens volatility characteristics, leads to an increase in stress test violation rates from 6.67% to 38.78%. SHAP analysis reveals that conditional volatility features are the core drivers of model optimization, with the importance accounting for more than 75%. This study provides methodological support for dynamic market risk forecasting and transparent model decision-making.