Enhancing Financial Tail Risk Forecasting: A Blending Ensemble Framework for Nonlinear Expectile Regression

As a prominent tool for tail risk measurement, expectile‐based value at risk (EVaR) has attracted growing interest due to its sensitivity to extreme risks. Existing approaches face two principal challenges: the inefficiency of conventional models under finite samples and the tendency of single machine learning models toward overfitting or underfitting. This paper introduces a nonlinear expectile regression model based on a blending ensemble framework, integrating neural networks, support vector machines, XGBoost, LightGBM, and random forests as base learners, with an expectile regression forest as the metalearner. Monte Carlo simulations confirm the method's robustness in finite samples. Applied to Chinese stock indices, the model outperforms both traditional linear specifications and each individual machine learning model in EVaR forecasting. Performance gains are statistically significant under stochastic volatility and TGARCH settings, as verified by Diebold–Mariano and Giacomini–White tests. SHAP analysis further shows that XGBoost and LightGBM contribute most to prediction, enhancing interpretability and offering insight into ensemble decision mechanisms.