基于GAS-CE-LGBM的“一带一路”指数收益率预测研究
Research on the Yield Prediction of the “Belt and Road” Index Based on GAS-CE-LGBM

研究“一带一路”指数收益率有助于投资者和政策制定者更好地理解和规划“一带一路”倡议相关的金融市场趋势，以支持有效的投资决策和制定经济政策，但由于其复杂性和非线性特征，传统的预测方法可能无法充分捕捉其动态变化。为了解决这一问题，本文提出了一种结合广义自回归得分(Generalized Autoregressive Score, GAS)模型、Copula熵(Copula Entropy, CE)特征选择和监督学习集成模型——轻量梯度提升机(Light Gradient Boosting Machine, LightGBM)模型的综合预测框架(Generalized Autoregressive Score-Copula Entropy-Light Gradient Boosting Machine, GAS-CE-LGBM)。首先，构建“一带一路”指数收益率的GAS波动率模型并估计参数；其次，计算“一带一路”指数及其成分股相应的Copula熵，并通过阈值进行筛选；最后，将所得成分股信息与GAS模型参数构成数据集输入LightGBM模型中建模预测。实验结果表明，GAS-CE-LGBM模型相较多层感知器神经网络(Multilayer Perceptron, MLP)、LightGBM、GARCH-LGBM (Generalized Autoregressive Conditional Heteroskedasticity-Light Gradient Boosting Machine)和GAS-LGBM (Generalized Autoregressive Score-Light Gradient Boosting Machine)模型在RMSE、MAE、MAPE和R²四个评估指标上表现最佳，RMSE、MAE和MAPE分别平均降低了19.09%、19.81%、62.48%，R²平均提高了12.05%。这表明该模型在“一带一路”指数的预测方面展现了良好的性能和潜力，能更好地捕捉到“一带一路”指数收益率的动态变化。
Studying the returns of the “Belt and Road” index contributes to a better understanding and planning for investors and policymakers regarding the financial market trends associated with the “Belt and Road” Initiative. This understanding supports effective investment decisions and economic policy formulation. However, due to its complexity and non-linear characteristics, traditional forecasting methods might not adequately capture its dynamic changes. To address this issue, this paper proposes a comprehensive predictive framework, the Generalized Autoregressive Score-Copula Entropy-Light Gradient Boosting Machine (GAS-CE-LGBM) model, which combines the Generalized Autoregressive Score (GAS) model, Copula Entropy (CE) feature selection and supervised learning ensemble model—Light Gradient Boosting Machine (LightGBM). First, build the volatility GAS model of the return rate of the “Belt and Road” index and estimate the parameters. Secondly, calculate the corresponding Copula entropy of the “Belt and Road” index and its constituent stocks and filter through the threshold. Finally, input the data set of constituent stock information and GAS model parameters into the LightGBM model for modeling and forecasting. Experimental results demonstrate that the GAS-CE-LGBM model outperforms Multilayer Perceptron (MLP), LightGBM, GARCH-LGBM (Generalized Autoregressive Conditional Heteroskedasticity-Light Gradient Boosting Machine), and GAS-LGBM (Generalized