【摘要】 时间序列预测在多个学科中被广泛运用，考虑到时间序列与被预测目标间的动态相关性对预测结果可能产生的影响，在XGBoost算法基础上设计出一种依据动态相关性的大小自动优化输入特征的GAS-Copula-XGBoost模型。引入能够衡量反向相关性的半旋转Clayton Copula和半旋转Gumbel Copula函数，利用GAS演化方程将半旋转Copula函数与常用的Copula函数拓展为时变Copula函数，对输入因子与被预测目标间的动态相关性进行测度；根据动态相关性设定阈值并设计动态程序，利用XGBoost算法预测，将模型应用于“一带一路”主题指数预测研究。结果表明：部分数据组合的相关性由半旋转Copula函数描述更为准确，GAS-Copula-XGBoost模型在分类预测精度上较Logistics、随机森林和XGBoost均有提升；在回归预测上，误差较BP神经网络、SVR和XGBoost分别降低37.875 3%、17.486 5%和5.361 2%。更多还原

【Abstract】 Time series prediction is widely used in many disciplines. Considering the possible impact of the dynamic correlation between time series and the predicted target on the prediction results, on the basis of XGBoost algorithm, a GAS-Copula-XGBoost model is designed to optimize input characteristics automatically based on the dynamic correlation. The semi rotating Clayton and Gumbel Copula functions which can measure the reverse correlation are introduced and extended to the time-varying form with the commonly used Copula functions together by using GAS model. Then, the dynamic correlation between the input factors and the predicted target is measured. According to the value of the dynamic correlation, the threshold is set and the dynamic program is designed. Then CSI One Belt& One Road Index is used to predict the theme. The empirical results show that the correlations of some data pairs are more accurately described by semi rotating Copula function. Compared with logistics, random forest and XGBoost, the accuracy of GAS-Copula-XGBoost model in classification prediction is improved; in regression prediction, error is 37.875 3%, 17.486 5% and 5.361 2%, lower than BP neural network, SVR and XGBoost, respectively.更多还原