可变权重的变系数分位数回归模型平均

Variable Coe?icient Quantile Regression Model Averaging for Variable Weights

摘要: 为了提高模型平均估计量的预测精度，基于分位数回归提出了可变权重的变系数分位数回归模型平均估计量的求取方法。首先，利用分位数回归函数计算出变系数回归函数的Jackknife估计量；接着，将得到的估计量代入分位数回归模型中得到第m个模型下解释变量的条件分位数的Jackknife估计值；最后，通过最小化局部Jackknife准则来取得权重估计量。数值模拟显示，所提估计量比传统的均值回归估计量更加稳健，同时所提可变权重估计量的预测精度也明显优于固定权重估计量的预测精度。最后，利用Boston房价信息进行实证分析，实证结果也进一步证明了所提估计量的优良性。数值模拟和实际数据分析证明了该方法具有更高的预测精度。

Abstract: In order to improve the prediction accuracy of the estimators of the model averaging, we propose a method for finding the estimators of variable coefficient model averaging with variable weights based on quantile regression. First, the quantile regression function is used to calculate the Jack-knife estimated value of the variable coefficient regression function. Then, the obtained estimators are substituted into the quantile regression model to obtain the Jackknife estimate of the condi-tional quantile of the explanatory variable under the mth model; finally, the weight estimators are obtained by minimizing the local Jackknife criterion. The simulations show that the estimators proposed by us are more robust than the traditional conditional mean regression estimators, and the prediction accuracy of the variable weight estimators proposed by us is significantly better than that of the fixed weight estimators. We use Boston house price information for empirical analysis, whose results highlight the merits of the estimators proposed by us. The simulation and empirical analysis prove that the proposed method has higher prediction accuracy.