实证研究中预测模型的选择:从逐步回归到信息标准

Forecasting Model Selection in Empirical ResearchFrom Stepwise Regression to Information Criteria

本文首先对显著性变量同变量显著性之间的关系予以讨论并区分,进而评价逐步回归模型选择法的缺陷性。在此基础上,我们对以AIC和B IC为代表的各种基于信息标准的模型选择法予以介绍和评论。同逐步回归法相比,信息标准模型选择法有着坚实的统计理论基础及清晰而优良的统计性质。本文通过基于近十年中国股市数据的实证检验说明,信息标准同逐步回归相比往往能产生具有更强预测能力的计量模型,因此值得在未来的实证研究中注意并推广。

We first differentiate two different concepts. They are, namely, significant variables and significance level. Then, the relative disadvantages of the widely used stepwise regression method are discussed,and the theory of information criteria as presented by AIC and BIC is introduced. As compared with stepwise regression,information criteria enjoy solid theoretical foundation and good statistical properties. A real data analysis is carried out based on ten years＇ Chinese stock market data. As it can be seen, the information criteria tend to produce models with better predictivability. Therefore,it should be suggested for future empirical sutdy.