报告题目：Supervised Expectation-Maximization method

报 告 人：李鲲鹏 教授

主 持 人：李仲飞 教授

时    间：10月22日下午14:45-15:30

地    址：中山大学岭南堂三楼陈荣捷讲学厅

摘要:

    The partial least square (PLS) method gains growing popularity in empirical asset pricing literature. Compared with the traditional dimensional reduction method such as the principal components analysis (PCA), the PLS uses the predicting target to supervise the direction of dimension reduction. The literature has long believed that the PLS can consistently and efficiently recover the relevant factors. In this paper, we show that this understanding is incorrect. The consistency of the PLS relies critically on what we call two-way orthogonal conditions which generally do not hold due to the fact that the relevant factor can be identified in the model. Due to failure of the PLS method, we therefore propose a new method, that is called supervised expectation-maximization method (SEM), to recover the relevant factors. In the SEM, zero restrictions are exploited to supervise the direction of dimension reduction. We establish the asymptotic properties of the SEM estimator. We run simulations to investigate the performance of the SEM as well as the PCA and the PLS. The simulation results present the bad performance of the PLS and the PCA, indicating that these two methods cannot extract the right direction even with the guidance of the predicting target. Our SEM method, as a comparison, works very well in the same scenario. We apply our method to the application of the extra return of bonds and find better forecasting performance.

​

报告人介绍：    

  李鲲鹏现为首都经贸大学国际经管学院教授、院长，国家级人才项目获得者，主要研究领域是数据降维方法、面板交互效应以及空间网络模型，在经济学、管理学、统计学国际高水平期刊上发表论文30余篇，主持国家自然科学基金4项，北京市教委高水平科研团队项目1项，目前担任Journal of the American Statistical Association和Journal of Business & Economic Statistics期刊的副主编，也是国内多个重要学术团体的副理事长或者常务理事。