Speaker：涂云东 (Yundong Tu)

　　Title：A Tale of Two Types of Structural Instabilities in High Dimensional Factor Models

　　Schedule: March 20, Mon 1:30-3:00 

　　Location: 诚明楼（Chengming Hall） RM 315 

　　Introduction：涂云东，美国加州大学河滨分校经济学博士，北京大学光华管理学院商务统计与经济计量系和北京大学统计科学中心联席教授，研究员。入选首批“日出东方”北大光华青年人才，教育部“长江学者奖励计划”青年长江学者，两次获评北京大学优秀博士学位论文指导教师。已有三十余篇学术论文发表在国际著名期刊 Journal of Econometrics，Econometric Reviews, Journal of Business and Economic Statistics，Oxford Bulletin of Economics and Statistics，Statistica Sinica，Journal of Empirical Finance，Journal of Management Science and Engineering，Computational Statistics and Data Analysis 。著作教材《时间序列分析》由人民邮电出版社于2022年9月出版。理论研究领域涵盖非参数/半参数计量经济模型，模型选择和模型平均，网络数据建模，金融计量，信息计量经济学，模型设定检验等；应用研究包含宏观经济预测，价格指数建模，网络数据分析，股票市场预测，生产率建模等。

　　Abstract：With the increasing availability of large data sets in economics and finance, the large factor model has become one of the most important tools to achieve dimension reduction in the statistical and econometric analysis. To capture the instability caused by economic condition shifts or policy reforms, factor models with structural breaks in the factor loadings are accordingly developed. On the other hand, recurring regime shifts that relate to higher frequency recurring fluctuation arise in situation where “history repeats”, and are conveniently described by threshold factor models, which allow recurring regime shifts in the factor loadings according to the magnitude of a (continuous) threshold variable. In practice, it is often difficult to decide whether structural break or threshold effect, or both types of instabilities one should employ to portray the observed data. This talk shall discuss how to model each type of instability in factor analysis separately first, and then provide a solution to distinguish the two categories in a model that simultaneously allows both types of structural instabilities. The proposed models are estimated by machine learning techniques such as group Lasso, backward elimination algorithms and information criterion-based model selection methods. The associated asymptotic properties are established and are corroborated by finite sample simulation results and empirical examples.