报告题目：Shattering Break Barriers: Inferential Theory for Group Factor Models Subject to Disruptive Breaks

报告人：涂云东教授

时间：10月22日（周三）下午3:00-4:00

地点：创新港涵英楼经济金融研究院8004室榆林研究院

报告人简介：

涂云东，北京大学博雅特聘教授，联合受聘于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。入选“日出东方”北大光华青年人才，北京大学优秀博士学位论文指导教师（2017,2021,2024），北京大学优秀研究生导师（2024），国家级青年人才计划入选者，国家杰出青年科学基金获得者。先后获武汉大学理学学士学位（2004）和经济学硕士学位（2006）、加州大学经济学博士学位（2012，河滨分校）。环亚太青年计量经济学者（YEAP）会议发起人和主要组织者。50余篇学术论文发表在多个国际国内知名专业杂志。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。研究领域涵盖时间序列分析、非参数计量方法、大数据分析、金融计量和预测等。

摘要：

The homogeneous group structure in factor analysis is widely recognized for streamlining the clustering of high-dimensional time series, a property that directly enhances estimation precision. This study focuses on a grouped factor model where two critical components---latent group memberships and factor loadings---experience abrupt structural changes often triggered by real-world disruptions, including technological advancements, economic or financial crises, and policy adjustments. A critical finding of this research is that neglecting structural instability leads to three severe inferential issues: inconsistent recovery of the factor space, overestimation of the number of factors, and overestimation of the number of groups. To address these challenges and `shatter the break barriers", we propose a four-step procedure tailored to achieve consistent and efficient estimation. We first identify a `pseudo group structure" and estimate corresponding pseudo factors within a pseudo linear factor model---one that temporarily sets aside structural breaks. We next detect structural break points using shrinkage techniques applied to the pseudo factor regression. Finally, we sequentially merge pseudo groups in an iterative manner, halting only when no further improvement in model fit is achieved as evaluated using an information criterion. Theoretical properties of the proposed procedure, including consistency in break point estimation, accurate recovery of the true group structure, and reliable inference for the factor space, are established. These results are corroborated with both Monte Carlo simulations and an application to studying U.S. macroeconomic datasets that demonstrate the superior performance of our procedure over alternative approaches in high-dimensional time series analysis.