统计与管理学院2018年学术报告第22期

【主 题】 IPAD: A Factor Approach to High-Dimensional Knockoffs Inference

【报告人】 Yingying Fan 副教授

University of Southern California

【时 间】 2018年06月12日（星期二）16:10-17:00

【地 点】 上海财经大学统计与管理学院大楼1208会议室

【摘 要】 This paper considers the problem of false discovery rate (FDR) control for variable selection in high-dimensional factor models, where the association structure of covariates is modeled using a latent factor model popularly exploited in economics and finance. To achieve the FDR control, we adapt the general framework of model-X knockoffs in Candes et al. (2017) and suggest the new method of intertwined probabilistic factors decoupling (IPAD) for knockoffs inference in high-dimensional factor models. Our new method and work differ from the existing literature in at least four aspects: 1) when constructing the knockoff variables, we estimate the covariate distribution from data while Candes et al. (2017) assumed that it is known; 2) our procedure does not need any sample splitting; 3) we provide theoretical justifications on the asymptotic FDR control; and 4) we establish theory for the power analysis. Our simulation examples and real data analysis also demonstrate that our newly suggested method has appealing finite-sample performance with desired interpretability compared to some widely used methods such as the Lasso and factor autoregressive models.