A1941
Title: Robust imputation for high-dimensional panel with a latent factor structure
Authors: Yiming Wang - Shanghai University of Finance and Economics (China) [presenting]
Long Yu - Shanghai University of Finance and Economics (China)
Abstract: Factor models are widely used in economics and finance to capture common variations in high-dimensional panel data. However, real-world data often suffer from heavy-tailed errors and missing observations, which can severely compromise the performance of conventional Principal Component Analysis (PCA). To address these challenges, a robust estimation framework for high-dimensional factor models with random missing data is proposed. Motivated by the equivalence between PCA and constrained least squares in the complete-data factor model, an element-wise Huber loss function defined only on the observed entries (i.e., where $W_{it}=1$) is minimized, which is solved by an iterative Huber regression algorithm. Theoretical convergence rates for the minimizer of the objective function are established and their asymptotic normality is derived under mild moment conditions. Furthermore, three criteria for robustly determining the number of factors are proposed. Extensive simulations and a real macroeconomic dataset (FRED-QD) demonstrate that the proposed method significantly outperforms existing approaches in imputation and prediction under various random missing mechanisms and heavy-tailed settings.
