A0743
Title: CSD-robust bootstrap prediction intervals for factor-augmented regression
Authors: Feifan Wang - University of Connecticut (United States) [presenting]
Jungbin Hwang - University of Connecticut (United States)
Abstract: A bootstrap method is developed that accounts for cross-sectional dependence in the idiosyncratic errors and constructs robust prediction intervals for factor-augmented regression models. Factor-augmented regression models are widely used in macroeconomic and financial forecasting and rely on latent factors extracted from large panel datasets. Standard asymptotic methods often ignore the asymptotic or finite-sample bias in such models and rely on restrictive Gaussian assumptions. Existing bootstrap approaches typically assume independence across cross-sectional units or focus only on inference for regression coefficients rather than constructing prediction intervals. The proposed method employs a hard-thresholding estimator of the idiosyncratic error covariance matrix to generate bootstrap samples that preserve cross-sectional dependence. It enables valid prediction intervals for both the future target value and its conditional mean without requiring normality of the regression errors. Simulation results show that the method improves coverage accuracy under dependence structure and various sample sizes. Theoretical justification is provided for the consistency of the intervals. The proposed approach will be illustrated through empirical examples in financial forecasting contexts.
