A1547
Title: Factor-adjusted knockoff inference for high-dimensional time series
Authors: Jie Wei - Huazhong Univ. of ST (China) [presenting]
Abstract: A new knockoff inference framework for high-dimensional time series data, termed FAR-knockoff, is proposed, which incorporates factor adjustment and accounts for serial dependence. Unlike most knockoff methods that require temporal independence on covariates or errors, FAR-knockoff leverages a weighted principal component analysis to estimate latent factors from possibly serially and cross-sectionally correlated covariates, followed by Fixed-X knockoff construction for the idiosyncratic components. This design improves power by reducing collinearity while preserving theoretical validity. Rigorous finite-sample and asymptotic guarantees for false discovery rate (FDR) control are established, showing robustness against estimation error in covariance structures. Simulation studies demonstrate substantial gains in power and FDR stability compared to existing approaches. Applying FAR-knockoff to U.S. Treasury bond risk premia prediction, improved out-of-sample performance and meaningful insights into maturity-dependent predictive structures are found. The method advances knockoff inference for dependent data environments common in many areas.
