Statistics seminar at CBS: Yoshimasa Uematsu

Speaker Yoshimasa Uematsu, Hitotsubashi University

Time and place Monday, November 23, 11:00. Copenhagen Business School (Solbjerg Pl. 3, 2000 Frederiksberg), room SPs03

Title Whittle spectral knockoffs for band-wise discoveries

Abstract

We propose a framework, called Whittle spectral knockoffs, for variable selection across frequency bands in stationary multivariate time series with false discovery rate (FDR) control. For each band, the framework identifies covariate series that are linearly associated with a response, after adjusting for the other series. In general, temporal dependence makes it challenging to construct knockoffs with the required exchangeability guarantees when the covariate distribution is unknown. We address this using a Whittle Gaussian working law for Fourier coefficients, which permits frequency-wise construction of exactly exchangeable oracle knockoffs. A feasible implementation combines factor modeling with regularized spectral estimation of the residual component. Under Gaussian assumptions, we derive approximate FDR bounds under the true data-generating law by controlling frequency-domain approximation and spectral estimation errors. We also investigate score constructions designed to improve power under strong covariate dependence. Simulations and semi-synthetic experiments based on U.S. macroeconomic data assess FDR and power.