Title: Frequency-domain cross-validation for determining the number of the common factors in factor models
Authors: Natalia Sirotko-Sibirskaya - University of Bremen (Germany) [presenting]
Abstract: A frequency-domain-based cross-validation (FDCV) criterion is proposed to determine the number of common factors driving the observable multivariate data process with respect to the appropriately defined loss function. The suggested method is based on the theoretical property of the coefficients of the Fourier transforms which are known to be approximately independent under certain conditions. An alternative scheme for dealing with missing values in cross-validation is suggested under the assumption that the empirical Fourier transforms of the time series are smooth functions of frequency. Properties of the proposed criterion are studied both at the theoretical level and in simulations. The performance of the method is tested on the U. S. macroeconomic and financial data and compared with other commonly used criteria on determining the number of common factors in static and dynamic factor models.