Identification of time-varying factor models

With the increasing availability of large datasets that span a long horizon of time, the assumption of constant loadings in conventional factor models is no longer realistic. The time-varying factor model (TVFM) is seen as a potential solution and has attracted a lot of interests from the literature. However, TVFM also suffers from the well-known problem of non-identifiability. This paper considers the situation under which both the factors and the loadings can be consistently estimated without rotations. Asymptotic distributions of the proposed estimators are derived. Theoretical findings are supported by simulations. Finally, we apply the proposed procedure to a large dataset of Canadian macroeconomic variables. We find substantive variations in the factor loadings that can be easily interpreted.