Abstract

We define Locally Stationary Factor models as Factor Models modulated by loadings that vary slowly over time. This specification allows to model non-stationary multivariate time series that are driven by common factors.
Principal Component Analysis (PCA) of the sample covariance matrix is a known technique for estimating the loadings of classical Factor Models. We show that local PCA, based on nonparametric estimation of the time-varying covariance matrix, is a consistent estimator of locally stationary factor models as the sample size and the dimension of the series increase.

This is a joint work with Christian Hafner and Rainer von Sachs.