Large Dimensional Time-Varying GMM Estimation : A New Approach

We develop a novel machine learning method to estimate large dimensional time-varying GMM models via our newly designed ridge fusion regularization scheme. Our method is a one-step procedure and allows for abrupt, smooth and dual type time variation with a fast rate of convergence. It outperforms state-of-the-art smoothed nonparametric methods for both small and large samples. We investigate the pricing ability of dynamic four-factor and consumption-based asset pricing models. We show that a dynamic SDF obtained by our method, which allows a rich set of conditioning variables to jointly determine time variation, has a significantly enhanced pricing ability while enjoying reduced time series errors