Variational Inference of Dynamic Factor Models with Arbitrary Missing Data

Dynamic factor models are typically estimated by point-estimation methods, disregarding parameter uncertainty. We propose a new method accounting for parameter uncertainty by means of posterior approximation, using variational inference. Our approach allows for any arbitrary pattern of missing data, including different sample sizes and mixed frequencies. It also yields a straight-forward estimation algorithm absent of time-consuming simulation techniques. In empirical examples using both small and large models, we compare our method to full Bayesian estimation from MCMC-simulations. Generally, the approximation captures factor features and parameters well, with vast computational gains. The resulting predictive distributions are approximated to a very high precision, almost indistinguishable from MCMC both in and out of sample, in a tiny fraction of computational time.