Forecasting US Commercial Property Price Indexes Using Dynamic Factor Models

The general purpose of a dynamic factor model (DFM) is to summarize a large number of time series into a few common factors. Here we explore a number of DFM specifications applied to 80 granular, non-overlapping indexes of commercial property prices in the US, quarterly from 2001 to 2017. We examine the nature and the structure of the factors and the index forecasts that can be produced using the DFMs. We consider specifications of 1, 2, 3 and 4 common factor trends. As a major motivation for the use of DFMs is their ability to improve out-of-sample forecasting of systems of numerous related series, we apply the DFM estimated factor returns in an Autoregressive Distributed Lag (ARDL) model to forecast the individual real estate price series. We compare the forecasted residuals to a conventional Autoregressive (AR) forecast model as a "benchmark" for two markets: Boston apartments and Dallas commercial. The results show that the ARDL model predicts the crisis and subsequent recovery really well, whereas the "benchmark" model typically follows the previous price trend. We find that the DFM forecasts are most precise with only one or two factors. The two prominent factors may reflect general economic conditions and the rental housing market, respectively.