Using Large Data Sets to Forecast House Prices:

Several Bayesian and classical models are used to forecast house prices in 20 states in the United States. There are two approaches: extracting common factors (principle components) in a factor-augmented vector autoregressive or factor-augmented Bayesian vector autoregressive models or Bayesian shrinkage in a large-scale Bayesian vector autoregressive models. The study compares the forecast performance of the 1976:Q1 to 1994:Q4 in-sample period to the out-of-sample horizon 1995:Q1 to 2009:Q1 period. The findings provide mixed evidence on the role of macroeconomic fundamentals in improving the forecasting performance of time-series models. For 13 states, models that include the information of macroeconomic fundamentals improve the forecasting performance, while for seven states they do not. This paper considers the dynamics of real house prices and the ability of different pure time-series models to forecast real house prices with two main foci. First, it considers how the researcher can incorporate large data sets into forecasting equations, using dynamic factor analysis or shrinking large-scale Bayesian vector autoregressive (BVAR) models. Second, it also considers the role of spatial priors in the BVAR models. The process is illustrated using house prices from the 20 most-populous states in the Unite