Advances in Forecasting Home Prices

Abstract

Numerous researchers have used various techniques to predict housing prices, but the results have been mixed. This article forecasts housing prices based on their stationary (level) and nonstationary (growth rate) presentations. Our study uses five classes of univariate time series techniques: autoregressive moving average (ARMA) modeling, generalized autoregression (GAR) modeling, generalized autoregressive conditional heteroskedasticity (GARCH) modeling, time-varying Kalman filtering with random autoregressive (KAR) presentation, and Markov chain Monte Carlo (MCMC) simulations. We assigned optimal weights to each technique to minimize the mean square error (MSE) of our forecasts. Our dynamic forecasting method shows superior out-of-sample performance based on the nonstationary presentation one to three quarters ahead, while reducing the average MSE by 37%. For four-quarter horizons, the average MSE of our dynamic forecasts decreased by 11% when we used a stationary presentation of housing prices and included lagged values for four economic leading indicators: the shadow federal funds rate, 1-year expected inflation, the 10-year Treasury Minus 3-Month Treasury Constant Maturity term spread (TERM), and the Brave-Butters-Kelley Leading Index.