Evaluating House Price Forecasts

Abstract

This study uses an autoregressive process to model a city-wide house price index. The model is used to produce one-quarter ahead forecasts for individual properties. We propose that managers use a battery of tests to compare prediction errors (PEs); in particular, their empirical distribution reveals important information.Transaction data from Dade County, Florida is used. PEs from two forecasting models, hedonic and repeat sales, show some departure from desirable properties of forecasts. Also, both show some informational inefficiency, but the hedonic is more efficient than the repeat. Nonparametric smoothing shows that the hedonic method dominates the repeat over an important range of PEs; thus, many risk-averse managers might prefer a forecast based on the hedonic method.