Modeling Spatial and Temporal House Price Patterns: A Comparison of Four Models

Abstract

This research reports results from a competition on modeling spatial and temporal components of house prices. A large, well-documented database was prepared and made available to anyone wishing to join the competition. To prevent data snooping, out-of-sample observations were withheld; they were deposited with one individual who did not enter the competition, but had the responsibility of calculating out-of-sample statistics for results submitted by the others. The competition turned into a cooperative effort, resulting in enhancements to previous methods including: a localized version of Dubin's kriging model, a kriging version of Clapp's local regression model, and a local application of Case's earlier work on dividing a geographic housing market into districts. The results indicate the importance of nearest neighbor transactions for out-of-sample predictions: spatial trend analysis and census tract variables do not perform nearly as well as neighboring residuals.

Keywords: kriging, out of sample prediction, data snooping, local polynomial regression, smoothing regressions, semiparametric models, cluster analysis, nearest neighbors, hedonic models