A statistical learning approach to land valuation: Optimizing the use of external information

Abstract

We develop a statistical learning model to estimate the value of vacant land for any parcel, regardless of improvements. Rooted in economic theory, the model optimizes how to combine common improved property sales with rare, but more informative, vacant land sales. It estimates how land values change with geography and other features, and determines how much information either vacant or improved sales provide to nearby areas through two levels of spatial correlation. For most neighborhoods, incorporating improved sales often doubles the certainty of land value estimates. Relative to conventional estimators, our method mitigates problems from excess variance and sample selection.