Apartment rent prediction using spatial modeling

This paper provides a new model to explain local variation in apartment rents by introducing the notion of a spatial process. The background section defines the difference between global trend and local variation, and explains why proximity variables and sub-market variables should be modeled in the mean portion of the model, and that the inclusion of such variables does not constitute a spatial model. Additionally, the background section provides an example of a simple model that incorporates both proximity and zonal (sub-market) association variables. A plot of the residuals illustrates the significant level of spatial association that remains even after controlling for spatial effects in the traditional manner. The paper also provides a review of the use of both proximity and zonal association variables, and discusses early attempts at spatial modeling, including conditional spatial autoregressive, simultaneous spatial autoregressive models, and kriging. A review of the data for the eight markets included in this study is followed by a detailed description of the spatial process model. The final section presents the results, which indicate significant prediction improvement over traditional hedonic rent models that only include indicator variables to capture spatial effects. A number of maps illustrate both continuous residual and rent surfaces, clearly illustrating the spatial patterns that exist in apartment rents.