Learning from man or machine: Spatial aggregation and house price prediction

Abstract

House prices vary with location. At the same time the border between two neighboring housing markets tends to be fuzzy. When we seek to explain or predict house prices we need to correct for spatial price variation. A much used way is to include neighborhood dummy variables. In general, it is not clear how to choose a spatial subdivision in the vast space of all possible spatial aggregations. We take a biologically inspired approach, where different spatial aggregations mutate and recombine according to their explanatory power in a standard hedonic housing market model. We find that the genetic algorithm consistently finds aggregations that outperform conventional aggregation both in and out of sample. A comparison of best aggregations of different runs of the genetic algorithm shows that even though they converge to a similar high explanatory power, they tend to be genetically and economically different. Differences tend to be largely confined to areas with few housing market transactions.