Genetic Programming for Spatiotemporal Forecasting of Housing Prices

Abstract

This chapter compares forecasts of the median neighborhood prices of residential single-family homes in Cambridge, Massachusetts, using parametric and nonparametric techniques. Prices are measured over time (annually) and over space (by neighborhood). Modeling variables characterized by space and time dynamics is challenging. Multi-dimensional complexities—due to specification, aggregation, and measurement errors—thwart use of parametric modeling, and nonparametric computational techniques (specifically genetic programming and neural networks) may have the advantage. To demonstrate their efficacy, forecasts of the median prices are first obtained using a standard statistical method: weighted least squares. Genetic programming and neural networks are then used to produce two other forecasts. Variables used in modeling neighborhood median home prices include economic variables such as neighborhood median income and mortgage rate, as well as spatial variables that quantify location. Two years’ out-of-sample forecasts comparisons of median prices suggest that genetic programming may have the edge.