Examining the Information Content of Residuals from Hedonic and Spatial Models Using Trees and Forests

Machine learning algorithms such as neural nets, support vector machines, and tree-based techniques (classification and regression trees) have shown great success in dealing with a number of complex problems (Hastie et al. 2009). However, real estate data exhibit both temporal dependence and high levels of spatial dependence (Pace et al., International Journal of Forecasting16(2), 229–246, 2000; LeSage and Pace 2009) that may make it harder to use with off-the-shelf machine learning procedures. We examine tree-based techniques (CART, boosting, and bagging) and compare these to spatiotemporal methods. We find that bagging works well and can give lower ex-sample residuals than global spatiotemporal methods, but do not perform better than local spatiotemporal methods.