Econometric and machine learning spatially varying models for the mass appraisal of residential properties

Abstract

The differences in predictive performance between econometric and machine learning (ML) spatially varying models for the mass appraisal of residential properties are largely unknown. Therefore, this article aims to identify the most accurate and precise model to serve as an automated valuation model for the housing market. The study examines global approaches such as multiple regression (LR) and eXtreme gradient boosting (XGBoost), but places greater emphasis on local methods. Among the latter, econometric linear models were analysed, particularly geographically weighted regression (GWR) and multiscale GWR (MGWR), as well as ML methods such as geographically weighted random forest (GWRF) and geographical XGBoost (G-XGBoost). The set of models was supplemented with a local hybrid model – geographically gradient boosting weighted regression (GWRBoost). The housing markets of Warsaw and Amsterdam were selected as case studies, with both rental and sales prices examined. The results revealed that GWRF and XGBoost performed best in terms of out-of-sample prediction, while offering reasonable computational time. The performance of GWRF and XGBoost was statistically equivalent for rental prices, whereas GWRF achieved significantly better results for sales prices. Analysis of partial dependence plots and local variable importance values revealed both nonlinear and spatially varying relationships between price and the predictors. In the rental market, modelling nonlinearity is largely sufficient to achieve high predictive performance. Spatial heterogeneity, on the other hand, is an additional advantage when predicting sales prices.