Hedonic house price models are increasingly applied in the process of mass appraisal, in which econometric specifications are used to obtain automated valuation of properties for taxation purposes. The predictive quality of such models is important, since it directly affects the revenue stream of local authorities. In this paper, we assess the relative predictive performance of different model specifications used in automated valuation. Specifically, we focus on the issue of spatial heterogeneity by comparing models that utilize different definitions of housing submarkets. In addition, we consider the inclusion of â€œspatialâ€ explanatory variables in the form of distance to various amenities as computed from a GIS. We apply this to data from the city of Bogot Ìa, Colombia, a pioneer in the application of mass appraisal techniques in a developing country context. We find that specifications that include the submarkets improve predictive performance and that the inclusion of the spatial variables is superior to the traditional models of homogenous zones. However, even the best models are still characterized by relatively poor performance in the form of a high degree of overprediction of the house value. In addition, the predictive performance of the models varied by socio-economic stratum in the city, which suggests that the dynamics of the housing markets in these strata would require closer and separate attention. These results may provide further guidance to enhance mass appraisal practice in the city of Bogot Ìa as well as potentially other Latin American cities.