Large-scale simultaneous market segment definition and mass appraisal using decision tree learning for fiscal purposes

Cadastral assessment aims at guarantying equity in the allocation of property taxes. Therefore, we must be able to massively determine property values through models that reflect, with the minimum error, the behaviour of land market in each region. Despite this imperative need, currently land valuation for cadastral purposes is plagued with subjectivity. A very extended bad practice for instance is to assume that variables of productive performance i.e. land use capacity, are the ones with the highest influence on land value formation in the rural sector. The former assumption largely ignores the plethora of rural land uses that exist nowadays. To open the door to less subjective methodologies of land mass appraisal we borrowed statistical methodologies from the field of data-mining and applied them to a dataset of 410 purchase-sale transactions (2003–2009) of land plots located in the rural sector of the Vilcabamba parish (southern Ecuador). Land market behaviour in Vilcabamba responds to a transition from a pure agricultural territory to a touristic one at which many second-homes are being built for leisure. Our results demonstrate the applicability of methodologies such as model-tress (M5P) and multivariate adaptive regression splines (MARS) to rural land mass appraisal. Both M5P and MARS allow defining market segments while simultaneously establishing the weights of predictor variables for land value formation. We also collected evidence supporting that removing variables of productive performance from land value prediction models do not hamper models predictive power at least in rural areas where gentrification is taking place.