Avaliação em massa com modelos de aprendizado de máquina aplicados aos terrenos urbanos do Município de Fortaleza

Abstract

This study concerns the mass appraisal of the market value of land in the city of Fortaleza. It is made by using machine learning models, from a sample with more than 8 thousand observations collected through an urban observatory of market values in the period from 2015 to 2019. An extensive exploratory analysis is carried out for the definition and choice of the explanatory variables of this evaluation, having as response variable the unit price of land. Subsequently, ordinary least squares regression is studied as a preliminary model to be outperformed by machine learning models, random forests and XGBoost. For each of these, the assumptions are assessed, mainly for the ordinary least squares model, due to its difficulty in meeting all its premises in real estate mass appraisal. The estimates of this model, with the unit price on a natural-log scale, are consistent with what is expected in practice and observed in the previous exploratory analysis. For machine learning models, random forests and XGBoost, the relationships among bias-variance trade-off, power of predictive generalization and overfitting are verified. The most important features to explain the unit prize of land are remarkably similar in both models. The XGBoost model outperforms the others in all the performance measures evaluated. At the end, a market value map is proposed for all georeferenced land parcels in Fortaleza.