Value Estimation of Properties Administered by the Brazilian Army Using Machine Learning and Spatial Components

Abstract

The valuation of an institution’s patrimony represents a necessary condition for an efficient management of its assets. The execution and analysis of real estate appraisal reports are essential to the achievement of some strategic objectives of the Brazilian Army, but they are also quite costly in terms of time, labor and financial resources. Sometimes, great effort is required for the aforementioned steps to take place and the market value finally obtained is inconsistent with what was initially imagined by the authorities, causing the technical study carried out to not be effectively used in negotiations by the organization. This work proposes the development of predictive models capable of building estimates of real estate values, so that the formal requests of the managers that imply the stages of execution and analysis of appraisal reports can occur with this information as an initial input. Counting on linear and nonlinear approaches and on machine learning techniques, the models have a reasonable level of assertiveness and national geographic coverage when generate estimated market values of Union real estate assets. Intrinsic and extrinsic variables to the properties were considered, including tests of aggregation of spatial components on some of them. As the interpretability of the proposed solution is an important requirement in both linear and nonlinear approaches, the Shapley value was adopted as a tool to support the guarantee of explainability and a PLS-SEM conceptual model was built to select attributes in a reasoned manner. These two considerations associated with modeling of real estate prices at a national level represent an innovation of this work in relation to the scientific literature analyzed.