Context:

The Multipurpose Cadastre public policy needs to consolidate real estate information from different sources for analysis, such as offers, transactions, and construction costs, among others. Real estate websites are part of these sources of information, although they have not yet been included in commercial analysis. In light of the above, it is necessary to review a methodology that allows optimal access to these web platforms and facilitates the analysis of the variables provided therein, which are crucial to a property’s commercial value. A study case was carried out in three Colombian cities: Fusagasugá, Manizales, and Villavicencio.

Method:

The method is implemented in two stages: (i) web scraping, which allows obtaining the information links from real estate web pages and downloading their data, and (ii) analyzing real estate data by developing a workflow that starts with data exploration and cleaning, continues with pre-modeling, and ends by modeling the crucial variables in the determination of real estate value using machine learning techniques.

Results:

By applying machine learning techniques, it was possible to automate the collection, cleaning, storage, and analysis of real estate data from web platforms, as well as to outline two models (Ridge Regression and Random Forest), which, according to their mean absolute percentage error (0,34 and 0,35, respectively), allow predicting the commercial value of a property while considering internal and external explanatory variables.

Conclusions:

Obtaining and analyzing real estate data from alternative sources such as web platforms through machine learning techniques contributes significantly to addressing the high information de-mand of the country’s cadastre. However, it is necessary to expand the supply of this information to rural areas, which have less access and availability to it.

        · resumen en Español     · texto en Español     · Español ( pdf )