Modeling of apartment prices in a Colombian context from a machine learning approach with stable-important attributes

Abstract

The objective of this work is to develop a machine learning model for online pricing of apartments in a Colombian context. This article addresses three aspects: i) it compares the predictive capacity of linear regression, regression trees, random forest and bagging; ii) it studies the effect of a group of text attributes on the predictive capability of the models; and iii) it identifies the more stable-important attributes and interprets them from an inferential perspective to better understand the object of study. The sample consists of 15,177 observations of real estate. The methods of assembly (random forest and bagging) show predictive superiority with respect to others. The attributes derived from the text had a significant relationship with the property price (on a log scale). However, their contribution to the predictive capacity was almost nil, since four different attributes achieved highly accurate predictions and remained stable when the sample change.