A Machine Learning Web Application to Estimate Listing Prices of South African Homes
Due to the heterogeneous nature of residential properties, determining selling prices which will reconcile supply and demand is difficult. Establishing realistic listing prices is vitally important for sellers to prevent prolonged time on market. Sellers have several resources available to assist in this endeavour, all of which involve understanding current market dynamics through analysing recent sales and listing data. Property portals which aggregate real estate agencies’ data, hosting it on online platforms, are one such resource, along with individual real estate agencies. Leveraging this data to develop solutions that could aid sellers in listing price decision making is a potential business objective that could not only add value to sellers but create a competitive advantage by increasing traffic to an online real estate platform. Using data provided by a South African online property portal, this paper creates a web application using machine learning to estimate listing prices for different types of homes throughout South Africa. This study compared log linear and gradient boosted models, estimating residential listing prices over a four-year period. The results indicate that although log linear models are suitable to account for spatial dependency in the data through the inclusion of a fixed location effect, the assumption of linear functional form was not satisfied. The gradient boosted models do not impose explicit functional form requirements, making them flexible candidates. Similarly, these models were able to handle the spatial dependency adequately. The gradient boosted models also achieved a lower out of sample error compared to the log linear models. The findings show that over observation periodperiod, larger properties consistently experience a diminishing return at some point over the marginal distribution of physical characteristics. The web application details how sellers are easily able to obtain mean listing price estimates and gauge the growth thereof, by simply inputting their property interest criteria.