Predicting take-up of home loan offers using tree-based ensemble models: A South African case study

Abstract

We investigated different take-up rates of home loans in cases in which banks offered different interest rates. If a bank can increase its take-up rates, it could possibly improve its market share. In this article, we explore empirical home loan price elasticity, the effect of loan-to-value on the responsiveness of home loan customers and whether it is possible to predict home loan take-up rates. We employed different regression models to predict take-up rates, and tree-based ensemble models (bagging and boosting) were found to outperform logistic regression models on a South African home loan data set. The outcome of the study is that the higher the interest rate offered, the lower the take-up rate (as was expected). In addition, the higher the loan-to-value offered, the higher the take-up rate (but to a much lesser extent than the interest rate). Models were constructed to estimate take-up rates, with various modelling techniques achieving validation Gini values of up to 46.7%. Banks could use these models to positively influence their market share and profitability.

The Changing Landscape of Financial Credit Risk Models

“…The results showed that several of the machine-learning techniques predicted credit risk significantly more accurately than the industry standard of logistic regression. One specific example is a publication by Verster et al (2021) which showed that bagging and boosting techniques outperformed logistic regression when predicting take-up rates of home loans in cases where banks offered different interest rates. Meaning in the future, more machine learning techniques will be used in financial credit risk predictive modelling.…”

“…One example is ensemble-based models, such as bagging (Maldonado et al 2014), random forests (Yiu 2019), and boosting (Verster et al 2021). For all three of these ensemble machine-learning techniques (random forest, bagging, and boosting), the underlying technique is a statistical technique called CART (classification and regression tree) developed by Breiman et al (1984).…”

“…For all three of these ensemble machine-learning techniques (random forest, bagging, and boosting), the underlying technique is a statistical technique called CART (classification and regression tree) developed by Breiman et al (1984). One example is ensemble-based models, such as bagging (Maldonado et al 2014), random forests (Yiu 2019), and boosting (Verster et al 2021). For all three of these ensemble machine-learning techniques (random forest, bagging, and boosting), the underlying technique is a statistical technique called CART (classification and regression tree) developed by Breiman et al (1984).…”