Analysis of ensemble classifiers for bank churn prediction

Today, with rising competition and customer migration, customer churn is an issue of great importance and relevance. Banks are facing the problem of retaining customers because clients tend to look for other options that may provide them with better financial services, decreased costs, better interest rates, etc. To solve this problem, we use several prediction models to predict which customers may choose to leave the service. In this study, we explore the use of 10 ensemble learners broadly categorized as bagging, boosting, voting, and stacking. To attain higher performance, ensembles generally use algorithmic diversity with a predetermined set of training instances. We conduct ETL and exploratory data analysis on five open-source datasets. Then we balance the datasets by resampling before we go on to conduct empirical validation on them. We observe that random forest and histogram based gradient boosting give better results when compared to other ensemble methods based on five different performance metrics, especially in specificity and geometric mean.