Enhanced Churn Prediction Model with Boosted Trees Algorithms in The Banking Sector

Abstract

Data science and machine-learning approaches assist banks in identifying the factors that encourage client churn and taking the necessary steps to reduce it. There has been a significant amount of study in the banking sector on customer turnover, but to the best of our knowledge, none of them employ the same data source, methodology, and machine learning algorithms to estimate the chance of attrition for individual customers. The purpose of this study is to research, improve, and evaluate several boosted-based techniques to develop the best possible customer churn model in a bank. In this study, we evaluate three boosting-based models to categorize customer turnover data on two scenarios: Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Category Boosting (CatBoost). These methods were applied on default and hyperparameters. The hyperparameters derived from Grid Search 10-Fold Cross-Validation work. The six models developed are assessed using four distinct machine learning criteria, and we offer a model with a comprehensive methodology based on customer turnover data from a bank. Our results demonstrate that utilizing LightGBM to perform on produces the best possible model with 91.4%, 94.8%, and 87.7% for both accuracy and AUC, precision, and recall scores respectively.