Predicting Credit Union Member Attrition using Machine Learning

This study aimed to help alleviate the problem of member attrition for a credit union that served as the study’s client. Like all other financial institutions, the client was looking to predict members with a high risk of leaving the credit union and the main reasons they do so. The study used all available potentially useful variables to predict a member leaving the credit union. These variables ranged from member descriptors such as tenure, age, and location to member behavior such as account balance counts, minimums, and maximums over time. The study built a decision tree and random forest to determine the predicted probability of a member leaving in the next six months. The decision tree performed well and helped to validate no underlying variable issues. The random forest performed very well with very high AUC values on two complete holdout test sets. The results could be used immediately for a marketing campaign and A/B/C testing. There is still much more that can be done to build off this study. Some ideas are creating models for prediction periods such as three and twelve months, testing models such as penalized regression and XGBoost, removing or adding features, and adding to a production environment for continual usage.