Machine Learning for Customer Churn Prediction in Retail Banking

Authors:
Publisher website
Google Scholar
Abstract
Predicting in advance whether a given customer will end his relationship with a company has an undeniable added value for all organizations, since targeted campaigns can be prepared to promote customer retention. In this work, six different methods using machine learning have been investigated on the retail banking customer churn prediction problem, considering predictions up to 6 months in advance. Different approaches are tested and compared using real data. Out of sample results are very good, even with very challenging out-of-sample sets composed only of churners, that truly test the ability to predict when a customer will churn. The best results are obtained by stochastic boosting, and the most important variables for predicting churn in a 1–2 months horizon are the total value of bank products held in recent months and the existence of debit or credit cards in another bank. For a 3–4 months horizon, the number of transactions in recent months and the existence of a mortgage loan outside the bank are the most important variables.

This publication has 36 references indexed in Scilit: