What we did We established a definition of customer churn that defines the moment of outflow of capital, based on a combination of savings depletion, transactional inactivity and account closure. We extracted features to create a complete picture of each customer based on personal characteristics (e.g. age, location), activity in the online banking environment, transactions (type, volume, frequency) and market information (e.g. competitiveness of interest rates). Before starting the analyses, we sanitized the data using robust statistical outlier detection methods.
While machine learning-based churn models, such as random forests and support vector machines, can capture nonlinear data, they typically yield poorly interpretable results. Classic methods, such as logistic regression or linear support vector machines, are easier to interpret, but cannot capture nonlinearity. However, we found a way to tweak linear models using I-splines so that nonlinear events are still captured and can be properly interpreted. I-splines fit polynomial functions to several subintervals for each variable. Customers that have features in multiple subintervals with an increased risk, adhere to the risk profile and can be targeted with marketing actions.
What we achieved We provided our client with essential insights into their customer churn. Combining risk-zones for savings volumes, transactional (in)-activity and app usage we were able to identify an at-risk subpopulation of customers. Using the results of our analysis, our client has started to increase customer retention in a focused manner.
What they said
“We have enjoyed working with IG&H. We were really on the same wavelength, without any need to explain where our challenges lie. The results of this investigation are interesting, especially because the results are practically applicable, adding concrete business value.”