A branch of a popular national bank in China had a retention problem with their private banking customers. Each account had a minimum daily deposit requirement of CNY 10 million or USD 1.5 million, so the bank understood that acquiring new customers would be several times more expensive than retaining existing ones. The bad news was that the churn rates, or the percentage of account holders who discontinued their services, were higher than expected over recent quarters.
The bank detected a churn rate of between 5% and 13% from their private banking customers who had been moving their savings accounts to other products or other banks over a period of a few months, with the highest number of churned customers coming from the Futures and Deposit department. The bank’s Marketing Department was tasked with running a promotional campaign targeting those customers that were most likely to withdraw their savings. The challenge was finding out which high-valued clients were most likely to churn and then use targeted marketing promotions to prevent these customers from moving their accounts elsewhere.
How SPM Helped
The project team used Salford Predictive Modeler’s (SPM) machine learning engine
TreeNet® to predict the potential customer churn. QY Datatech Inc, Minitab’s Authorized Partner in China, introduced the team to Salford Predictive Modeler (SPM). They chose TreeNet because of its flexibility and accuracy as well as its ability to deal with a data structure that has many observations.
TreeNet was able to help the project team quickly narrow down the variables that make the most impact, and visually illustrate the relationship between variables and probability of churn with these identified variables. The TreeNet model can help visualize the different layers of interaction between complex variables. In this case, these variables included the amount of savings, the purchase of gold and other investments, transaction time, transaction amount, company, job position, social insurance, debt, credit card limit, real estate and car ownership.
Through eliminating low-response customers from the list and then validating the models, a targeted group of VIP customers was identified. The models were able to predict the responsiveness of the customers to marketing campaigns based on previous marketing campaigns.
The team used the TreeNet analysis because within a few clicks and the help of automation, the initial model was ready with reliable results with more complex analysis powered within a few days.
The double variable dependency plot of the TreeNet model shows how two variables, ‘marriage’ and ‘household by region’ interact with the target variable or the probability of customer response to the market campaigns at a glance. The project team can then pinpoint a higher or lower response range with ease.
The variable importance summary in TreeNet shows which variables are of interest to the project team in terms of target interaction. The higher-rated variables are more likely to contribute to the target.
The TreeNet summary report shows model performance in one window. The quantified statistics are listed, available for comparison with other models, even the ones generated by other tools such as ROC, Lift, K-S Stat and Misclassification Rate.
As a result, the bank was able to predict with two months’ notice that 1,700 private banking customers out of 12.9 million could be lost to another bank or financial service provider with an accuracy of between 80% to 90%.
The bank then conducted a marketing campaign for these customers and achieved the goal of reducing the loss of these VIP customers and increased the funds retained by the bank. In addition, the customer asset under management (AUM) at the bank, which includes deposits, futures, stocks, and gold was increased by 16%.