Bank Churn Model
For a domestic bank, we created an early warning system using artificial intelligence that enables targeted intervention by identifying the customers most prone to dropout.
Industry
Function
ProductContext
The expectation was to reduce the attrition rate among the bank's customers. This was obviously of paramount importance because if the bank could retain a fraction of its churn rate, that would generate significant revenue on an annual basis. The expectation was that the system should identify customers who were likely to drop out or terminate their contracts, since the marketing effort required to retain them would be significantly lower than the cost of acquiring a new customer.
Approach
In order to apply modern machine learning algorithms, we had to start from the ground up: from the design of the data relationship, through the analytical sorting and cleaning of the database, to the definition of the software environment and infrastructure requirements of the final program. The solution we implemented had to comply with the data protection requirements of the domestic banking sector and the European Union.
During the project, we assessed and categorized the customer's data assets by examining the contents of more than 1500 data tables. The random forest algorithm we selected is at the forefront of international research - by applying it, we significantly reduced the time spent on data cleaning, recoding and transformation. The final model also uses customer transaction and card usage patterns, account metainformation, and anonymized customer demographics to make predictions. During the implementation, we made sure that neither the data nor the model leaves the bank's on-premise infrastructure, thus guaranteeing our client's business integrity. The project was built in a modular way, so that new data sources can be integrated in the future to produce the prediction.
Results
On average, our model successfully predicted 40% of clients who drop out each month and was wrong for less than half a percent of clients. In contrast, the benchmark model used to predict customer attrition using logical conditions and rules only found 25% of customers who attritted and misclassified 6% of the total customer base (false positives). Such accurate identification of attrition-prone customers represents a business opportunity and thus a potential competitive advantage for the bank using the model.