A bank wanted to improve its capability for early identification of customers who may get into financial difficulty at a later date. This question was of interest as regulators in the UK and elsewhere are increasingly concerned with questions of identifying and supporting vulnerable customers. There is also commercial value in early flagging of at-risk customers to help them avoid running up debt. Finally, the bank was interested in exploring the value of customer data more widely, using this as a case study of modern data science capabilities.
Seeing financial difficulty through the customers’ eyes
First, we worked with the bank to define financial difficulty through a customer lens. In this case, we ended up with a definition that incorporates both the length of use and the amount borrowed via unarranged overdrafts. We then set out to answer three key research questions:
how predictable is financial difficulty and to what extent is it due to unexpected shocks?;
what are the key early indicators of financial difficulty?; and
can we improve on using credit scores to predict financial difficulty?
Our research revealed:
Depending on the relevant actions the bank wanted to take (and on the cost of false positives), we were able to correctly flag up to 90% of customers who would go into financial difficulty.
We identified the most promising additional data sources (e.g. card transactions) which could increase the accuracy and predictive power of the model.
Machine learning for future success
The project demonstrated the potential value of integrating machine learning techniques into the bank’s core capabilities (e.g. credit risk assessments).
A new team recently established by the bank to focus on machine learning continued to progress the work, looking to apply the insights with real time data and specific interventions for customers identified at risk of financial difficulty.