The problem

A leading US bank is responsible for evaluating and assigning a risk score for every customer applying for a loan with the bank. Recently partnered with a hospitality merchant and has procured customer data relating to raw transactions, product affinity, purchases, and demographics from the partnering merchant. Using the new rich data, the bank wants to improve on the existing credit risk model that will eventually improve the performance and reduce customer defaults in the long run.

The cognitive solution

With the procurement of new data, the data had to be ingested into the data mart for consistent use and develop a lineage for frequent updates.

1. Merchant data ingestion

• Data touchpoints across affluence and spends, product affinity, recency-frequency metrics, customer demographics, in total, over 1800 data fields were ingested
• 200+ new features were deemed significant and derived from the merchant data to be used in the risk model

2. Benchmarking against current risk model

• The new features were benchmarked with the two champion -challenger framework models – prophet model and a neural network model
• New segments were defined based on the customer preference in premium segments from merchant data enabling a cleaner separation of low-risk and high-risk customers

Key insights

The new credit risk model showed an improvement of 5% with an expected benefit of over $35M annually. The customer distribution across risk score ratings are also more balanced after the incorporated enhancement. Some interesting customer insights allowed the bank to understand the customer in a different lens:

• Premium package spending customers have a lower chance of defaults as opposed to the other segments
• Customers who avail self-drive car rentals surprisingly have a lower chance of defaults
• Non-credit card recorded transactions (cash/cheque) show a higher risk for defaults