Objective

Use a retail grocer’s loyalty shopper data as an alternative data source for predicting credit risk among credit invisible applicants, aiming to grow financial inclusion, especially within the lower-income demographic.

Overview

Recognizing the potential of loyalty shopper data, especially from a retail grocer catering to a lower-income demographic, we set out to overlay this data with banking datasets. The goal was to identify mutual customers and understand the shopping behavior of “good borrowers.” This would allow banks to establish a risk profile based on this behavior and offer credit to credit invisible applicants, who typically would be declined.

The grocer and the banks used Omnisient’s desktop application to anonymize and tokenize their customer datasets with Omnisient’s US patented Crypto-IDs before uploading to Omnisient’s secure and neutral “escrow-like” environment in the Cloud for overlapping and analysis:


• Loyalty Shopper Data: This data provided insights into the shopping behaviour of millions of anonymized individuals. With the initiative targeting individuals without credit, this data set was particularly relevant, because it tracked cash-based purchases mostly made by shoppers without credit.
• Bank Loan Repayment Data: This data provided an overview of bank clients who had taken out a loan. This dataset was overlapped with the retailer’s shopper data to reveal the shopping patterns of the bank clients considered “good borrowers”. This enabled the bank to build and test predictive models for loan repayment based on the shopping behavior of their “good borrowers”.

Omnisient’s data scientists played a critical role in this initiative. Recognizing the limitations of existing tools in identifying relevant retail-specific data and patterns for credit risk assessment, they leveraged their retail experience to develop custom code. This code allowed them to extract and summarize relevant information from raw data, such as identifying patterns based on the retail calendar, weekend dynamics, month-end spending trends, and correlating transaction types (cash vs. card) with credit default likelihood. This human-driven data preparation enhanced the accuracy and quality of information used by our embedded machine-learning tools for predicting customer behavior and credit risk.