Credit Risk Estimation Using Various Graphs

Abstract

For banks and financial companies lending to individuals, understanding the creditworthiness of new customers is essential. It is challenging for people with limited/no past credit data. Most credit risk models extract features about existing users and aggregate them to infer for new ones. There are not enough known features about such customers to be extracted satisfactorily. Axio Digital Pvt Ltd is an NBFC lending to millions of Indians. This project uses anonymised datasets to develop various customer networks. Each node and edge has attributes reflecting various aspects of the customers and the various shared pieces of information among them, like mobile numbers, email IDs, address components, etc. These provide valuable insights into their credit behaviour. Specifically, networks based on customer email IDs and mobile numbers were developed in the first part of the project. Then, addresses from each pin code region were clustered to derive address cluster-based features for an individual. Additionally, address unigram co-occurrence graphs were created. Using these, several features were calculated for each customer to predict their default/non-default status using binary classifiers. The models achieved appreciable predictive power on both in-time and out-of-time test datasets. They perform comparably to standalone inference models built on the same dataset. They can provide an alternative way to determine creditworthiness, especially for customers with limited/no past credit history. Overall, this study enhances Axio’s capability to make sound credit decisions, thereby fostering customer trust.