The historic models of credit risk modeling proven on the bank of structured financial parameters like credit ratings, salaries and repayment records have failed to reflect the dynamics of the new consumer behavior. As an economy grows more digital, such borrower insights are held in unstructured and behavioral data, such as transaction stories, talk to apps, customer support calls, which are routinely discarded by traditional models. The paper uses deep learning to look into the possibility of using such non-conventional data sources to extract credit risk information using deep learning architectures, specifically Long Short-Term Memory (LSTM) networks and Transformers. The paper suggests using client footprints left in various digital solutions as the basis of a revolutionary credit assessment method by converting them into one seamless, high-resolution borrower print. This paper also looks into why all these neural networks may not be a model but also question concerns of fairness, interpretability, and regulatory compliance. The end result is to come up with a data-based framework, which promotes greater access to credit, particularly among the underbanked and credit invisible client groups, without compromising risk integrity.
Godwin David Akhamere (2022) studied this question.