AI-Enhanced Loan Default Prediction Models

Abstract

Financial institutions face increasing challenges with respect to loan default prediction because their existing simple models have become outdated. Past due accounts have been left on bank balance sheets longer, and as a result, older default models are not accurately predicting charge-offs. Keenly interested in new and advanced techniques for assessing loan default risk, banks are turning to artificial intelligence, which enhances default prediction by taking into account many more factors than can be analyzed or weighted by a loan officer. Loans are the largest asset at most banks, and the ability to predict whether or not those loans will be paid is critical to sustaining institutional finances. Similarly, the prediction of which loans will default greatly affects the nation’s capital markets, with implications domestically as well as on the global economies. U.S. banks presently use logistic statistical models, typically either a stand-alone or blended credit score model, whose primary output is a score, i.e., probability, predicting the likelihood of loan default. While useful, it is not effective to approach loan default prediction strictly from a quantitative perspective. Qualitative factors, unique to individual borrower behavior, in conjunction with abundant large-scale data—such as established industry performance data—are important. The addition of borrower account activity to improve the discriminatory ability of blended credit score logistic models would be invaluable. In the domain of consumer lending, numerous models offer consumer-behavioral predictive factors with varying degrees of reliability. In small business lending, however, generating alternative data that a consumer or commercial bank might use to assess default risk is very difficult. Instead, the chosen methodology for this research focuses on unique account activities identified as important in previous unsuccessful small business financial investigations. In comparing recent studies exploring conventional statistical versus alternate AI approaches for loan risk prediction, no research is known to directly compare a blended credit score logistic model to its AI counterpart. Although data mining is often compared with conventional models in the literature, decision trees implemented by AI techniques are very rarely seen. Instead, AI-enhanced linear logistic models, neural networks, or genetic algorithms are conventionally studied. Moreover, easily interpreted decision tree methodology is rare in research, specifically for the domain of loan default prediction.