AI-Enhanced Credit Scoring Models

Abstract

Credit scoring refers to predictive analytics using consumers' credit reports and other application information for assessing creditworthiness. Under pre-algorithm and scorecard models, expert rules and a small number of explanatory variables or "features" were used. Automated underwriting models, such as those combining decision trees with logistic regression, as well as newer approaches, are models that include many features. But the bulk of the literature on credit scoring has been focused on the pre-algorithm models using expert rules.

The field of expert-rule-based credit scoring has been transformed in the last 20 years by new technology, such as speedier software and the sale and analysis of "big data" from credit bureaus and credit card and other companies. As markets and technology have changed for credit scoring, academics, data scientists, and others have sought to improve predictive accuracy. They have been particularly interested in at least two offshoots of machine learning. One is methods of "semi-supervised learning" to develop predictive machine learning models for the "thin file" or "no file" populations—that is, those customers tracked with few or no previous characteristics, behavior, or outcomes. It is true that many people who have "no file" can be scored with traditional models using other evaluative information such as application data. However, studies have found that "branches with a big share of thin file borrowers can run into serious credit scoring problems." Another development in machine learning about which many are excited is boosting techniques where a combination of many weaker models together forms a strong model.