Machine Learning-Based Profit Modeling for Credit Card Underwriting - Implications for Credit Risk

Retail credit issuers traditionally rely on cutoffs derived from risk-based scorecards for acquisition and account management purposes. However, in recent years, advances in data and increased comfort with advanced modeling practices such as machine learning have given rise to more sophisticated behavioral and profit-based modeling approaches, which attempt to estimate the projected NPV of a customer instead of their likelihood of going delinquent or charging-off. Financial institutions can potentially adopt these profit-based models for underwriting and loan management purposes to more precisely target profitable but potentially risky consumers. Using a unique proprietary panel dataset of credit cards combining data from many major banks, I construct both traditional risk and ML-based profit models and find that, using only information known at customer acquisition, these models can rank-order customers according to their risk and profit reasonably well. I find that profit score cutoffs generally target wealthy, high spending, "revolving" customers, while risk score cutoffs target low-activity "transacting" customers. Absent risk-based guardrails, profit-based underwriting could potentially cause an increase in riskiness in a card portfolio. However, this is highly dependent on portfolio composition and increased risk would mostly be for portfolios concentrating on "revolvers" in the lower end of the credit spectrum, reinforcing the importance of risk guardrails in these portfolios for institutions utilizing such new modeling techniques