Prediction of consumer credit card risk from an analysis of spending categories: a hidden Markov model

Purpose The purpose of this paper is to investigate whether the dynamic evolution of a credit card customer’s latent risk state helps in predicting delinquency. Further, we assess whether an analysis of customers’ spending categories is useful to understand delinquency. Design/methodology/approach The study develops a Hidden Markov model (HMM) to predict whether a customer will be delinquent or not. Delinquency is defined as two or more months of non-payment of the minimum amount due. The explanatory variables in the model include customer demographics, past delinquent behavior and dollar amounts spent in different categories. Data were obtained from a large US bank for a large sample of customers. The data had customers' spending and payment behaviors in each month for a period of 25 months. Some demographic variables were also available. Findings Empirical findings show that the dynamic evolution of customers’ latent risk state predicts delinquency quite well. Customers who move from a high-risk state to a low-risk state are the most profitable with a moderate probability of delinquency and constitute over 53% of the customers. On the other hand, customers in high-risk states who stay in high-risk states have a high probability of delinquency even though they are highly profitable. This segment consists of 27% of the customers. Further, we find that spending categories predict the latent risk state. High-risk customers tend to overspend on entertainment, while low-risk customers use cards to make payments and spend on travel and discount shopping. We show that spending in discretionary categories is associated with higher delinquency, while spending in non-discretionary categories is not. Research limitations/implications The results are useful to banks in identifying and managing risky credit card customers. Practical implications Bank managers would benefit from understanding the latent risk state of customers that is unobserved. The HMM provides a method for calibrating the evolution of the risk state over time. Some patterns in the evolution of the hidden state predict delinquency quite well. Based on the model, bank managers could take remedial actions on customers that stay in a high-risk state for a long time. They could monitor risky accounts much more carefully, increase the annual percentage rate or lower the credit limit for such customers. Social implications By managing risky customers, banks and financial institutions can lower their costs, improve profitability and serve customers more effectively. Originality/value Past research has used demographic and transaction variables to predict delinquency. However, these variables have limited explanatory power. While HMM has been applied to detect credit card fraud, it has not been used to predict credit card delinquency. Further, the use of spending categories to predict delinquency is a new idea.