Risk Factors for Consumer Loan Default: A Censored Quantile Regression Analysis

The most widely-used econometric technique for analyzing default behavior in consumer credit markets is the proportional hazard model, which assumes that borrower characteristics increase or decrease default probability in a similar way over the life of a loan. In this paper, I employ an alternative method using censored quantile regression (Portnoy (2003)) on a data set of over 17,000 loans. This approach evaluates the effect of a borrower’s financial characteristics on the entire distribution of default times, allowing co-variates to influence early and late defaulters in different ways. I find that several typical predictors of credit-worthiness influence default probability differently depending on the amount of time that a loan has been active. I illustrate the importance of this heterogeneity by comparing predicted default probability and expected profit across the two models. Because the quantile regression model takes into account the effect of characteristics on the timing of default, rather than the overall probability of default, it finds a significantly higher expected profit for lowand medium-risk ∗University of Michigan Stephen M. Ross School of Business. Email address: [email protected]. I would like to thank Roger Koenker for his generous advice and guidance. I am grateful to an anonymous referee whose comments substantially improved this paper. This paper also benefited from helpful discussions with Dan Bernhardt and Darren Lubotsky and comments from seminar participants at the University of Illinois, Urbana-Champaign.