Predicting Loan Defaults Using Logistic Regression

Abstract

We used anonymized data from a loan company to analyze correlations between loan defaults and other characteristics of loans or borrowers of loans. We performed an exploratory data analysis of the different factors and how they correlated with loan defaults. Using observations made in the EDA, we proceeded to use logistic regression to predict the odds of loan defaults with several loan characteristics as predictor variables. Different models were evaluated and cross-validated using AIC, AUC, and predicted accuracy. Weighted accuracy was also measured because the loan dataset was a stratified sample. We concluded that the interest rate most accurately predicted the odds of a loan default and that the most useful model was both simplistic and accurate. Research was limited by the variables that were not analyzed during EDA, the limited variables the loan dataset contained, and the modeling technique used.