Date of this Version
Financial lending institutions continuously look at improving their credit risk models. This study examines the performance of three estimation methods: logistic regression, decision tree, and neural networks, in terms of their misclassification rates of credit default. The study uses 17,328 loans of grain producers for the period of 2006 - 2010. Those loans belong to the category of “diversified loans / core standard” originating from a large financial lending institution. The data has been split into nine different sets to acknowledge three factors: the shift in price of grains to a higher plateau after 2006, the contamination effect on defaulting on more than one loan, and the lack of information provided by the borrower at the time the loan is initiated. Findings show that credit default predictions vary slightly depending on the model used. In addition, when excluding the data for the loans that were refinanced and matured in 2006 there are a different set of significant variables that affect the prediction of default. The results also show the importance of having separate models for borrowers with one loan versus those borrowers with more than one loan.