Commercial Mortgage Default: A Comparison of Logit with Radial Basis Function Networks

Abstract

This manuscript explores the use of artificial neural networks in the modeling of foreclosure of commercial mortgages. The study employs a large set of individual loan histories previously used in the literature of proportional hazard models on loan default. Radial basis function networks are trained on the same inputs as those used in the logistic, and performance is assessed in terms of prediction accuracy. Neural networks are shown to be superior to the logistic benchmark in terms of discriminating between "good" and "bad" loans. Sensitivity analysis performed on the average loan demonstrates the use of neural networks as an analytical tool. Finally, the study offers suggestions on further improving prediction of defaulting loans.

JEL Classification: G21