Default Prediction and Generalized Covariate Effects Based on Ann Methods

This paper suggests a multilayer artificial neural network (ANN) method to predict the probability of default (PD) within the survival analysis framework. ANN structures consider hidden interconnections among covariates determining the PD which can lead to prediction improvements, compared to logit models. To measure covariate effects on the PD, we propose a method allowing for composition effects among covariates and use stochastic dominance analysis to rank their importance across the ANN and logit models. The application of the method to a large data set of small business loans demonstrates prediction gains, relative to the logit and skewed logit models. These gains are more apparent for short term prediction horizons and for the type I misclassification error of loan default events, which has important implications for credit risk management. Concerning the covariate effects, our results indicate that the most important influence on the PD comes from the covariates capturing behavioral attitudes. Furthermore, we show that the ANN method stochastically dominates the logit models for most of the covariates considered