Interpretable credit model development via artificial neural networks

This research examines techniques for interpreting the connection weights in an artificial neural network (ANN) in order to develop an initial credit risk model that can be used to explain an adverse credit decision and that is more accurate than credit risk models developed from traditional statistical techniques. Traditionally, consumer credit risk models have been developed using linear statistical techniques. ANNs are non-linear statistical techniques that lack the ability to provide an explanation for the output of the model. ANNs are being used with much success in statistical modeling situations that do not require interpretation of the decisions of the models. The lack of explanatory power of ANNs has prevented their use in consumer credit risk modeling.; This research shows that ANN credit risk models are viable alternatives to traditional models. While the results do support the use of ANNs in credit risk model development, care must be exercised when determining which model is the optimal model, as various techniques develop models with performance differences across different data sets and loss ratios and across alternate false positive rates.; The research contributes to the literature by extending prior research on credit model development and ANN connection weight interpretation. Non-linear ANN credit models are interpreted using several connection weight interpretation techniques in order to develop linear first-order credit models that can be interpreted and that can be used to explain an adverse credit decision, which is essential for credit risk assessment decisions.