An improved method for developing neural networks: The case of evaluating commercial loan creditworthiness

Neural networks have proven to be a worthy alternative to traditional statistical techniques, such as regression and discriminant analysis, for prediction and classification problems. Unfortunately, neural network architectures are often chosen based upon conventional rules-of-thumb which limit the predictive power of the resulting model. As a means of overcoming the poor development of neural network models, this study describes and uses a systematic neural network development methodology. The methodology is presented via the study of a particular application of neural networks-determining the creditworthiness of commercial loan applications. The ability of humans to evaluate creditworthiness accurately is poor, and statistical techniques only help slightly. A neural network model is well suited for this type of problem. The results indicate that the proposed development methodology produced a neural network model that does a respectable job of determining creditworthiness in a very difficult problem situation.