Refining a neural network credit application vetting system with a genetic algorithm

Abstract

This paper describes how a simulated genetic process is used to automate the configuration and training of a back propagation trained multi-layer perceptron network used for credit application vetting. The network is trained on past loan case data, and is then used to classify the suitability of issuing credit on new loan applications. A prototype scheme for using a genetic algorithm to choose the network geometry and back propagation parameters so as to optimize classification accuracy and speed of convergence is described. This optimization relies upon the genetic algorithm assessing a fitness criteria. The novel fitness criteria that has been developed for this application is described with the associated problems, and some suggestions for future research. The particular genetic algorithm used and its mechanisms are detailed. The performance of the final system is compared with the performance of a manually configured system over common data. The genetic algorithm refined system is seen to outperform the manual system in terms of accuracy, whilst requiring a minimum of operator effort by comparison. Results indicate the successful automation of this aspect of the optimization for such a credit application vetting system, although further investigation into the most suitable fitness criteria is still warranted, so as to incorporate further business information.