Consumer Loan Classification Using Artificial Neural Networks

Problems that are difficult to model do not lend themselves to classic algorithmic approaches to their solution. This has fostered the growth of novel techniques that can provide acceptable results based on experience; one such technique uses Artificial Neural Networks (ANNs) that can be “trained” to produce reasonable outputs by presenting them with examples of input data and their corresponding expected outputs. This paper presents the experiments and results obtained by using an ANN to classify the consumer loan and identify potentially risky loans, that is, loans that might not be repaid on schedule. The network was used as a “second level” filter, by supplying it with data of loans that already had been approved by the bank officers. The network optimization technique used is based on the Reactive Tabu Search (RTS), which relies only on the integer arithmetic, does not require computationally expensive derivatives, and is directly executable on the TOTEM neural chip. The combination of TOTEM and RTS is used to train an ANN by presenting 400 loan application cases and their corresponding repayment history. The ANN thus configured is then used for the evaluation of the risk of 600 loans, where the prediction of the ANN is compared with the loan repayment history. The ANN classifies 597 of those loans correctly detecting 35 of them (that the bank had granted) as having problems in their repayment schedule.