A Novel Hybrid Learning Approach to Train Hierarchical Mixture of Neural Networks Model for Credit Scoring in Bank Industry

Credit risk in bank industry is the probability of non-repayment of obligations by customers at specific time. It is one of the most important hazards for banks and private institutes. Due to huge bulk of banks’ overdue receivables, establishment of a Credit Scoring (CS) system is one of the most important means of controlling such a risk. This paper uses the powerful Neural Networks in predicting and mixing them, which can classify customers in two groups of customers who pays their debts on time and customers who don’t. The used model, which has modular based structure and training, is named Hierarchical Mixture of Neural Networks (HMNN) . In mentioned model, for the decomposition of problem among networks and combining results to achieve the final prediction and also the method of training of it uses new approach. The purposed approach applies Binary Particle Swarm Optimization (BPSO) for dimension reduction and decomposing the problem among modules at first, then using the modulation of the training rules specific to each module and the general training rule of this network. Results are achieved in compersion with Multi-Layer Perceptron and Laterally Connected Neural Network.Based on observed results, the suggested model could predict customers’ behaviour with punctuality.