Consumer Loan Analysis Using Neural Networks

Abstract

Neural networks show promising results in areas where decisions have to be made on the basis of fuzzy or incomplete information. Their structure is quite simple and the research in this area is rapidly increasing; however, little is known about the optimal network configurations, network size, number of layers, number of neurons, best type of network and about optimal training techniques. Hence, the designer still has to rely on empirical evidence or even on trial and error to develop neural networks for practical applications. This chapter presents the basic facts about neural networks and introduces the features of the PC-based network BrainMaker. It also presents classical techniques for consumer loan evaluation and the results of empirical studies based on a sample of 1.000 consumer loans. Empirical results based on a sample of 1.000 loan tuples containing 20 loan characteristics allowed the comparison of the neural network approach to other classification techniques. It turned out that neural networks are highly competitive and most of the time performs better than classical algorithms. A PROLOG system is used for some sort of initial screening. Loans passing this test are then fed into a combination of two neural networks. If the results of both networks agree, then the loan is considered as automatically classified. Otherwise, further checks need to be performed by the loan officers. Based on the sample used 82% of the loans could be classified automatically.