Hybrid neural network-driven reasoning approach to bankruptcy prediction: comparison with MDA, ACLS, and neural network

The objective of this paper is to propose a new neural network-based approach to bankruptcy prediction problem, named HYNEN (hybrid neural network-driven reasoning) model which is based on two types of neural networks: unsupervised and supervised neural network. Accordingly, it consists of two stages: 1) clustering neural network (CNN) stage, and 2) output neural network (ONN) stage. CNN categorizes input sample into an appropriate cluster, which is identical to finding a relevant rule to be fired in knowledge base. Then in the ONN stage, ONNs are built based on information about the clusters derived from CNN stage, and used to make a final decision: “bankrupt” or “non-bankrupt”. CNN uses two types of unsupervised neural network models for pattern clustering, the self-organizing map and learning vector quantization, and then learns the clusters in a supervised manner. ONN utilizes a supervised neural network. We performed comparative experiments with Korean bankruptcy data using HYNEN, MDA (Multivariate Discriminant Analysis), and ACLS (Analog Concept Learning System) conventional neural network approach