Performance comparison of machine learning approaches to bond rating

In this paper, we compare the prediction performances of analog concept learning systems (ACLS) and neural networks in real world bond rating problem. For ACLS experiments, we revise the ACLS, which pruning methods-cost complexity pruning, reduced error pruning, pessimistic error pruning, and production rule-are incorporated into. For neural networks experiments, we conduct experiments for two cases: one is for using the original whole data. We call it the conventional neural networks (CNN). This CNN approach is to categorize a new instance as one of the predefined bond classes, but it has limitations in dealing with the ordinal nature of bond rating problem and as the number of classes to be recognized increases, the predictive performance decreases. To alleviate such difficulties, we partition the original whole data in a pairwise manner for training the neural networks. We call it partitioned neural networks (PNN). Experimental results show that the predictive performance of PNN is best among other machine learning methods as well as statistical methods. The PNN has two computational methods and we discuss under which circumstances one method performs better