A comparative analysis of rule-based, neural network, and statistical classification systems for the bond rating problem
Jun Woo Kim · 1992
Bond rating is characterized as a qualitative decision making process and its problem domain is ill defined. Numerous efforts for simulating judgement using mostly statistical analysis tools have been made during the past decade. However, the results are characterized by low prediction performance. Thus, a new and different approach in this problem domain is needed. The goal of this research is to find a better approach to the bond rating problem. The study objective is to find a comparative measure for classification and prediction performance of statistical classifiers, artificial neural network (ANN) systems, and rule based expert systems using past bond rating data. The performances of the classifications are evaluated with respect to training, classification, and prediction. For this purpose, three random samples of 1988 and 1989 data were collected from Standard and Poor's Compustat financial tapes. The six highest bond ratings (AAA, AA, A, ..., B) were considered. The Belkaoui model (1980) was employed as the classification platform. To meet the research objective, five analytical tools were selected from the three different tool families: three statistical tools (a regression analysis, an ordinal logistic regression analysis, and a discriminant analysis); an artificial neural network system; and a rule based system. The performance criteria applied to the classifiers were (1) the percentage of correct classifications, and (2) the percentages of one (or more) rating difference(s) between the desired and the predicted ratings. To compare these performance data, a series of nonparametric 2 x 2 contingency tables was prepared, and the corresponding $\chi\sp2$ statistics were calculated. In most cases, the $\chi\sp2$ statistics were significant at the.1 level. Based on these results, it is concluded that the artificial neural network system is, in fact, a better tool than the other tools considered for the bond rating problem in terms of the prediction and the classification performance.