Using Probabilistic Neural Networks and Rule Induction Techniques to Predict Long-Term Bond Ratings

Previous studies have implemented Backpropagation Feedforward Neural Networks (BPNNs) as an alternative technique to Linear Discriminant Analysis (LDA) to predict bond ratings using relatively small data sets due to the limited number of bond ratings data available. Although these studies report the superiority of BPNNs over the LDA to predict bond ratings, they seem to ignore that BPNNs suffer from the problem of overfitting if the data set is small compared to the free parameters of the network. We suggest instead that the Probabilistic Neural Network (PNN) model should be implemented if the data set is relatively small and boundary groups are considered due to the way it uses functions around clusters of the training data. On the other hand, we suggest that the Ripper-Rule Induction (RRI) algorithm should be implemented if the data set is relatively large because not only classifies better than the LDA and PNN models, but also offers interpretable rules that are easy to visualize and understand, as opposed to the inherent inability of Artificial Neural Networks (ANNs) to explain in a comprehensible form the process by which a given decision has been reached. After experimentation, we found that both the PNN and the RRI algorithm classify significantly better than the LDA model while at the same time not suffering from departures from multivariate normality as opposed to the LDA model.