The Prediction of Bond Ratings using Statistical and Neural Network Techniques

Student Name: Shariful Alam

In this thesis an attempt was made to model the CRISIL ratings of long-term bonds issued by manufacturing companies in India in calendar year 2001, using sixteen financial fundamentals covering six different dimensions. Three approaches were adopted for this modeling effort, namely Multiple Discriminant Analysis, Multinomial Logistic Regression and Artificial Neural Networks, after subjecting the sixteen independent variables to principal component analysis to reduce redundancy, which yielded six significant principal components. All the six principal components were found to be useful in predicting the CRISIL rating in all the three models. These models were next validated using the 2002 data. Though the performance of the logistic regression model was comparable to that of the more complex feed forward neural network with one hidden layer, both these models out-performed both Fisher's and Quadratic Discriminant Analysis models in terms of the accuracy of prediction in both the training and validation sets.