Prediction of bond ratings using statistical and neural networks techniques

Abstract

Are the performances of these models good enough to be used as bond rating tools? Apparently classification range 51% to 58.7% is definitely improvable figures. One of the reasons for limited success could be that, these models are too simplistic to capture the relationship between bond ratings and various financial ratios. Conjecturing, these relationships might be so complex that, no functional form is capable of capturing it. Moreover, the data constraint does not provide us the luxury of experimenting with higher order terms. And finally, there are many subjective elements like top management quality, market growth of the firm’s products, which decide the rating of a bond, whose proxy variables are not included in these models due to non-availability of data. All these effectively mean loss of vital information, which might have resulted in such classification percentages.