Statistical Models of Bond Ratings: A Methodological Inquiry

Abstract

Published financial statements of corporations are supposed to provide relevant information to a variety of users. Empirical studies indicate that earnings numbers of corporations are included in the information set of investors when setting security prices of a firm's common shares (see Kaplan [1978] for a survey of these studies). No generally accepted formal models exist, however, to show how to use the information in a firm's financial statements to determine what good investments might be. The efficient market theory implies that one cannot use publicly available financial data to choose common stocks which will outperform the stock market after controlling for risk effects. Financial statement analysis, which traditionally has consisted of extensive computations of a myriad of ratios and statistics to detect underor overvalued securities, is presently in a less than satisfying state. What helpful information can be gleaned from systematic examination of firms' financial statements? Practitioners emphasize the complex and subjective nature of their analysis which cannot be reduced to mathematical formulas. Nevertheless, research has indicated that formal analysis of accounting and Previously developed models for predicting bond ratings are summarized and criticized. A statistical procedure appropriate to the ordinal nature of a bond rating is applied to a recent sample of seasoned and newly issued bonds. A simple linear model using a subordination dummy variable, total assets, the long-term debt to total-assets ratio, and the common stock systematic risk measure can correctly classify two-thirds of a holdout sample of newly issued bonds. Further analysis reveals that the model may be predicting the actual risk of a bond better than the rating agency in about half of the misclassifications.