A statistical approach to modeling the behavior of bond raters

Abstract

Since the early 1970s the literature concerning statistical models of bond ratings has burgeoned. The thrust of these studies has been the attempt to predict as accurately as possible the rating assigned to a particular firm’s bond issue based on readily available data concerning that firm’s financial characteristics.’ In concentrating on accurate prediction, however, these studies have tacitly assumed that the bond rater’s assessment of default risk is perfectly accurate. As a result, those variables that are theoretically important in directly affecting default risk are the ones employed as predictors, and the statistical model judged as best is the one having the highest percentage of correct ratings predictions. We have no objection to this approach per se; correctly predicting bond ratings is clearly important, and these studies have contributed dramatically to our understanding of the bond-rating process. Perhaps because of their interest in accurate prediction, however, these studies have paid little attention to several questions of interest and potential importance both to rating agencies and to the firms whose debt is rated. These questions are as follows: What variables significantly affect a rater’s assessment of a bond’s quality?* What is the relative importance of each of these variables in determining the position of the bond in the rater’s implicit quality hierarchy? What is the magnitude of the effect of an incremental change in a particular financial variable on the probability of a firm’s issue receiving a given rating? Answers to these questions should be of interest to the financial manager of a debt-issuing company to the extent that the ability to control key financial variables implies some ability to control the rating assigned to its bonds. Likewise, answers to these questions should be of interest to rating agencies for establishing objective norms for their raters.3