Abstract:

Kamstra et al. (2001) developed a bond rating classification model that was based on a similar model developed by Ederington (1985). While both studies use Moody's bond ratings as dependent variables, the studies differ with respect to the independent variable data source, that is, Kamstra et al. (2001) use financial statement data extracted from Moody's Industrial Manual (now known as Mergent) while Ederington (1985) uses financial statement data extracted from Compustat. Given this, and given the divergent results of the two studies, the following question must be addressed: Do different data sources yield models that differ considerably with respect to overall performance of bond rating classification?

New bond issues for the period January 2004 to June 2006 that are common to both the Moody's bond rating database and the Standard and Poor's (S&P) bond rating database are included in the analysis. The most recent annual financial statement data reported prior to the issuance of each issue were extracted from both the Mergent database and the Compustat database. Using ordered logit models, I determine the predicted probabilities of the bond ratings, as well as the correct classification rates using the Kamstra et al. (2001) bond rating model for each of the following four data source combinations: Moody's/Mergent; Moody's/Compustat; S&P/Compustat; and, S&P/Mergent.

The results of the Wilk's Lambda test show data source dependency while the results of McNemar test did not show data source dependency. The difference in test results may relate to the level of precision of the tests, that is, the Wilk's Lambda test focuses on the predicted probabilities of the bond rating while the McNemar test focuses on the bond rating categories. Since the predicted probabilities of the bond ratings represent compositional data, I transformed the data in order to avoid spurious interpretations. The need to transform compositional data was not an issue in the previous literature since that research focused on bond ratings and not the predicted probabilities of the bond ratings.