Previous studies have examined different statistical models to predict corporate bond ratings. However, these papers use agency ratings as the benchmark to assess models and ignore the evidence that agency ratings may not be accurate in a timely manner. In this paper, we propose a new approach which incorporates ex-post bond returns to evaluate rating prediction models. Relative rating strength portfolios, formed by buying under-rated bonds with agency ratings lower than model ratings and selling over-rated bonds with agency ratings higher than model ratings, are employed to test the performance of different statistical models in rating predictions. Our results show that one version of multiple discriminant analysis model can generate a statistically significant abnormal return of 5% over a 5-year horizon. The ordered probit model which is believed to possess theoretical advantages in classifying bonds does not perform better. This suggests that using traditional measures to evaluate models can be misleading. The existence of a profitable trading strategy also raises the concern of market efficiency in the corporate bond market.
Review of Pacific Basin Financial Markets and Policies, 7(2), 153-172