A corporate credit rating model with autoregressive errors

In this paper we propose a longitudinal credit rating model which accounts for the serial correlation in the ratings. We achieve this by imposing an autoregressive structure of order one on the errors of a multivariate ordinal regression model. The longitudinal structure of the model improves significantly both the goodness-of-fit and predictive performance compared to static models. By modeling the joint distribution of the ratings over time, the framework allows us to obtain predictions conditional on the past rating history of a firm, which clearly out-perform the unconditional predictions both in- and out-of-sample. This shows the importance of incorporating past rating information in the prediction. Another upside lies in the framework’s ability to deal with missing rating observations. A real data example is provided by using a sample of US publicly traded corporates rated by S&P for the years 1985–2016. The determinants of corporate credit ratings are pre-selected using the ordinal version of the least absolute shrinkage and selection operator (LASSO). Additionally, as a model extension we allow the regression coefficients of the selected variables to vary over time in the longitudinal model. This allows us to gain a better understanding of the drivers and evolution of the rating behavior over the sample period. Finally, based on the longitudinal model with LASSO selected variables, we find evidence that S&P exhibits procyclical aspects in their rating behavior.