Predicting Corporate Credit Ratings withMachine Learning Algorithms

Abstract: This thesis investigates the performance of machine learning models in predicting long-term issuer credit ratings, relative to the that of traditional statistical modeling approaches. Our dataset consists of 3,992 ratings by S&P, Moody's and Fitch of American non-financial, non-governmental companies, in the period 1 January 2010 through 1 September 2016. 20% of the dataset is used strictly as an out-of-sample set, in order evaluate the models' performance. We find that our best-performing machine learning model, the ExtraTrees algorithm, achieves an accuracy of 37% when predicting over 16 classes - significantly better than our highest performing statistical method, multiple discriminant analysis, which had 27% accuracy. When predicting over 6 and 2 separate classes, the best-performing models achieved accuracies of 70% and 92%, respectively. These results are in line with previous research on the topic, but are the results of training on a significantly larger dataset. Whereas our results, and past studies show that a relatively high degree of accuracy is possible, the specific implications and possible applications are still unclear.