Predicting Corporate Credit Ratings: A Comparative Study Between Ordered Probit, Neural Network and Random Forest

Abstract: This thesis compares the prediction accuracy for corporate credit ratings between three different models. The two first models, a traditional statistical model called ordered probit and a machine learning model called artificial neural network has been used with success before. The third model, a machine learning model called random forest is implemented and compared to the previous models. The random forest model accurately predicts 66% of all credit ratings in a holdout samples outperforming ordinal probit (58%) and artificial neural network (63%). McNemar's test validates that the accuracy of the random forest is significantly different from the ordinal probit at a 0.1% significance level and artificial neural network at a 5% significance level. The random forest also provides evidence that market value and equity volatility are important when predicting S&P credit ratings.