A Multiclass Machine Learning Approach to Credit Rating Prediction

Abstract

Corporate credit ratings are important financial indicators of investment risks. Traditional credit rating models employ classical econometrics methods with heteroscedasticity adjustments across various industries. In this paper, we propose using machine learning techniques in predicting corporate ratings and demonstrate, empirically, that multiclass machine learning algorithms outperform traditional econometrics models in exact, 1-notch, or 2-notch away rating predictions. We use three years of CompuStat data from four very different industries and compare corporate credit rating prediction tasks across linear regression, ordered probit model, bagged decision tree with Laplace smoothing, multiclass support vector machines (SVM), and multiclass proximal support vector machines (PSVM). Our findings show that with the proper multiclass and heteroscedasticity adjustments, the computationally inexpensive multiclass PSVM can be utilized in making viable automated corporate credit rating systems for todaypsilas vast marketplace.