«They'll just go to Moody's» : Investigating Corporate Credit Rating Updates Using Machine Learning Techniques

Abstract

Credit Rating Agencies («CRAs») play an important role in the global debt market. They influence the credit spread and thus the borrowing costs for major corporations. An inherent problem is the conflict of interest that arise when the CRAs are paid by issuers. This is not a recent concern, and numerous studies have looked into this and other issues with CRAs. In this master's thesis, we extend this area of research by applying machine learning («ML») models for predicting credit rating updates. For this task, we construct a prediction model using financial ratios, for which we have 20 years of data for two major agencies; Moody's and Fitch. We also include ratings for an investor-paid agency: Egan-Jones. In the model, we change the soft factor in the CRAs' assessment with a new factor that both theoretically and, as will be shown, empirically explain rating updates; trailing stock returns. We apply the XCBoost algorithm to provide more accurate predictions of credit rating updates. Moreover, we analyse SHAP values to interpret different features' contributions to the predictions of rating updates. We evaluate our approach on a dataset of credit ratings in the US and EU and obtain an accuracy of 84.25%. We find that the total return 12 months before the update is the most important when predicting, which suggests stale credit rating updates. Most excitingly, we find that for CRAs with an investor-paid model, the total return three months before the update is the most important when predicting. For the issuer-paid revenue model, twelve months' total stock return turned out to be important: This suggests that investor-paid revenue models are more proactive in updating credit ratings than issuer-paid agencies. The model is applied to the rating downgrade of Wirecard in 2020, which allows for an interesting interpretation of local SHAP values. We also discuss the potential limitations of using ML in credit rating predictions, such as loss of interpretability, unreliable accounting data and the sensitivity of SHAP values.