Forecasting Sovereign Credit Ratings With Machine Learning: A Transparent and Data-Driven Approach

Abstract

Sovereign Credit Ratings (SCRs) are critical risk indicators in global finance, and their opacity and limitations, motivate the development of independent forecasting models. This thesis uses a data­driven Machine Learning (ML) approach to forecast SCRs assigned by Fitch Ratings. It starts with a baseline dataset based on Fitch’s published rating methodology, then progressively expands it with macroeconomic features and, finally, high ­frequency financial data. A central contribution is the evaluation of cross­-validation strategies for panel data. We demonstrate that commonly used methods like random hold­out significantly inflate performance estimates due to data leakage, advocating for and utilizing a recursive window approach to provide a more robust assessment of out-of-­sample accuracy. Our empirical results showed that ML models consistently outperformed the econometric baseline in forecasting SCR, particularly when including broader macroeconomic features. High frequency financial features improved some models but did not consistently enhance forecasting accuracy. Feature importance analysis confirmed that key predictors, such as institutional quality, sovereign size and wealth and financial indicators like CDS spreads, align with established economic and financial theory. Surprisingly none of the ML models beat the no­ change benchmark, due to the static nature of SCR. We access rating transitions, where ML models showed mixed performances. Therefore, while ML holds promise as a complementary tool in SCR forecasting, future research should explore their integration into early warning systems and the use of calibrated probabilities to support, rather than replace, expert­based SCR forecasting.