Abstract
This study examines the performance of fourteen (14) Machine Learning (ML) techniques in comparison to the traditional statistical methodology, Logistic Regression (LR), for analysing and predicting future sovereign debt ratings (SDR). The negative impact of speculation and pessimistic expectations regarding sovereign ratings can have detrimental effects on macroeconomic indicators and lead to disruptions in the monetary or fiscal system, ultimately resulting in financial instability. This situation highlights the importance for developing nations to anticipate and forecast changes in sovereign debt ratings to mitigate the negative consequences of a downgrade. By analysing macroeconomic variables and SDRs in South Africa's Quarterly data from 1999 to 2022, this study aims to determine the model with higher precision and superior analytical capacity. The data for South Africa's sovereign debt rating was obtained from the three major debt rating agencies (DRAs), including Moody Ratings, Fitch Ratings, and Standard & Poor's. Macroeconomic indicators were obtained from Thomson Reuters, the South African Reserve Bank (SARB), Statistics South Africa (Stats SA), and Quantec Easy Data. The data was divided into a test set and a training set, with a ratio of 75:25, respectively. The comparison between statistical and ML techniques was conducted using performance measurements such as accuracy, sensitivity, specificity, precision, and the area under the curve (AUC). The study discovered that ML techniques exhibit superior precision and outperform traditional statistical models. However, the effectiveness of these techniques is contingent upon the specific analysis and data employed. ML techniques offer a wide range of possibilities, but the excessive reliance on assumptions in traditional models can hinder their performance. The study highlighted that DRAs employ various methodologies and variables when evaluating sovereigns.