Predictive Modelling of Local Currency Sovereign Debt yields using Machine Learning

Abstract

Listen

Translate article

This paper explores applying Statistical Machine Learning (ML) models to forecast local currency 10-year bond yields in 17 Emerging Markets (EM) economies, using Economic & Financial Indicators as feature parameters. The models considered include Random Forest (RF), Gradient Boosting (GB), Lasso Regression, Ridge Regression, and Linear Regression (LR). The performance is assessed using three key metrics: Mean Squared Error (MSE), R2 on the testing set, and R2 on Cross-Validation (CV). The results show that RF and GB models generally outperform traditional linear models, showing lower MSE and higher R2 values, indicating better accuracy in their predictions. By exploring the practical application of those models in the context of EM economies, this paper contributes valuable insights into the predictive modelling of sovereign debt yields, which is crucial for investors and policymakers. The findings suggest that ML models, with their ability to capture non-linear relationships, offer a more reliable approach to predicting 10-year bond yields in EM economies.