A Survey-Driven Ensemble Approach to Predicting Sovereign Debt Distress in Bangladesh

Abstract

Operating out of a new, perception‐aware machine‐learning method of predicting probability of sovereign debt crisis, we integrate survey‐based predictors along with existing ensemble classifiers. Out of a purposive sample of 650 Bangladeshi government officials, financial analysts, academics, and students, we extract demographic information, acquaintance with debt concepts, and multi dimensional risk perceptions in fiscal, political, institutional, and financial dimensions. After categorical conversion via label encoding, treatment of outliers using an interquartile‐range filter, Min–Max normalization, and training and testing XGBoost, LightGBM, Random Forest, and weighted soft‐voting ensemble via five‐fold time‐series cross‐validation regimen, we demonstrate that the ensemble model has the highest cross‐validated training accuracy (0.9528), the same as optimal test accuracy (0.8481), and has weighted F1 score of 0.847, outperforming individual learners and having narrow train–test gap (0.1046). Exploration of the confusion matrix reveals high classification in all five classes for crisis likelihood with specific strengths in classification of “Moderately Likely,” “Likely,” and “Very Likely” outcomes. Adopting the direct incorporation of stakeholder judgments in prediction algorithms, the present study generalizes beyond the usual, data‐driven sovereign‐risk models and offers an early‐warning system via the incorporation of quantitative as well as qualitative characteristics of debt distress. Our research is summed up with the policy implications for proactive risk management as well as sketching the future perspectives, e.g., the leveraging of alternative data streaming in real‐time as well as federated learning architecture.