Predicting Sovereign Debt Distress in Africa: A Machine Learning Early Warning System

Sovereign debt distress has re-emerged as one of the most pressing development challenges in Africa. When a government can no longer service its external obligations, the costs fall on citizens rather than markets: public services are cut, currencies weaken under reserve pressure, and the fiscal room needed for development investment disappears. Identifying which countries are approaching distress, early enough for preventive action, is therefore a problem of direct social consequence. The frameworks that currently anchor multilateral surveillance rely on fixed threshold indicators that tend to signal late and to confuse long-run structural vulnerability with acute, near-term liquidity stress. This paper develops a machine learning early warning system (EWS) for sovereign debt distress across 49 African countries over 2011 to 2023. Its central methodological contribution is a three-layer procedure for constructing the distress label, combining recorded distress events (default and restructuring, IMF emergency programmes, and G20 Common Framework engagements), an independent debt-service ratio screen, and a structured expert-review layer that reconciles disagreements between the two. The procedure yields a composite label of 91 distress observations (14.3 percent) across 637 country-years. Using temporal walk-forward cross-validation to prevent information leakage, a tuned Gradient Boosting model with synthetic minority oversampling attains an area under the ROC curve (AUROC) of 0.940 on 343 out-of-sample predictions, a gain of about six points over the same model trained on event-only labels (0.881) and well above logistic regression. Debt service relative to total reserves is the single dominant predictor (50 percent of model importance, ranked first in all seven temporal folds). Applied to 2023 data, the model places twelve countries at critical risk for 2024. Nigeria remains in the low-risk tier (seed-averaged probability about 10 percent), though its deteriorating debt-service-to-reserves trajectory is visible in the model and warrants monitoring.