Reproducibility package for Predicting Debt Distress In Low-Income Countries

We develop an empirical model to predict episodes of debt servicing difficulties (“debt distress”) in low-income countries, with three main contributions to the existing literature. First, we develop more refined measures of external debt distress episodes that allow us to time the onset of distress episodes with increased precision. Second, we develop a systematic algorithm to comprehensively assess the out-of-sample predictive performance of more than 550,000 candidate binary prediction models using J-K-fold cross-validation. Third, we test whether more sophisticated machine learning algorithms can outperform simple probit models. We find that simple single-equation probit models have better predictive power for debt distress than more sophisticated algorithms and are comparable in terms of predictive performance to important policy benchmarks such as the current IMF and World Bank debt sustainability framework for low-income countries.