Predicting Emerging Market Sovereign Debt Crises Using Data Mining Techniques

As any stand-alone indicators are proving ineffective in estimating country's fiscal soundness and particularly debt crisis probability, we identify the need for building a model capable of capturing the indicators' context dependence and interactions. We utilize three data mining modeling techniques – Classification and Regression Trees (CART), Random Forests (RF) and Stochastic Gradient Boosting or Boosted Trees in search for optimal model for predicting sovereign debt crisis. We compare their predictive performance and find Boosted Trees model dramatically outperforming others with overall accuracy of 95% and 98% overall and debt crisis episode prediction accuracy, respectively. Macroeconomic and solvency variables show the highest predictive power (i.e. importance) in the most successful model – in particular reserves over total external debt, external public debt over GDP, M2 over reserves, total external debt over GDP and official exchange rate depreciation