Governance and Currency Crises in Latin America Post-Nineties: A Machine Learning Approach

The paper provides evidence of currency crises in 11 countries in Latin America from 2002 to 2020. The research compares four machine learning classifiers to identify which better classify a new observation as a currency crisis when considering governance and macroeconomic variables. The random forest model outperforms other models in the assessment, even in identifying imbalanced data within our interest class. The model yields a misclassification rate of 1.82%, a balanced accuracy of 99.04%, and a kappa value of 84.76. The results reveal the governance aspects associated with the likelihood of currency crises: poor perception of political stability, corruption, and government effectiveness. Moreover, it is confirmed that export growth, foreign reserves growth, and external debt remain relevant macroeconomic predictors of currency crises post-nineties.