This paper provides a comprehensive early warning system (EWS) that balances the classical signaling approach with the best-realized machine learning (ML) model for predicting fiscal stress episodes. Using accumulated local effects (ALE), we compute a set of thresholds for the most informative variables that drive the correlation between predictors. In addition, to evaluate the main country risks, we propose a leading fiscal risk indicator, highlighting macro, fiscal and institutional attributes. Estimates from different models suggest significant heterogeneity among the most critical variables in determining fiscal risk across countries. While macro variables have higher relevance for advanced countries, fiscal variables were more significant for Latin American and Caribbean (LAC) and emerging economies. These results are consistent under different liquidity-solvency metrics and have deepened since the global financial crisis.