In this paper we face the fitting versus forecasting paradox with the objective of realizing an optimal Early Warning System to better describe and predict past and future sovereign defaults. We do this by proposing a new Regression Tree-based model that signals a potential crisis whenever pre-selected indicators exceed specific thresholds. Using data on 66 emerging markets over the period 1975-2002, our model provides an accurate description of past data, although not the best description relative to existing competing models (Logit, Stepwise logit, Noise-to-Signal Ratio and Regression Trees), and produces the best forecasts accommodating to different risk aversion targets. By modulating in- and out-of sample model accuracy, our methodology leads to unambiguous empirical results, since we find that illiquidity (short-term debt to reserves ratio), insolvency (reserve growth) and contagion risks act as the main determinants/predictors of past/future debt crises