The main purpose of constructing "Early Warning Systems" (EWSs) for financial crises is to provide policy makers with some lead time to take pre-emptive actions that would help avoid, or at least mitigate, the damages of an approaching crisis. Accordingly, this study empirically evaluates and compares the effectiveness of the econometric models developed so far to construct EWSs. In addition, a more accurate (dynamic-recursive) forecasting technique is developed to generate better out-of-sample warning signals for currency, banking, and sovereign debt crises in the different regions of the world. The empirical analysis shows that the predictive performance of the EWS is significantly improved when using simple pooled models that account for the heterogeneity of the signalling indicators across the different regions. Moreover, including the entire crisis period in the sample outperforms the more common practice of dropping post-crisis-onset periods or using a multinomial specification of the crisis variable. In addition, the findings reveal that our dynamic-recursive technique provides more accurate out-of-sample forecasts for logit models. Finally, the dynamic signal extraction approach is recommended for policy makers who value avoiding financial crises at all costs, while the binomial logit model is more suitable for less conservative policy makers who consider the economic and social costs of false alarms.