The object of this paper is to develop an Early Warning System of Macro Vulnerability for several Latin American countries based on previous work of Kaminsky, Lizondo and Reinhart (1997) and Kaminsky (1998). We build a composite leading indicator that signals macroeconomic vulnerability, showing that, historically, crises tend to happen in these “vulnerable” situations. The main differences with Kaminsky’s approach are, 1) we use a reduced set of variables to generate the signals, 2) the variables were first aggregated and then the signal was issued depending on the behavior of the composite index, as opposed to Kaminsky’s procedure of generating signals with each individual variable and then aggregating these. The results are very satisfactory on the statistical and operational fronts: on the statistical side, Type I and Type II errors are smaller than those reported in previous papers. Operationally, this system of leading indicators is less costly to maintain given the reduced set of primary variables involved, their wide-spread availability and the timeliness with which they are reported. The models’ out-of-sample predictive ability are tested with plenty of crises cases occurring after the first stage of this project was finished: Colombia (September, 1998), Brazil (January, 1999), and Ecuador (February,1999). All the cases were correctly anticipated by the signals of vulnerability issued by the models. Additionally, Mexico’s models , estimated with information available two years before the crisis, show that these signaling devices would have been useful anticipating the 1994 crisis.
