Previous literature on using economic variables to predict currency crises tend to be very sample specific. In other words, what one dataset confirms to be a predictor of a crisis tends not to work when applied to a different dataset. The authors use extreme value theory (EVT) to determine which of the 18 variables named in the literature correlate with the exchange rate in extreme situations, such as a currency crisis. The exchange rate is measured in three different ways: the rate of return on the exchange rate, the exchange rate pressure index (includes the interest rate differential and M2 measures), and the real exchange rate market pressure index (same as the previous measure but adjusted for inflation).

Two variables emerge as having some correlation with the exchange rate return in EVT: the lagged real interest rate (1 to 15 months, lagged 8 months performing the best) and the lagged real interest rate differential (lagged 4 months). EVT differs from other methods in that a distribution does not need to be defined for the variables, and the measure of an extreme value is defined by the data and not by a multiple of the standard deviation. This distinction is critical when looking at out-of-sample data because the predictors favored in EVT differ from the best predictors used in the previous distribution-defined methods. Out of sample, the predictors that EVT favors perform better but do lack a certain degree of precision because false alarms are still very frequent.

The EVT logic is very appealing in that the standard definition of correlation fails when the correlation really matters (i.e., in a crisis situation). For example, passengers on a boat can be very uncorrelated but then become very correlated if the boat starts to sink. But the methodology requires a great amount of data, and although the theory finds better predictors of currency crises, these predictors are not very accurate. But knowing which of the 18 possible predictors of a crisis is best is still valuable information.

Given the lack of “very good” predictors of currency crises, investors should be aware of this risk in portfolio construction.

The authors use monthly data from 46 countries between 1974 and 2008. Initially, all of the data are pooled to produce an in-sample EVT analysis. To produce the out-of-sample analysis, they generate the EVT procedure for (pooled) data between 1974 and 1994 and then use the results from the 1974–94 sample to predict crises occurring between 1995 and 2008. Note that the out-of-sample analysis is only performed on emerging markets.

The two variables that EVT favors perform the best in predicting crises out of sample but do have a high frequency of false alarms. Different attempts to change crisis threshold values or to combine variables to reduce the frequency of false alarms with the EVT preferred variables deteriorate the prediction power. Consequently, no real solution to reduce the frequency of false alarms given by the EVT preferred variables is found.

The analysis is interesting, and the logic behind EVT is very appealing. But the demands on the data make me question whether there can be a practical implementation of EVT. Further research in this area may produce such a result.