Determining real exchange rate misalignments and predicting currency crises in Eastern Europe: Statistical and artificial intelligence methods

The dissertation is devoted to estimating the long run equilibrium real exchange rates and the corresponding short run misalignments, as well as analyzing and predicting the currency crises in the late 1990's in the four emerging market economies of Bulgaria, The Czech Republic, Romania, and Russia. The long run equilibrium exchange rates are estimated with co-integrating equations; the movements of the exchange rates in the short run are estimated with suitable ARIMA and GARCH error correction specifications. Misalignments are calculated as the short run deviations of the exchange rates from the long run equilibrium values corresponding to the sustainable values of the macroeconomic fundamentals. As an exercise, the post-crisis exchange rates are forecasted based on the pre-crisis error correction specifications. The currency crises are explained and predicted by adopting two alternative approaches. First, a clustering algorithm (sequential exception technique) is implemented on the individual indicators to generate 'signals' prior to the crises. This is further modified in combination with a genetic algorithm to determine the set of indicators that could be taken as the best in predicting the crises. Second, different classification methods are applied to predict the currency crises. In particular, the K-nearest neighbor classifier, the Bayes' classifier and the multi-layered feed-forward neural network classifier are employed as an alternative to the logit/probit models. The experimental results suggest, however weak the sample data, currency crises in all the countries under consideration can be predicted well in advance, based on both within-country and cross-country time series. Both the methods do reasonably well with respect to the percentages of crises called and percentages of right alarms. Under the second approach, neural network performs the best. It is found that none of the three generations of theoretical models alone can be taken to be explaining the crises. Further experiments and investigations are warranted in consideration of possible alternatives to the 'signal approach' and the approach based on probit and logit models.