Predicting the Asian Currency Crises with Artificial Neural Networks: What Role of Function Approximation?

Abstract

Indicators approach and probit and logit models are the two most widely used early warning systems for currency crisis.

Several studies however suggest that despite their strengths these systems have high error rates in prediction. In this paper a recurrent neural network model is proposed as an alternative. Consistent with the notion of model uncertainty the system does not rest on a pre-specified mathematical relationship between the macroeconomic fundamentals and probability of crisis. Under such conditions the neural network model, by virtue of its unique property, serves the purpose of approximating a non-linear function that is assumed to be unknown. The system is implemented for prediction of currency crises in the emerging market economies of Africa, Asia, and Latin America, and the results are compared to those obtained with several probit models by Berg and Pattillo (1999). The recurrent neural network model is found to perform better than the probit models with respect to goodness of fit measures, both within sample and out of sample. In particular, the probit models are outperformed in predicting the1997 Asian crises out of sample. The best of probit models and the neural network (NN) model perform almost equally well in predicting the Asian crises with a probability of 20% -- but the NN model, in addition, predicts a significant percentage of crises with a probability of 50%, whereas the probit models fail completely. We conclude that the NN model, which approximates an unknown non-linear function, has greater chances of performing well out of sample in terms of distinguishing between a ‘tranquil’ month and a ‘pre-crisis’ month. However, both systems fail to predict the crises systematically.