Building a System of Leading Indicators for Forecasting the Currency Crisis

This research is devoted to the analysis of financial crises. We examine different classifications of crises, methods of forecasting, approaches to building systems of early warning indicators. To better understand the potential for predicting f inancial crises, we conduct our own empirical research, comparing logit model and random forest to predict currency crises in developing countries. We also identify the most relevant variables, whose dynamics may signal the currency crisis is approaching. We aim to compare the accuracy of econometric models and machine learning techniques in predicting currency crises in developing countries, and to identify a set of relevant indicators that could be used in a warning system. We use logit regression and random forest models. We compare the predictive power of these models using the ROC curve. The significance of variables in a random forest model is determined by the Shapley values. We found that the random forest model has slightly more accurate predictive power than the logit approach. Both models indicate that oil prices and commercial bank deposits are the most robust predictors of currency crises. The results obtained can be taken into account by economic institutions involved in financial system regulation, as we indicate the variables, which should be primarily taken into account when forecasting currency crises in developing countries.