CURRENCY CRISES IN GEORGIA: A MULTIVARIATE LOGIT MODEL
After the collapse of the Bretton Woods system, developing countries, including Georgia, experienced several currency crises followed by severe recessions and deteriorated macroeconomic stability. This creates incentives for policymakers to predict currency crises in a timely manner, and avoid them or mitigate their negative consequences. This paper aims to identify episodes of the currency crisis in a panel of the Post-Soviet countries (to create evidence for Georgia), and access predicting power of the various economic, structural and institutional variables. Based on the different versions of the foreign exchange market pressure indices and their critical values, we identified three periods of the currency crisis: 2008-2009, 2015-2017 and 2020 years (with multiple episodes of the crisis). Among the reasons behind these episodes of currency crises, we can highlight: global financial crisis, monetary expansion of the United States, reduced crude oil and commodity prices, armed conflicts between countries in the region, political instability and imposed sanctions, and COVID-19 pandemic. Early warning indicators were chosen based on desk research of the theoretical models, and meta-analysis of the empirical papers. The optimal forecast horizon is 1 year and predicting ability of indicators are assessed employing multivariate logit model. One-year lag of the annual export growth, crude oil price and credit to GDP ratio are significantly correlated with the probability of currency crisis. These early warning indicators have an ability to collectively predict currency crises one year prior. The results of the multivariate logit model are robust under different specifications of the model. In contrast to the theoretical foundation, the lag value of the crude oil prices is positively correlated with the probability of the currency crisis, but narrowing the predicting corridor changes the sign of the correlation coefficient from positive to negative. The most reliable specification of the models successfully predicts 34% of the crisis episodes. Moreover, the model has low Quadratic Probability Score (QPS) and Logarithmic Probability Score (LPS), indicating high level of reliability of the model's outcomes.