Machine learning versus traditional methods in exchange rate forecasting: applications for the fragile five economies

In 2013, the Federal Reserve Bank (Fed) profoundly impacted the global economy by announcing the termination of its quantitative easing policy. The countries referred to as the "Fragile Five" in economic literature (Brazil, Indonesia, South Africa, India, and Turkey) were severely affected by this decision due to their high dependence on financial capital flows. Fluctuations in interest rates, a key determinant of financial capital flows, have led to increased exchange rate volatility over time. The Fragile Five countries experience deviations from their inflation targets due to the adverse reflection of exchange rate increases on domestic prices stemming from their high import dependencies. These deviations bring critical issues such as reliability, credibility, and reputation for monetary authorities and disrupt macroeconomic balance. Therefore, making reliable forecasts of exchange rates is vital for both monetary authorities and other financial actors. Reliable forecasts are necessary to enhance the success of inflation targeting, ensure the effectiveness of monetary policy, and maintain macroeconomic stability. In this context, modeling based on advanced forecasting methods will guide managing the effects of exchange rate volatility and in the processes of formulating economic policies. In this study, univariate models were constructed using lagged values of the exchange rate, while multivariate models were developed based on the sticky-price monetary model. In univariate models, point and directional forecasts of the exchange rate were emphasized, while in multivariate models, the interaction of macroeconomic variables from Dornbusch's Overshooting Exchange Rate model with the nominal exchange rate and their contributions to the forecasting process were examined. Exchange rate forecasts were conducted using Naïve Drift, Theta, Holt's Winter Exponential Smoothing, ARIMA, Ridge Regression, RNN, LSTM, GRU, CNN, and XGBoost methods. Advanced optimization techniques were employed at every stage of modeling. According to the findings, machine learning techniques demonstrate superior performance in both point and directional aspects in univariate models. Similarly, even when past values of the exchange rate are not included in the model, machine learning techniques exhibit superior directional performance in multivariate models as well. In conclusion, reliable forecasts can be made for both monetary authorities and other financial actors using machine learning methods with advanced optimization techniques without departing from economic theories and the macroeconomics of reserve currency countries.