PERFORMANCE COMPARISON OF MACHINE AND DEEP LEARNING METHODS IN USD/TRY EXCHANGE RATE FORECASTING

Abstract

Accurate estimation of exchange rates is of great importance in terms of economic and financial analysis. Turkey has been faced with serious exchange rate fluctuations, especially in the recent period. At this point, accurate estimation of exchange rates is of critical importance for both individual and institutional investors. The aim of this study is to make a comparative performance analysis of different machine and deep learning methods used in USD/TRY exchange rate estimation. In the study, USD/TRY exchange rate estimation was performed using 149 months of data between January 2012 and May 2024. Total opened USD deposits, M3 money supply, total imports, total exports, unemployment rate, gold price, CPI, PPI and central bank net dollar reserves were used as input variables. Estimates were made with XGBoost, Random Forest, LightGBM, LSTM and SVR methods. In addition, the generalizability of the results was tested using the five-fold cross-validation method. According to the obtained results, the best estimation performance was produced by the Random Forest model in the training, test and cross-validation data sets. This study contributes to the literature by comparing the strengths and weaknesses of different methods in USD/TRY exchange rate forecasting.