Foreign exchange rate forecasting with deep learning approach: The case of Turkey

This thesis investigates the effectiveness of deep learning models for exchange rate forecasting in the context of the Turkish economy. The exchange rate is a critical variable in economic stability and macroeconomic policy decisions and is determined by a complex process with nonlinear dynamics. Since traditional econometric methods cannot adequately capture this complexity, deep learning techniques have been increasingly preferred in exchange rate forecasting in recent years. In this study, RNN-based models such as LSTM, GRU, Bi-LSTM, Bi-GRU as well as hybrid models combined with CNN are used to predict USD/TL and EUR/TL exchange rates. The data set used in the study consists of the daily closing prices of the Central Bank of the Republic of Turkey and is analyzed in three periods according to the structural transformations in the Turkish economy: The transition to a flexible exchange rate regime in 2001, the introduction of financial stability policies in 2010 and the economic fluctuations after 2016. The findings show that the inclusion of sudden fluctuations in the training set improves the performance of the models in short-term analyses, while in long-term analyses, the models are more insensitive to sudden changes. Especially in long-term and medium-term analyses, CNN-based hybrid models provide higher accuracy than traditional RNN-based models thanks to their feature extraction capabilities. When the performance of the models is evaluated with MAE, MSE, RMSE and R² metrics, it is found that hybrid approaches such as CNN/Bi-LSTM and CNN/Bi-GRU provide the best prediction results. This study reveals the potential of deep learning in exchange rate forecasting and sheds light on future studies in this field.