Analysis of variables determining the exchange rate with a hybrid method based on conventional time series and artificial neural networks

The study aims to propose a hybrid forecasting method with Conventional Time Series and Jordan Feedback Artificial Neural networks for the variables determining the exchange rate. In this direction, fifteen variables that determine the exchange rate have been reached. The data set for the variables were taken from the Central Bank of the Republic of Turkey's Electronic Data Distribution System (EVDS). The series has consisted of 47 quarterly observations from 2009Q1 to 2020Q3. In the study, stationarity control was made from the Seasonal Research and Conventional Time Series Analysis (CISA) methods. The Toda-Yamamoto Causality Analysis method was used to determine the relationship between the variables. As a result of causality analysis, exchange rate estimation was made using Artificial Neural Network (ANN) with the related variables. Three different models were created for prediction with ANN and the prediction performance of the models was calculated according to the RMSE Error Measurement Criteria. The estimation error of the proposed model (Model 3-1) within the scope of the study was calculated as 0.3062, and the estimation errors of the other models were calculated as 2.7960 and 0.3184. The proposed model has contributed to the literature due to its high predictive power. As a result, it has been seen that CPI, interest, unemployment, current account deficit, GDP, and exports are the most critical variables that determine the exchange rate. In the first part, the study was introduced, in the second part the theories explaining the exchange rate change were discussed, in the third part the literature was examined, and in the fourth part, the subject, purpose, importance, assumptions, scope, and limitations were investigated. In addition, the research was determined, research methods were discussed and empirical findings were included. In addition, research methods have been introduced and empirical findings are included. In the conclusion part, the prediction performances of the models used were compared, and the similarities and differences between the literature and the findings were discussed.