A comparative analysis of artificial neural networks and time series models in exchange rate forecasting

Abstract

This study aims to compare the performances in intra-period forecast values of the exchange rate, which is one of the classic linear time series models, ARIMA model, and one of machine learning methods, artificial neural network models. This study uses monthly data for the USD/TRY exchange rates for Turkey over the period from January 2010 to October 2024. Exchange rates are sensitive to unexpected events and political uncertainties in the economy. It is, therefore, highly considerable that the estimation of the future values of exchange rates is important to central banks and companies concerning risk management. ARIMA is a linear univariate time series model that makes in-sample and/or out-of-sample value estimates based on the AR and MA components in the relevant time series data. ARIMA models are models based on a linear component of past data that makes estimates and can yield statistically reliable results. However, all these models are based on the assumption of data having a linear structure and do not give successful results in estimates against nonlinear series. In the literature, it was indicated that the ANN model is more successful than ARIMA models in capturing and predicting the nonlinear structure of the data structure. Since the USD/TRY exchange rate might have a nonlinear structure due to this fact, the use of the ANN model, methods were preferred in this study. According to the empirical findings in the study, the intra-period forecast values obtained with the ANN model on the USD/TRY exchange rate were determined to be more successful than the ARIMA model according to the forecast performance criteria.