Forecasting Exchange Rate Depending On The Data Volatility: A Comparison Of Deep Learning Techniques

Abstract

Abstract The prediction of the foreign exchange rate is critical for decision makers since international trade is a vital task, and an accurate prediction enables effective planning of the future. To model the exchange rate behavior over time, a deep learning methodology is used in this study. Deep learning techniques can uncover indeterminate complex structures in a dataset with multiple processing layers. Traditional artificial neural networks (ANNs) do not consider the time dependence between data points in time series data. To overcome this problem, deep learning tools, such as recurrent neural networks (RNNs), consider long-term time dependency in the data. In this study, among the types of RNNs, long short-term memory (LSTM), bidirectional LSTM, and gated recurrent units (GRUs) are used to predict time series data of USD/TRY and EUR/TRY. This prediction is conducted for three different periods in the last 11 years in Turkey. One period includes near-steady data, and two periods have volatile exchange rate data. The prediction performance of the models is evaluated based on the mean absolute error (MAE), root square error (RMSE), and mean absolute percentage error (MAPE) metrics. After the comparison of different models, the bi-LSTM and GRU models are found to yield the most accurate predictions in volatile periods, depending on the nature of the volatility. This study proposes new models for exchange rate estimation and compares the performance of each model based on the volatility of the data.