Deep learning with encoder-decoder architecture for exchange currency rates model predictions

Abstract

Exchange currency rate becomes one of the most important things on country economic growth. In some countries, the rate may affect seriously to economic growth and political stability. Governments need to have some actions in order to stabilize the rate. Knowing the pattern of exchange currency rate is a mandatory by governments for determining their future policies or as consideration in future decision-making. Therefore, government needs an analysis tool for modelling the rate prediction. Our previous study emerged that using deep learning method with encoder-decoder architecture has succeeded to model the rate prediction with teacher forcing and give only a one-step prediction. In this study, the authors aim to predict the pattern of exchange rate using the previous method with zero input to decoder to obtain multiple time steps prediction. After training and testing over 4,344 daily data series of IDR to USD exchange currency rate over 17 years taken from Bank Indonesia official website, the result found that the method still showed an interesting model to predict the next 40 sequence data series of exchange rate. In conclusion, the proposed method can be utilized to tackle the others prediction problem and the resulted model can be applied as an analysis tool by government for predicting their rate for the coming months or years.