News Sentiment-Enhanced GRU Encoder-Decoder for Advanced USD/IDR Exchange Rate Forecasting

Accurate prediction of a currency's exchange rate movement is crucial to various stakeholders, especially in the financial sector. Successful prediction could provide an insight into the market's movements and thus give a competitive advantage to traders. Unfortunately, this task proves to be challenging due to the numerous factors affecting a currency's exchange rate. This study explores the impact of news headline sentiments towards exchange rate prices. This study focuses on improving USD/IDR exchange rate predictions using Recurrent Neural Networks (RNN) and its variations by adding sentiment data. To achieve this goal, the research would process relevant financial news headlines using FinBERT to obtain sentiment data. The sentiment data would then be integrated into historical USD/IDR close prices. Afterwards, several multistep forecasting models are trained to predict the exchange rate prices 12 days ahead. 6 different RNN models were trained, two sets of LSTM, GRU and encoder-decoder GRU models are each trained on sentiment-added and historical price data. The results are evaluated using three evaluation metrics which are RMSE, MAPE and MAE. The results demonstrate that models trained on the added sentiment data obtained lower errors when forecasting USD/IDR prices, with the best model achieving an RMSE of 299.0. In particular, the encoder-decoder models achieved impressive long-term forecasting performance where it managed to maintain a stable error score.