Evaluation of Deep Learning Optimizers for Predicting JISDOR Exchange Rates Using LSTM Networks

This research explores the application of four optimization algorithms—Adam, Nadam, RMSProp, and SGD—on a Long Short-Term Memory (LSTM) model to forecast the Jakarta Interbank Spot Dollar Rate (JISDOR). The volatile nature of exchange rate data, influenced by global and domestic economic dynamics, necessitates the use of models like LSTM that excel in capturing both short- and long-term dependencies. Performance was assessed using metrics such as Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). Among the optimizers, Nadam proved to be the most effective, achieving the lowest RMSE of 62.767 and a MAPE of 0.003, indicating its capability in managing complex fluctuations in the dataset. Despite Nadam's promising results, opportunities for improvement remain, including the inclusion of additional input variables, fine-tuning model parameters, and expanding the training dataset. This study underscores the critical role of selecting appropriate optimization algorithms for enhancing the accuracy of LSTM models in forecasting volatile financial time-series data, particularly for currency exchange rates