<![CDATA[This study investigates the performance of two deep learning architectures, namely Simple LSTM and Stacked LSTM, for Indonesian gold price forecasting, with a particular focus on evaluating the effect of optimizer selection and learning rate configurations. An experimental framework is implemented using daily Indonesian gold price data from 2021 to 2024. Model performance is assessed using five-fold expanding window time series cross-validation to ensure robustness and avoid data leakage. Four adaptive training optimizers (Adam, Nadam, Adamax, and RMSprop) are evaluated across three learning-rate settings as part of a systematic sensitivity analysis of training hyperparameters. The results indicate that the Simple LSTM consistently outperforms the Stacked LSTM. The best performance is achieved by the Simple LSTM using the Adam optimizer with a learning rate of 0.01, yielding an RMSE of 9.235, MAE of 7.060, and MAPE of 0.71%. These findings demonstrate that simpler architectures combined with appropriate training configurations can provide superior forecasting accuracy for volatile financial time series.]]>
Copyrights © 2026
