Optimizing Stock Price Forecasting using Elman RNN with Distributed Training and Hyperparameter Tuning

The fluctuation of stock prices is influenced by various internal factors, such as tax policies and earnings per share, as well as external factors including economic conditions and political situations. These nonlinear and volatile characteristics present significant challenges for accurate prediction. This study applies a Recurrent Neural Network (RNN), specifically the Elman architecture, to forecast the daily closing prices of Bank Rakyat Indonesia (BBRI) stock using historical data from 2003 to 2024.To enhance computational efficiency, a distributed training strategy using data parallelism was employed, allowing faster model training on large-scale datasets. Additionally, hyperparameter tuning was carried out to optimize model performance. The best-performing model, optimized through extensive tuning, uses 9 time steps, 16 hidden neurons, a learning rate of 0.00011, ReLU activation, RMSProp optimizer, and Xavier Normal initialization. Evaluation results show that the model achieved a Mean Absolute Error (MAE) of 79.18 IDR, a Root Mean Square Error (RMSE) of 107.20 IDR, and a Mean Absolute Percentage Error (MAPE) of 1.58%. Furthermore, distributed training significantly accelerated the training process up to 33 times faster com- pared to conventional single-machine setups. These findings demonstrate the importance of distributed computing and thorough hyperparameter optimization in enhancing the performance of deep learning models for financial time series forecasting.