<![CDATA[Stock price prediction in the Indonesian capital market is highly complex due to the significant influence of market volatility and the non-linear nature of time-series data. Errors in predicting price trends can significantly increase investment risks. This study aims to enhance the accuracy of stock price prediction for blue-chip banking companies (a case study on one of the largest state-owned banks) by addressing the limitations of single models. The proposed method is a Hybrid Deep Learning architecture combining Bidirectional Long Short-Term Memory (Bi-LSTM) to capture long-term dependencies and Bidirectional Gated Recurrent Unit (Bi-GRU) for computational efficiency. To ensure maximal model performance, automatic hyperparameter optimization was performed using the Hyperband algorithm, which utilizes an adaptive resource allocation strategy. The data employed consists of 10 years of historical daily trading data (2014–2024), which underwent Min-Max normalization and a 60-day window size formation. Experimental results demonstrate that the Hyperband algorithm successfully identified the optimal configuration of 128 Bi-LSTM units and 32 Bi-GRU units with the tanh activation function. Model evaluation on the test data indicated a high level of accuracy, with a Root Mean Squared Error (RMSE) of IDR 171.41 and a Mean Absolute Error (MAE) of IDR 142.53. These results confirm that the systematically optimized hybrid approach is capable of minimizing prediction errors significantly better and is reliable for modeling fluctuating stock price dynamics.]]>
