Comparative Deep Learning Analysis: Unveiling the Power of LSTM, BiLSTM, GRU, and BiGRU for Agricultural Stock Price Forecasting on the Indonesian Stock Exchange

This study aims to analyze the performance of deep learning algorithms in predicting agricultural sector stock prices on the Indonesia Stock Exchange (IDX) by comparing four models: Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), and Bidirectional GRU (BiGRU). Daily historical data of six agricultural sector stock issuers (AALI, BISI, DSNG, LSIP, SIMP, SSMS) for the period 2017–2025 was used as the dataset. The research methods included data pre-processing (normalization, 80:20 training-test data split), model training with optimal hyperparameters (unit=512, dropout rate = 0.3, epoch = 50–150, learning rate = 0.0001), and evaluation using Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), R² Score , and computation time metrics. The results show that BiGRU is the most accurate model with the lowest RMSE (7.43–17.20) and the highest R² (0.99 on BISI and SSMS), thanks to the Bidirectional architecture that processes bidirectional data to capture complex temporal patterns. However, GRU is more efficient with a training time of 40–43 seconds, suitable for real-time applications . LSTM and BiLSTM have lower accuracy, especially on volatile stocks such as DSNG (RMSE LSTM = 130.51). This study provides practical recommendations: BiGRU for long-term investment strategies that prioritize accuracy, while GRU for quick decisions based on efficiency. Theoretical implications strengthen the effectiveness of the Bidirectional architecture in financial time series analysis