Modelling Time Series Data for Stock Prices Prediction Using Bidirectional Long Short-Term Memory

The dynamic nature of stock markets, characterized by intricate patterns and sudden fluctuations, poses significant challenges to accurate price prediction. Traditional analytical methods are often unable to capture this complexity. This requires the use of advanced techniques capable of modelling non-linear dependencies. This study aims to build a model using recurrent neural network and predict the Indonesian stock prices. PT Gudang Garam Tbk.'s (GGRM.JK) stock was selected due to its significant role in the Indonesian stock market and its contribution to national revenue through excise tax. The method used in this research involves training the BiLSTM (Bidirectional Long Short-Term Memory) model using historical stock price data with training and test data ratios of 90:10, 80:20 and 70:30 to determine the optimal configuration. The evaluation results showed that the 90:10 data ratio gave the best performance with a MAPE of 1.51%, MAE of 343.55 IDR and RMSE of 522.30 IDR. These results indicate that the BiLSTM model has high accuracy and minimal prediction errors. Further analysis showed that the model performed optimally with a batch size of 32 and higher epochs, such as 200 and 250, providing greater stability and prediction accuracy. These results demonstrate the potential of the BiLSTM model as an effective predictive tool to support strategic investment decisions, particularly for high volatility stocks. Future research is recommended to test this model on other stock data and to consider external factors to improve its generalizability.