Indonesian Stock Index Price Prediction Using the Stacked Bidirectional Unidirectional Long Short-Term Memory (SBU-LSTM) with the GDELT News Sentiment

Sentiment has been empirically proven to be a significant contributor to excess market volatility, yet extracting relevant information from diverse news sources remains challenging. Leveraging the GDELT event dataset, our study overcomes this hurdle by using a structured repository of news events. Previous LSTM models achieved commendable MAPE scores (1.5% to 5%) in stock price prediction but face challenges when applied to comprehensive stock price index. This research enhances stock price index predictions using the Stacked Bidirectional Unidirectional Long Short-Term Memory (SBU-LSTM) model with sentiment variables from the GDELT dataset. The SBU-LSTM model can capture complex pattern and both long and short term dependecies using the bidirectional LSTM (BDLSTM) and LSTM stacked as a neural network. The optimal model configuration includes a primary layer of BDLSTM with 256 units and a final LSTM layer with 256 units. Using the RMSprop optimizer, a learning rate of 0.0001, a window size of 10, a batch size of 64, a 3-day time lag, and 25 epochs, our configuration achieves an average MAPE of 0.81% in 48.01 seconds. Notably, using only the optimism sentiment variable results in a slightly lower MAPE score of 0.80%. Intriguingly, our study outperforms the SBU-LSTM research paper (MAPE of 5.67%), highlighting the SBU-LSTM model’s efficacy in accurate stock price prediction.