Enhancing Stock Market Forecasting with LSTM and BiLSTM Models: A Deep Learning Approach

Stock market prediction is challenging due to the market's non-linear and volatile nature. Traditional methods, such as Mean Baseline and Linear Regression, often fail to capture the complexities of stock price movements. To address this, Long Short-Term Memory (LSTM) networks, a type of Recurrent Neural Network (RNN), were employed for stock price forecasting, leveraging their strength in modeling time- series data. Our findings reveal that LSTM models outperform traditional approaches, while Bidirectional LSTM (BiLSTM) models achieve superior accuracy by processing data in both forward and backward directions. Performance evaluation of the LSTM_6 model yielded RMSE of 299, MSE of 89,528, MAE of 218, and Directional Accuracy (DA) of 0.504. Incorporating intermarket features, such as E-mini Nasdaq 100 Futures, E- mini Dow Jones Futures, Volatility Index (VIX), and Dollar Index (DXY), further improved outcomes, with the LSTM_6_Intermarket model achieving RMSE of 138, MSE of 19,282, MAE of 110, and DA of 0.506. Remarkably, the BiLSTM model outperformed both, achieving RMSE of 124, MSE of 15,411, MAE of 95, and DA of 0.512 without intermarket data, highlighting its ability to capture patterns more effectively. These results demonstrate that LSTM-based models, particularly BiLSTM, are promising tools for stock closing price prediction in regression tasks, with the potential for further enhancement through dataset size, feature selection and intermarket data integration.