Improving stock price forecasting with M-A-BiLSTM: a novel approach

Source abstract

Stock price prediction plays a crucial role in investment, corporate strategic planning, and government policy formulation. However, stock price prediction remains a challenging issue. To tackle this issue, we propose a novel hybrid model, termed M-A-BiLSTM, which integrates Attention mechanisms, Multi-Layer Perceptron (MLP), and Bidirectional Long Short-Term Memory (Bi-LSTM). This model is designed to enhance feature selection capabilities and capture nonlinear patterns in financial time series. Evaluated on stock datasets from Apple, ExxonMobil, Tesla, and Snapchat, our model outperforms existing deep learning methods, achieving a 15.91% reduction in Mean Squared Error (MSE) for Tesla and a 5.95% increase in R-squared ( R 2 ) for Apple. Meanwhile, the MSE on the ExxonMobil dataset decreased to 1.8954, showing a significant reduction, while the R 2 increased to 0.9887. These results demonstrate the model's superior predictive power, offering a robust and interpretable approach for financial forecasting.