Enhancing GRU With Long-Short Attention: A Case Study of Saudi Stock Market Prediction

Predicting stock market movements accurately is a challenging task due to the inherent complexity and volatility of financial markets. In recent years, deep learning techniques, particularly the Gated Recurrent Unit (GRU), have shown promising results in financial time series prediction. However, capturing the most relevant information from large volumes of data remains a critical challenge for improving prediction accuracy. This work introduces an attention-enhanced GRU model for forecasting the Saudi stock market. The proposed model integrates attention mechanisms to dynamically prioritize different input patterns, enabling the network to focus on the most relevant information for prediction. Two types of attention are incorporated: short attention, which focuses on recent time steps to capture fine-grained details and immediate context, and long attention, which captures long-term dependencies across the entire sequence, helping to identify broader trends and patterns. The proposed model was empirically evaluated using historical data from the Saudi stock market and compared with traditional GRU, LSTM, RNN, and Transformer models. The results show that the enhanced model outperforms the baseline models in terms of prediction accuracy. In addition, a comparison with results reported on other public datasets was conducted, demonstrating improved performance in terms of RMSE and R2 metrics.