CAGTRADE: Predicting Stock Market Price Movement with a CNN-Attention-GRU Model

Abstract

Accurately predicting market direction is crucial for informed trading decisions to buy or sell stocks. This study proposes a deep learning based hybrid approach combining convolutional neural network (CNN), attention mechanism (AM), and gated recurrent unit (GRU) to predict short-term market trends (1 day, 3 days, 7 days, 10 days) across different stock indices (BSE, HSI, IXIC, NIFTY, N225, SSE). The architecture dynamically weights the input sequence with the AM model, captures local patterns through CNN and effectively models long-term dependencies with GRU thus aiming to accurately classify either "buy" or "sell" positions of stocks. The model is assessed using classification and financial evaluation metrics involving accuracy, precision, recall, f1-score, annualized returns, maximum drawdown, and return on investment. It outperforms benchmark models, and different technical indicators including average directional index, rate of change, moving average convergence divergence, and the buy-and-hold strategy, demonstrating its effectiveness in various market conditions. The proposed model achieves an average accuracy of 98% in predicting the 1 day-ahead direction, and an average accuracy of 88.53% across all prediction intervals. The model was also validated using the wilcoxon signed rank test that further supported its significance over the benchmark models. The CAG model presents a comprehensive and intuitive approach to stock market trend prediction, with potential applications in real-world asset decision-making.