关键词: 股价趋势预测, 深度学习, 注意力机制, 双向长短期记忆网络

Abstract: Current conventional models, such as support vector machine (SVM), recurrent neural network (RNN), and long short-term memory (LSTM), have limitations in handling nonlinear, multi-scale, and highly noisy stock time-series data, often failing to effectively improve the accuracy of stock price trend prediction. To address this issue, an innovative deep learning prediction model named SDAE-CNN-BiLSTM-CM, based on channel attention mechanism and multi-head self-attention mechanism, is proposed. This model integrates a stacked denoising auto-encoder (SDAE) and a CNN-BiLSTM model to effectively model noisy stock data, and also introduces a channel attention mechanism (CAM) and a multi-head self-attention mechanism (MSA) to better capture the short-term and long-term dependencies of the time series. Finally, a joint optimization layer is utilized to achieve hierarchical aggregation of temporal information, adapting to the highly volatile nature of financial time series. Empirical results show that, compared with traditional models, the proposed model has significant advantages in improving prediction accuracy, and the trading strategy based on this model has also achieved higher returns and lower risks in backtesting performance.

Key words: stock index trend forecasting, deep learning, attention mechanism, bi-directional long short-term memory