AE-ACG: A novel deep learning-based method for stock price movement prediction.

• A CNN-GRU block is first designed and embedded into AE as fundamental structure to extract advanced, abstract, significant temporal features. • We introduced attention mechanism (AM) to highlight the meaningful features while reducing the weight of irrelevant information during the decoding process. • Different from existing methods, we design an encoding-decoding structure based on attention mechanism and skip connection. The skip link connects the decoder and encoder directly, allowing hierarchical features to be used more directly in the decoding process and improving the efficiency of information utilization. This paper proposes a method named AE-ACG for stock price movement prediction. In AE-ACG, the convolutional neural network (CNN) and gated recurrent unit (GRU) are combined to design a base layer, which is embedded in the autoencoder (AE) framework, to efficiently extract features from financial time series data. Furthermore, skip connection links encoding and decoding to leverage hierarchical features. Attention mechanism (AM) also distinguishes the importance of historical data across periods. Extensive experiments demonstrated that the proposed model is effective in predicting price movements, showing advantages over some mainstream methods. [ABSTRACT FROM AUTHOR]