联合多尺度分层注意力的股价走势预测模型OA

Traditional stock prediction models fail to effectively capture multi-scale temporal features in financial time series data.This paper proposes a stock price trend prediction model that integrates multi-scale hierarchical attention mechanisms.The model employs an expert knowledge-driven feature selection method based on modern financial theory to identify key indicator features.Multi-scale convolutional neural networks(CNN)extract local pattern features,while bidirectional long short-term memory networks(BiLSTM)capture temporal dependencies.The model innovatively fuses hierarchical attention and cross-temporal attention mechanisms to achieve adaptive integration of information across different time windows.The experiment uses a sliding window dynamic alignment algorithm,takes the past 45 time points as the input sequence,predicts price fluctuations in the next three days,and uses five leading stocks in different industries to verify the effectiveness of the model.Experimental results demonstrate that the proposed model achieves optimal performance across overall accuracy,F1-score,and Matthews Correlation Coefficient compared to 11 baseline models.The model exhibits stable performance across different stocks,confirming its superior accuracy,stability,and adaptability in predicting price trends of leading stocks.