基于多视图注意力机制的多维度价格预测模型研究

传统的股票价格预测模型只针对单一维度价格进行预测，忽略了多维度价格之间的复杂关系。因此，为了更好地对股票价格进行准确预测和为决策者提供前瞻性信息，提出了一种新的基于多视图注意力机制的多维度价格预测模型。通过多视图的深度可分离卷积网络学习多维度股票价格潜在的复杂的输入—输出关系，更好地提取股票价格的时空特征，实现时空数据的智能关联，并使用注意力机制进一步提升模型的预测性能，进而通过时空多维度的股价历史数据来预测单和多时间步长股票价格。该模型与其他四种模型在中国银行股价数据集上进行实验和比较，发现所提模型在不同预测时长下相比于表现最好的模型，平均绝对误差分别降低了0.4%、0.5%、4.2%、3.9%，均方误差分别降低了0.8%、2%、1.9%、1.9%，平均百分比误差分别降低了0.15%、0.21%、1.24%和1.34%。因此所提模型预测精度最高，预测性能最好，并且在对其他维度的股票价格预测上具有普适性。

Research on multidimensional price prediction model based on multi-view attention mechanism

Traditional stock price forecasting models only predict prices in a single dimension, ignoring the complex relationship between prices in multiple dimensions. Therefore, to better predict stock prices accurately and provide forward-looking information for decision-makers, this paper proposed a new multidimensional price prediction model based on multi-view attention mechanism. It learned the underlying complex input-output relationship of multidimensional stock prices through multi-view depthwise separable convolution networks, better extracted the spatial-temporal features of stock prices, realized the intelligent association of spatial-temporal data, and used the attention mechanism to further improve performance, and then predicted stock prices of single and multiple time steps through spatial-temporal and multidimensional stock price historical data. The model was tested and compared with other four models on the bank of China stock price datasets. Comparing with the best performing model under different forecasting time periods, it show that the mean absolute error of the model is reduced by 0.4%, 0.5%, 4.2%, and 3.9%, the mean squared error is reduced by 0.8%, 2%, 1.9%, and 1.9%, and the mean absolute percentage error is reduced by 0.15%, 0.21%, 1.24%, and 1.34%, respectively. Therefore, the proposed model has the highest prediction accuracy and the best prediction performance, and has universality in predicting stock prices in other dimensions.