计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (9): 319-328.DOI: 10.3778/j.issn.1002-8331.2205-0270

• 工程与应用 • 上一篇    下一篇

基于指数成分股关联的图卷积指数走势预测

王昌海,梁辉,王博,崔晓旭   

  1. 1.郑州轻工业大学 软件学院,郑州 450000
    2.中国政法大学 发展规划与学科建设处,北京 102249
  • 出版日期:2023-05-01 发布日期:2023-05-01

Graph Convolutional Index Trend Prediction Based on Correlation of Index Constituent Stocks

WANG Changhai, LIANG Hui, WANG Bo, CUI Xiaoxu   

  1. 1.College of Software Engineering, Zhengzhou University of Light Industry, Zhengzhou 450000, China
    2.Development Planning and Discipline Construction Office, China University of Political Science and Law, Beijing 102249, China
  • Online:2023-05-01 Published:2023-05-01

摘要: 利用历史交易数据预测股市指数未来走势是金融领域的重要问题,使用图卷积网络融合指数间走势关联性是该领域的前沿热点。针对当前图卷积指数预测中历史与未来动态图不一致的问题,提出一种基于指数成分股构建图结构的图卷积指数走势预测方法G-Conv。该方法提取传统量化特征和一维卷积网络的深度特征作为预测样本的特征。使用指数的成分股数据构建指数图结构,并对不同指数样本特征做图卷积以得到指数预测结果。使用A股中42个常用指数验证该方法的有效性。实验使用MAE和MSE作为模型训练的损失函数,选取GC-CNN、AD-GAT等经典方法作为比较基准,结果表明在两种误差评价标准下,G-Conv分别降低平均预测误差5.10%和4.20%,且表现出较好的泛化性能。

关键词: 金融数据分析, 股市指数预测, 数据归一化, 一维卷积神经网络, 图卷积神经网络

Abstract: Predicting future trends of stock market indexes with historical transaction data is an important issue in the field of financial investment. Fusing the trends correlations between indices with the graph convolutional network is a research hotspot. Aiming at the inconsistency of historical and future dynamic graph structures in the current graph convolutional index prediction, a graph convolutional index trend prediction method termed G-Conv is proposed, in which the constituent stocks are applied to construct the index graph. The traditional quantitative features and deep features of one-dimensional convolutional networks are extracted as the features of predict samples. Then, the graph between indices is constructed using the constituent data of the index, and graph convolution is performed on features of different index samples to obtain the index prediction result. Finally, 42 commonly used indices in the A-shares market are used to evaluate the performance of this method. In the experiment, MAE and MSE are used as loss functions of model training, and classical methods such as GC-CNN, AD-GAT are selected as comparison benchmarks. The results show that G-Conv reduces the average prediction error by 5.10% and 4.20% under these two error evaluation criteria respectively, and shows better generalization performance.

Key words: financial data analysis, stock market index prediction, data normalization, one-dimensional convolutional neural network, graph convolutional neural network