基于注意力的卷积神经网络金融时序数据预测  

Financial time series data prediction by attention-based convolutional neural network

作　　者：孙启森 张建新 程海阳 张强[1,3] 魏小鹏 SUN Qisen;ZHANG Jianxin;CHENG Haiyang;ZHANG Qiang;WEI Xiaopeng(Key laboratory of Advanced Design and Intelligent Computing,Ministry of Education(Dalian University),Dalian Liaoning 116622,China;School of Computer Science and Engineering,Dalian Minzu University,Dalian Liaoning 116600,China;School of Computer Science and Technology,Dalian University of Technology,Dalian Liaoning 116024,China)

机构地区：[1]先进设计与智能计算省部共建教育部重点实验室(大连大学),辽宁大连116622 [2]大连民族大学计算机科学与工程学院,辽宁大连116600 [3]大连理工大学计算机科学与技术学院,辽宁大连116024

出　　处：《计算机应用》2022年第S02期290-295,共6页journal of Computer Applications

基　　金：国家自然科学基金辽宁省联合基金资助项目(U1908214);国家自然科学基金资助项目(61972062)。

摘　　要：针对现有金融时序数据预测方法在构造金融特征图像的过程中因忽视市场环境变化导致的数据密度分布差异问题,提出一种基于滑动窗口标准化的金融数据预处理方法。所提方法将滑动窗口截取的数据使用独立的标准化转换为金融特征图像,使得依赖价格特征进行训练的卷积神经网络(CNN)模型能够学习到正确的映射关系;同时,针对金融特征图像的特征表达问题,为更好地捕捉其动态变化特征,将注意力机制引入CNN中,进而构建出一种注意力CNN金融时序数据预测模型。对标普500指数未来1天涨跌进行预测的准确率和F1分数分别为61%和0.7397,模拟交易实验投资回报率为23.04%,优于买入并持有策略。此外,消融实验结果也证明了预处理方法、注意力模块引入的有效性。To address the problem that existing financial time series data prediciton methods ignore the differences in data density distribution caused by market environment changes in the process of constructing financial feature images,a financial data preprocessing method based on sliding window standardization was proposed.The sliding window intercepted data was converted into financial feature images by independent standardization,so that the Convolutional Neural Network(CNN)model trained on price features can learn the correct mapping relationship.Meanwhile,for the feature representation problem of financial feature images,the attention mechanism was introduced into CNN to better capture the dynamic changing features,and an attentional CNN model for financial time series data prediction was constructed.The accuracy and F1-score of predicting the rise and fall of the S&P 500 index in the next day were 61%and 0.7397,respectively,and the return of investment in the simulated trading experiment was 23.04%,indicating that the proposed method is better than the buy-and-hold strategy.In addition,the ablation experiment results demonstrate the effectiveness of the preprocessing method and the attention module introduction.

关 键 词：卷积神经网络 核密度估计 注意力机制 时间序列 股票指数预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]