融合经验模态分解与线性Transformer的高频金融时间序列预测  

Fusion of empirical modal decomposition and linear transformer for high⁃frequency financial time series forecasting

作　　者：文馨贤 WEN Xinxian(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500

出　　处：《现代电子技术》2022年第23期121-126,共6页Modern Electronics Technique

摘　　要：随着深度学习的发展,神经网络模型已被广泛应用于期货等金融资产价格序列预测研究工作中。当前的研究以低频数据为主,针对非线性、非平稳、高噪声的高频数据的预测准确率还有待提升。因此,提出CEEMDAN_Linformer模型,通过引入“分解⁃重构”方法,使用自适应噪声完全集成经验模态分解(CEEMDAN)方法对高频交易数据进行去噪预处理;通过引入时间戳进行特征融合,为输入编码提供了全局特征;使用线性Transformer提升模型的预测准确率,同时降低原始Transformer的复杂度,使其更适用于在当前的期货高频数据预测任务上。以贵金属期货品种——黄金期货、白银期货的5 min、1 h价格序列为例,实现了多步价格序列预测。实验对比了LSTM、CONVLSTM、TCN、Transformer四个基准模型,提出的模型在三个评价指标上均优于以上基准模型,取得了较好的预测效果。With the development of deep learning,neural network models have been used widely in the research on price series forecasting of financial assets like futures.However,the current research is focused on low⁃frequency data,and the prediction accuracy rate of high⁃frequency data with non⁃linearity,non⁃smoothness and high⁃noise is yet to be improved.Therefore,a CEEMDAN_Linformer model is proposed,which employs the method of complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)to denoise high⁃frequency trading data by introducing the″decomposition⁃reconstruction″approach.The timestamp is introduced for feature fusion,so as to provide global features for the input encoding.The linear transformer is used to improve the prediction accuracy of the model while reducing the complexity of the original transformer,making it more suitable for the current task of predicting high⁃frequency futures data.In this paper,two species of precious metal futures(gold futures and silver futures)are taken as examples and multi⁃step forecasts are implemented for their 5⁃minute and 1⁃hour price series respectively.The experiments were conducted to contrast the four benchmark models of LSTM,CONVLSTM,TCN and transformer,and the proposed model outperforms the above benchmark models in all three evaluation metrics and achieved better prediction results.

关 键 词：金融时间序列 经验模态分解 神经网络 深度学习 线性Transformer 高频数据 价格预测 期货 

分 类 号：TN711-34[电子电信—电路与系统] TP3[自动化与计算机技术—计算机科学与技术]