基于多尺度1D-CNN和注意力机制的汇率多步预测研究  

Multi-step forecasting of exchange rate based on multi-scale 1D-CNN and attention mechanisms

作　　者：王刚[1,2] 陈红[1] 马敬玲 王珏 WANG Gang;CHEN Hong;MA Jingling;WANG Jue(School of Management,Hefei University of Technology,Hefei 230009,China;The Ministry of Education Key Laboratory of Process Optimization and Intelligent Decision,Hefei 230009,China;Academy of Mathematics and Systems Science,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]合肥工业大学管理学院,合肥230009 [2]过程优化与智能决策教育部重点实验室,合肥230009 [3]中国科学院数学与系统科学研究院,北京100190

出　　处：《系统工程理论与实践》2024年第6期1934-1949,共16页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(72071062,72371096);安徽省杰出青年基金(2208085J12)。

摘　　要：深度学习在处理时间序列数据上具有优势,在汇率时间序列的应用研究中,目前深度学习主要关注于单步预测,即利用以前时点的数据预测下一个时点的汇率数据.然而,在实际应用中,这种单步预测方式往往无法为决策者提供足够的决策信息;同时,由于汇率时间序列呈现出非平稳、复杂度高等特点,直接利用传统深度学习方法进行预测无法充分挖掘汇率序列的特征及规律.为此,本研究提出一种基于多尺度一维卷积神经网络(1D-CNN)和注意力机制的汇率多步预测方法,该方法在自适应的融合多尺度特征的同时,差异化的融合汇率不同时刻的时间序列特征,实现汇率的多步预测.通过实验发现,所提方法相较于基准方法,如差分整合移动平均自回归模型、支持向量回归、随机游走、极限梯度提升算法、长短期记忆网络等具有更高的预测精度,表明该方法能够为外汇市场投资者提供决策支持.Deep learning has advantages in processing time series data.Currently,in the applied research of exchange rate time series,deep learning mainly focuses on single-step prediction,which only uses data from previous time points to predict exchange rate data for the next time point.However,this one-step prediction method often fails to provide sufficient decision-making information for decision makers.At the same time,due to the characteristics of non-stationary and high complexity of exchange rate time series,using traditional deep learning methods for forecasting cannot fully explore the characteristics of exchange rate series.Therefore,this study proposes a multi-step prediction method of exchange rate based on multi-scale one-dimensional convolution neural network(1D-CNN)and attention mechanisms,which adaptively integrates multi-scale features and differentiated time series features of exchange rate at different moments to achieve multi-step prediction of exchange rate.Experiments show that the proposed method has higher prediction accuracy than the benchmark methods such as autoregressive integrated moving average model(ARIMA),support vector regression(SVR),Random Walk(RW),eXtreme gradient boosting(XGBoost),long short term memory(LSTM),etc.,which indicates that the proposed method can provide decision support for foreign exchange market investors.

关 键 词：汇率多步预测 深度学习 多尺度1D-CNN 注意力机制 

分 类 号：F062.4[经济管理—政治经济学]