基于多级注意力机制的大豆期货预测  

Forecasting for Soybean Futures Based on Multistage Attention Mechanism

作　　者：张杰[1] 甄柳琳 翟东升[1] ZHANG Jie;ZHEN Liu-lin;ZHAI Dong-sheng(School of Economics and Management,Beijing University of Technology,Beijing 1o0124,China)

机构地区：[1]北京工业大学经济与管理学院,北京100124

出　　处：《系统工程》2023年第2期148-158,共11页Systems Engineering

基　　金：北京市自然科学基金资助项目(9182001)。

摘　　要：考虑影响大豆期货价格的外部因素,针对国内外期货市场间波动的动态相关性特征,提出一种基于多级注意力机制的MAN预测模型(Multistage Attention Network)。模型利用LSTM编码器-解码器结构,整合局部注意力机制与全局注意力机制建模编码器,使用时间注意力机制与外部影响因素融合模块建模解码器,据此建立深度学习框架。以中国、美国、巴西、日本大豆期货为研究对象,提取国内外大豆期货历史交易信息及外部影响因素的潜在特征,用于对我国大豆期货收盘价的预测。通过实证研究验证了MAN模型的性能及预测的有效性。In view of the dynamic correlation characteristics between domestic and foreign futures markets,and considering the external factors that affect the price of soybean futures,this paper proposes the MAN model based on multistage attention mechanism.The model uses LSTM encoder-decoder structure,integrates local attention mechanism and global attention mechanism to model encoder,and uses temporal attention mechanism and external influencing factors fusion module to model decoder.Based on this,a deep learning framework is established.Taking soybean futures of China,the United States,Brazil and Japan as the research objects,this paper extracts the historical trading information of domestic and foreign soybean futures and the potential characteristics of external influencing factors,which can be used to forecast the closing price of Chinese soybean futures.The performance of the MAN model and the effectiveness of prediction are verified by empirical research.

关 键 词：大豆期货 预测模型 多级注意力机制 Encoder-Decoder结构 

分 类 号：F762[经济管理—产业经济]