主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2026, Vol. 34 ›› Issue (10): 28-38.doi: 10.16381/j.cnki.issn1003-207x.2022.1040

• • 上一篇    下一篇

基于多特征融合深度神经网络的玉米期货价格预测

张大斌1,2,3, 曾芷媚1, 凌立文1,2(), 林锐斌1   

  1. 1.华南农业大学数学与信息学院,广东 广州 510642
    2.华南农业大学乡村振兴研究院,广东 广州 510642
    3.广东白云学院大数据与计算机学院,广东 广州 510440
  • 收稿日期:2022-05-12 修回日期:2022-12-15 出版日期:2026-10-25 发布日期:2026-10-09
  • 通讯作者: 凌立文 E-mail:linglw@scau.edu.cn
  • 基金资助:
    国家自然科学基金项目(71971089);国家自然科学基金项目(72001083);国家自然科学基金项目(72471097);广东省自然科学基金项目(2022A1515011612);广东省自然科学基金项目(2024A1515010941)

Prediction of Corn Futures Price Based on Multi-feature Deep Neural Network Model

Dabin Zhang1,2,3, Zhimei Zeng1, Liwen Ling1,2(), Ruibin Lin1   

  1. 1.School of Mathematics and Information,South China Agricultural University,Guangzhou 510642,China
    2.Rural Development Institute,South China Agricultural University,Guangzhou 510642,China
    3.School of Big Data and Computer Science,Guangdong Baiyun University,Guangzhou 510440,China
  • Received:2022-05-12 Revised:2022-12-15 Online:2026-10-25 Published:2026-10-09
  • Contact: Liwen Ling E-mail:linglw@scau.edu.cn

摘要:

在我国金融体系中,农产品期货市场具有指导市场进行自我调节以及为监管者提供高效信息的重要作用。期货价格的有效预测能够协助指导农业生产,监控价格大幅波动产生的经营风险,同时增强国家实施宏观调控政策的预见性和针对性。本文对粮食期货的主要品种——玉米期货进行研究,综合考虑其价格序列的非平稳非线性特征及复杂的市场和非市场影响因素,以玉米期货历史交易数据和相关新闻文本作为源数据,提出一种多特征融合的深度神经网络模型。利用该模型使用双向长短期记忆神经网络(BiLSTM)和文本卷积神经网络(textCNN)分别提取价格特征和文本特征,再融合SnowNLP提取的新闻情感特征,对提前一步的玉米期货收盘价进行预测,通过设置横纵双向实验,对所提出模型的预测有效性进行验证。实证结果表明,在不同的特征组合方案中,同时融合三种特征对模型预测准确性的提升最为明显,且相比于四种基线模型,本文模型在MAE、RMSE和R2上均表现出显著的性能优势。

关键词: 玉米期货价格, 新闻文本, 多特征融合, 深度神经网络模型

Abstract:

In China's financial and economic system, the futures market of agricultural products plays an important role in guiding the market to self-regulate and providing efficient information transmission for regulators. Effective prediction of futures prices can help guide agricultural production, monitor the operating risks arising from large price fluctuations, and enhance the predictability and pertinence of the national macro-control policies. In this paper, corn futures are taken as the research object,which is the main variety of grain futures. Considering the non-stationary and non-linear characteristics of its price series and the complex market and non-market influencing factors, a deep neural network model with multi-feature fusion is proposed based on the historical trading data of corn futures and relevant news texts. The model uses Bi-directional Long-term and Short-Term Memory neural network (BiLSTM) and text Convolution Neural Network (textCNN) to extract price features and text features respectively, and then fuses the news emotion feature extracted by SnowNLP to predict the closing price of corn futures one step ahead. The effectiveness of the proposed model is verified by setting horizontal and vertical experiments. The empirical results show that the fusion of the three features improves the prediction accuracy of the models most obviously in different feature combination schemes. And compared with the four baseline models, the proposed model shows significant performance advantages on MAE, RMSE and R2.

Key words: corn futures price, news text, multi-feature fusion, deep neural network model

中图分类号: