基于多源多任务自动编码器的原油期货价格预测  

Crude oil futures price prediction based on multi-source and multi-task autoencoder

作　　者：蒋锋[1] 胡成雨 王辉[2,3] JIANG Feng;HU Chengyu;WANG Hui(School of Statistics and Mathematics,Zhongnan University of Economics and Law,Wuhan 430073,China;National Science Library(Wuhan),Chinese Academy of Sciences,Wuhan 430071,China;Hubei Key Laboratory of Big Data in Science and Technology,Chinese Academy of Sciences,Wuhan 430071,China)

机构地区：[1]中南财经政法大学统计与数学学院,武汉430073 [2]中国科学院武汉文献情报中心,武汉430071 [3]中国科学院科技大数据湖北省重点实验室,武汉430071

出　　处：《系统工程理论与实践》2025年第2期702-716,共15页Systems Engineering-Theory & Practice

基　　金：教育部人文社会科学研究项目(22YJAZH038);国家自然科学基金(61773401);科技大数据湖北省重点实验室项目(E3KF291001);国家社会科学基金重大项目(20&ZD132)。

摘　　要：新闻文本是反映国际金融市场价格波动的重要信息.为了量化原油期货价格预测中的不确定性,本文提出了一种新的基于新闻文本和结构化指标的多源多任务自动编码器(multi-task autoencoder,MTAE)方法,并用于原油期货价格的预测.首先应用Word2Vec方法提取新闻文本中的潜在特征;针对新闻文本词向量的高维性问题,引入MTAE方法对词向量进行降维和去噪;其次,利用MTAE网络拓扑结构,对新闻文本词向量和原油期货每日涨跌信息进行融合,以增强文本特征的可预测能力.最后,使用长短期记忆神经网络将文本特征与经济发展、能源、气候环境等原油相关指标进行集成,预测原油期货价格.结果表明,本文提出的多源多任务自动编码器能够很好地提取新闻文本中的非线性特征,具有较好的水平精度和鲁棒性.News texts can reflect important information in international financial markets.In order to quantify the uncertainty in crude oil futures price forecasting,a novel multi source and multi-task autoencoder(MTAE)method with news text and structured indicators is proposed to predict crude oil future price.First,the underlying features are extracted from news text using Word2Vec.For the high dimension of news text vectors,MTAE is introduced for dimension reduction and denoising.Then,to enhance the predictability of text features,news texts and crude oil daily news with movement information are fused by using network topology of MTAE.Moreover,the long short-term memory neural network(LSTM)is employed to integrate text features with economy,energy and climate environment indicators for crude oil futures price prediction.Experimental results show that the proposed MTAE can extract nonlinear features well and have good horizontal accuracy and robustness.

关 键 词：原油期货价格 新闻文本 多任务自动编码器 特征融合 长短期记忆神经网络 

分 类 号：F764.1[经济管理—产业经济] F713.35F224