多源异构数据图像整合预测方法研究——以黄金价格预测为例  

Convolutional Neural Network Applied to Gold Price Forecasting with an Image Integration Methods Based on Multi-Sources and Heterogeneous Data

作　　者：高丽珺 蒋雨芯 盛培根 魏先华 GAO Lijun;JIANG Yuxin;SHENG Peigen;WEI Xianhua(School of Economics and Management,University of Chinese Academy of Sciences,Beijing 100190)

机构地区：[1]中国科学院大学经济与管理学院,北京100190

出　　处：《系统科学与数学》2022年第11期3073-3093,共21页Journal of Systems Science and Mathematical Sciences

基　　金：国家自然科学基金项目(71932008);中国科学院大学数字经济监测预测预警与政策仿真教育部哲学社会科学实验室(培育)基金资助课题。

摘　　要：黄金是一种拥有商品、货币、投资等多重属性的特殊大宗商品,而且黄金市场的交易范围和参与者涉及全球,所以其价格会受到多种因素的影响.同时,由于这些影响因素来源分散,数据异构性强并且有各自的规律可循,所以黄金价格本身具有复杂性、非平稳性以及非线性等特征,这也直接导致了预测黄金价格的变动趋势非常具有挑战性.基于黄金的这一复杂特性,文章提出一种全新的基于图像挖掘技术的多源数据融合预测方法—MODⅡ(Multiple Original Data Images Integration).通过2种数据可视化方法将黄金价格的不同影响要素数据转化为4种类型的图像,将这些来自不同市场的不同影响因素通过图像整合在一起,并利用卷积神经网络(Convolutional Neural Network,CNN)的多通道输入结构将多种类型的图像进行融合,之后由CNN自动学习各图像特征及所占权重并进行预测,样本外预测准确率可达81.06%,一定程度上提高了预测准确率.Gold is a special commodity with multiple attributes,such as commodity,currency and investment.With wide range of trade and participants across the globe,gold price may be affected by various factors.Furthermore,scattered sources of these influencing factors,coupled with strong heterogeneity of the data structure with respective regularity makes gold price complex,non-stationary and non-linear,which directly leads to the difficulties of predicting the trend of the gold price.By leveraging the complex characteristics of the gold price,this paper proposed a new multi-source data fusion prediction method based on image mining technology-MODII(Multiple Original Data Images Integration) to accurately predict gold price.Specifically,we collected various of influencing factors of the gold price across markets and transformed them into four types of images.Then,we built a convolutional neural network(CNN) by taking these cross-batch images as input to learn the features and predict corresponding gold price.Our method could reach 81.06% accuracy for out-of-sample prediction and outperform the state-of-the-art methods.

关 键 词：卷积神经网络 图像挖掘 趋势预测 多源数据 黄金价格 

分 类 号：F831.54[经济管理—金融学]