摘要
针对期货收盘价预测问题中样本量偏少(对于深度学习模型来说)的问题,本文设计了一个多输入多输出的深度学习模型。该模型通过多个预测目标训练共享网络,提高了特征提取的效率。文中以PTA期货为算例,证明了该模型的优越性:(1)对于收盘价预测,该模型的预测误差显著小于CNN-LSTM混合模型;(2)对于收盘价涨跌预测,本文根据预测结果模拟投资,获得了正的收益。另外,本文在模型训练之前,使用协整检验分析了现货价格和现货库存与期货价格的相关性,从而保证了数据质量,也弥补了深度学习模型可解释性较弱的缺点。
To solve the problem that the amount of futures daily prices is small(for deep learning models),a deep learning model with multiple Inputs and Outputs is designed in this paper.The model can improve the efficiency of feature extraction by training the shared network with multiple prediction targets.Taking PTA futures as an example,this paper proves the superiority of the model:(1)for the closing price prediction,the prediction error of the model is significantly smaller than that of the CNN-LSTM hybrid model;(2)For the closing price change value prediction,this paper simulates the investment according to the forecast results,and obtains positive returns.In addition,this paper enriched the input features of the training set.In addition,before model training,this paper uses co-integration test to analyze the correlation between input variables and futures price,so as to ensure data quality and make up for the weak interpretability of deep learning model.
作者
林杰
李英超
LIN Jie;LI Yingchao(School of Economic and Management,Tongji University,Shanghai 200092,China)
出处
《上海管理科学》
2022年第1期30-37,共8页
Shanghai Management Science
作者简介
林杰(1967—),男,四川渠县人,博士,教授,同济大学经济与管理学院博士生导师,主要研究方向:管理信息系统、数据挖掘。E-mail:jielinfd@163.com。
