摘要
在股价预测领域,预测的准确率比估计的相合性更有价值,因此保证相合估计的传统线性模型正逐渐被长短期记忆神经网络(long short-term memory,简称LSTM)等深度学习方法替代.然而,影响股价的因素是多源的,不仅包括股市历史交易信息,还包括企业基本面信息和宏观经济信息等,这些不同来源的信息间有长期确定关系,而关于此关系的数据记忆会被传统LSTM模型在学习过程中抛弃.构建“集成式长短期记忆神经网络模型”即ensemble LSTM,应用动态网络生成机制保证不同来源数据间的长期均衡关系不会被遗忘,且采用多个LSTM并联,让各神经网络独立处理单来源数据,再通过稠密层融合,因此该模型具有节约运算资源的能力.随机选取了16支个股,对比LSTM和ensemble LSTM在预测股价涨跌方面的性能,发现后者在节约运算资源上具有优势,且准确率也大多高于前者.
In the field of stock price prediction,the accuracy of prediction is of far more value than the consistent estimation.Therefore,the traditional linear model guaranteeing the consistency of estimation was gradually replaced by deep learning such as long short-term memory.However,the factors affecting stock price were diversified.Such multi-source information system contained not only historical trading information of stock market,but also fundamental information of each individual enterprise and macroeconomic information with a striking feature that there was long-term equilibrium relationship between them,nevertheless the data memory about this relationship would be abandoned by the traditional LSTM model in the learning process.In this paper,we built the ensemble long short-term memory model(ensemble LSTM),which applied the dynamic network generation mechanism to ensure that the long-term equilibrium relationship was never forgotten during process.The multiple LSTM was connected in parallel,each neural network processed single-source data independently,and then fused through dense layers,therefore the model could save the computing resources.By selecting 16 stocks randomly and comparing the performance of LSTM and ensemble LSTM in predicting stock prices fluctuation,it was found that the latter had advantages in saving computing resources and the accuracy was mostly higher than that of the former.
作者
赵丽君
王峻楠
程建华
ZHAO Lijun;WANG Junnan;CHENG Jianhua(School of Economics,Anhui University,Hefei 230601,China)
出处
《安徽大学学报（自然科学版）》
CAS
北大核心
2021年第4期17-26,共10页
Journal of Anhui University(Natural Science Edition)
基金
国家自然科学基金青年基金资助项目(71701001)
安徽省社科基金资助项目(AHSKF2019D019)。
关键词
多源数据
长短期记忆神经网络
长期均衡关系
股价涨跌预测
multi-source data
long short-term memory neural network
long-term equilibrium
stock price movement prediction
