基于DPSO-LSTM超参数调优的股市价格预测  

Stock market price forecasting based on DPSO-LSTM hyperparameter tuning

作　　者：张成军 李琪[1,2,3] 王梅 乔译 陈亚当 余文斌 ZHANG Cheng-jun;LI Qi;WANG Mei;QIAO Yi;CHEN Ya-dang;YU Wen-bin(Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology(CICAEET),Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Computer Science,Nanjing University of Information Science and Technology,Nanjing 210044,China;Engineering Research Center of Digital Forensics,Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学,江苏大气环境和装备技术协同创新中心,南京210044 [2]南京信息工程大学计算机学院,南京210044 [3]南京信息工程大学数字取证教育部工程研究中心,南京210044

出　　处：《信息技术》2024年第5期1-7,共7页Information Technology

基　　金：国家自然科学基金(61501247,61703212,61802197)。

摘　　要：长短期记忆网络(Long Short-Term Memory,LSTM)适合处理和预测时间序列中间隔和延迟较长的重要事件。由于其复杂的网络结构、不确定的超参数和耗时的网络训练,使得人工寻找高效的网络配置成为一项具有挑战性的工作。文中采用分布式粒子群算法(Distributed Particle Swarm Optimization,DPSO)来有效解决LSTM的超参数调优问题,研究LSTM中最优的隐藏元个数、激活函数以及学习率等超参数的选择,寻找高性能的LSTM。基于沪深300历史交易数据进行价格预测,实验结果表明该方法是有效的,这为超参数调优与股市价格预测提供了新的思路和方法。Long Short-Term Memory(LSTM)is suitable for processing and predicting important events with very long intervals and delays in time series.It is a challenging task to manually find an efficient network configuration due to its complex network structure,uncertain hyperparameters,and time-consuming network training.In this paper,Distributed Particle Swarm Optimization(DPSO)is used to effectively solve the hyperparameter tuning problem of LSTM,and research the selection of the optimal number of hidden elements,activation function and learning rate in LSTM to find a high-performance LSTM.Based on the historical transaction data of CSI300 for price prediction,the experiment results show that the method is effective,which provides new ideas and methods for hyperparameter tuning and stock market price prediction.

关 键 词：长短期记忆网络 人工神经网络 分布式粒子群优化算法 超参数调优 股市预测 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]