基于生成式对抗网络的股价预测研究  

Research on stock price forecasting based on FWGAN model

作　　者：许飞飞 胡月[1] 汪召兵 XU Feifei;HU Yue;WANG Zhaobing(School of Sciences,Zhejiang University of Science and Technology,Hangzhou 310023,Zhejiang,China)

机构地区：[1]浙江科技学院理学院,杭州310023

出　　处：《浙江科技学院学报》2022年第3期207-215,共9页Journal of Zhejiang University of Science and Technology

基　　金：浙江省科技计划项目(2015C33088)。

摘　　要：为了降低股票市场中噪声信息和投资者情绪对股票价格的影响,以便给投资者带来较高的投资回报并降低交易风险,特提出一种基于金融双向编码器表征和瓦瑟斯坦距离的生成式对抗网络(financial bidirectional encoder representation from transformers and Wasserstein generative adversarial networks,FWGAN)股价预测模型。本模型首先采集东方财富网股评数据,并利用自然语言处理预训练模型将股评数据量化为情绪值,然后将情绪值连同历史股票交易数据、技术指标数据输入由长短期记忆网络(long-short-term memory,LSTM)为生成器和卷积神经网络(convolutional neural network,CNN)为判别器组成的FWGAN模型中进行训练。对比LSTM模型、门控神经网络(gated recurrent units,GRU)模型和生成式对抗网络(generative adversarial networks,GAN)模型对山西汾酒股价的预测性能,结果表明,FWGAN模型的均方根误差为2.572,达到最低,预测效果最好。试验结果验证了本模型对股票时间序列预测的有效性和优越性,可以为投资者进行股价预测提供参考。In order to reduce the impact of noise information and investor sentiment on stock prices,bring investors higher investment returns and lower transaction risks,a generative adversarial networks stock price forecasting model was proposed on the basis of financial bidirectional encoder representation from transformers and Wasserstein distance(FWGAN).This model firstly collected the stock evaluation data of eastmoney.com,and then quantified those data into sentiment values by virtue of the natural language processing pre-training model.Finally,the sentiment values,together with the historical stock transaction data and technical index data,were input into the long-short-term memory network(LSTM)to conduct training in the FWGAN model consisting of a generator and a convolutional neural network(CNN)as a discriminator.Comparing the forecasting performance of the LSTM model,the gated recurrent units(GRU)model and the generative adversarial networks(GAN)model on Shanxi Fenjiu stock price,the results show that the root mean square error of FWGAN model is 2.572,which is the lowest,achieving the best forecasting effect.Experimental results have verified the effectiveness and superiority of this model for stock time series forecasting,which can provide reference for investors to forecast stock prices.

关 键 词：股价预测 股民情绪 生成式对抗网络股价预测模型 时间序列 

分 类 号：F830.91[经济管理—金融学]