基于LSTM的股票预测研究

【作者】 林升；

【导师】 綦科；

【作者基本信息】 广州大学 ， 教育技术学， 2019， 硕士

【摘要】 股市是国家经济的重要组成部分,随着近年来人民生活水平的提高,进行股票投资的人不断增多。股票预测是每个投资者都在进行尝试的一项研究,普通投资者通过技术分析手段确认选股进行交易,技术分析师通过基本面、技术面和消息面多方结合进行推荐,而科研人员则是通过建立数学模型来对股票数据进行分析。随着深度学习的爆发以及循环神经网络在时间序列中取得良好的表现,LSTM作为循环神经网络中的经典模型受到了广泛的关注,具有广阔的应用前景。股票数据表现为经典的金融时间序列,利用神经网络对股票数据进行预测是近年来的研究热点。随着算法交易、量化投资等理念的兴起,越来越多的人开始利用神经网络对股票数据进行预测。但神经网络隐藏层的构建至今没有较好的指导理论,众多研究人员都是依靠自己摸索,或者在前人的经验上获取模型的结构以及参数设置。本文基于“历史总会重演”的观点,对同行业中的股票间常出现“同涨同跌”现象进行研究,通过结合Pearson相关系数和动态时间规整两种算法来对股票相关性特征的提取进行设计。在线性关系中利用Pearson对股票中存在的长、短周期进行获取,而在非线性关系中则利用动态时间规整,并将获取到的信息转化为相关性特征。在此基础上,设计了结合相关性特征的LSTM股票预测方法,利用Dense、PReLU、Dropout等多种神经网络算法结构构造出多种不同的LSTM模型,并探讨了不同模型结构及参数设置的股票预测效果。实验结果表明,本文提出的分类预测方法比传统的SVM、BP模型在正确率有3%以上的改进;而回归预测方法则比传统的LinearRegression、BP模型在RMSE、R~2、误差值以及自设计的盈利值等多个指标上均有更好表现。更多还原

【Abstract】 The stock market is an important part of the national economy.With the improvement of people’s living standards in recent years,more and more people are investing in stocks.Stock forecasting is a study that every investor is trying.The ordinary investors confirm the stock selection through technical analysis,and the technical analysts recommend through the combination of fundamentals,technical aspects and news,while the stock data is analyzed by researchers through establishing a mathematical model.With the outbreak of deep learning and the Recurrent Neural Network has achieved good performance in time series.LSTM has to attract broad attention as a classical model in the Recurrent Neural Network and has broad application prospects.Stock data is expressed as a classic financial time series.Predict stock data with using neural network is a research hotspot in recent years.With the rise of ideas such as algorithmic trading and quantitative investment,more and more people are using neural networks to predict stock data.However,there is no good guiding theory for deep learning so far.Many researchers rely on their own groping,or obtain the parameter settings of the model from the experience of their predecessors.Based on the viewpoint of "History repeats itself",this thesis studies the phenomenon of "same rise and fall" between stocks in the same industry,and extracts stock correlation characteristics by combining Pearson correlation coefficient and Dynamic Time Warping.In the linear relationship,Pearson is used to acquire the long and short periods existing in the stock,while in the nonlinear relationship,the dynamic time warping is utilized,and the obtained information is transformed into the correlation feature.On this basis,the LSTM stock prediction method combining correlation characteristics is designed.We used multiple neural network algorithm structures such as Dense,PReLU and Dropout to construce many different models,and discusses the influence of different model structure and parameter settings on stock forecasting.The experimental results show that the classification forecast method proposed in this thesis has more than 3% improvement rate than the traditional SVM and BP models.The regression forecast method is better than the traditional LinearRegressionand BP models in terms of RMSE,R2,error value and self-designed profit value.更多还原