基于深度学习的股票价格趋势预测方法研究

发布时间：2018-01-04 06:33

本文关键词：基于深度学习的股票价格趋势预测方法研究　出处：《云南财经大学》2017年硕士论文　论文类型：学位论文

【摘要】：当今股票市场不仅为优秀挂牌企业提供融资,同时让一些有投资意识的股民提供资金出路。从而使得社会资源得到更好的配置和宏观经济得以调控,然而由于股市的不确定性,每个投资人对股市认知的异同性、技术分析的复杂性等因素,使得广大股民投资的回报率达不到预期,有的甚至血本无归。所以一直以来股票市场都被无论是政府、企业还是投资者所高度关注。股票价格趋势的预测更是股票研究中的热点。众所周知,由于股市的波动具有极强的非线性、高噪声等特点,所以对股票价格趋势预测极其困难,传统股票预测方法往往收效甚微。因此如何建立新的股票价格趋势预测的模型来提高预测的准确度,从而帮助金融投资者有效规避风险,投资获利最大化,具有重要的理论意义和应用价值。本文首先阐述了传统股票预测方法,大体分为:基本分析法主要是从宏观微观经济、相关公司的财务报表和现金流等信息角度,通过相对估值和折现估值等等,对该股票的内在价值进行估值。不足之处:信息不对等性及相关挂牌公司披露信息延时性、准确性等导致估值困难。大盘分析法主要是依据统计图表,如K线图,其形态可分为整理形态和趋向线等,根据对其特定的形态来判断股市的未来动向。不足之处:此类分析方法繁多,且各个投资人判断习惯不同,方法之间存在巨大差别。统计学分析法主要是采用最小二乘构建各种回归,例如混合回归模型、自回归模型等进行股票价格趋势预测,此类模型的预测预测准确率较前两类预测方法要高。不足之处:这些回归模型通常假设前提太多,且对非线性强的问题处理能力,而股票价格趋势的预测问题影响因素众多且非线性极强。基于人工神经网络的预测模型具有高度自组织、自调整和自学习的能力、是一个复杂度极高的非线性系统,模型预测结果通常也要优于上述传统方法。不足之处:基于神经网络的股票预测模型容易陷入局部最小值的问题,且多层神经网络在对复杂事物的描述时,往往要增多隐含层的层数,这样会导致梯度扩散的问题,从而影响准确率。本文正是从基于人工神经网络预测模型的缺点,如梯度扩散和局部最小值等问题,从而提出了采用受限的玻尔兹曼机模型构建基于深度学习的股票价格趋势预测模型。深度学习是基于神经网络基础上发展而来,不仅继承了神经网络方法的优点,而且很好的克服了神经网络方法的不足之处。本文预测模型采用受限的玻尔兹曼机来构建深度置信网络,学习方法是采用K步吉布斯采样后,结合对比散度算法,来训练整个深度置信网络。最后利用收集的格力空调的股票价格信息来训练本文预测模型并对本文模型预测准确率进行了检验。选用基于BP神经网络的股票价格预测模型作为本文预测模型的对比模型,并用采用实例贵州茅台和比亚迪的股票价格信息来检验两个模型的预测准确率,实验结果表明:基于深度学习的股票价格趋势预测模型效果良好,且准确率要优于BP神经网络预测模型。本文创新点:(1)本文采用了基于受限的玻尔兹曼机构建深度置信网络的股票价格趋势预测模型,学习方法采用了经过K步的吉布斯采样后的对比散度算法(CD算法)来训练预测模型。最后给出实例验证。(2)将本文预测模型与基于BP神经网络股票价格趋势预测模型的预测准确率进行了实例比较。
[Abstract]:The stock market not only provides financing for outstanding listed companies, while some investment minded investors to provide funds to make way. A better allocation of social resources and macroeconomic regulation to, however due to the uncertainty in the stock market, stock market investors on the similarities and differences of each cognitive complexity, factors such as technical analysis, making the stock investment the rate of return is not up to expectations, some even lose everything. So since the stock market has been highly concerned by both the government, enterprises and investors. The stock price trend forecast is a hot stock research. As everyone knows, because of the volatility of the stock market has a very strong nonlinear, high noise characteristics, so the stock price trend prediction is extremely difficult, the traditional stock forecasting methods often have little effect. So how to establish a new prediction model of stock price trend To improve the prediction accuracy, so as to help investors to avoid financial risk effectively, investment profit maximization, and has important theoretical significance and application value. This paper describes the traditional prediction methods of stock, can be divided into: basic analysis mainly from the macro and micro economy, financial statements and cash flow information related to the company's point of view. Through the relative valuation and valuation discount and so on, the intrinsic value of the stock valuation. Disadvantages: information asymmetry and listed company information disclosure delay, as a result of the valuation accuracy difficult. Large disk analysis method is mainly based on the statistical charts, such as the K map, its shape can be divided into the consolidation pattern and the trend line. According to the future trends of its specific form to determine the stock market. The shortcomings of such analysis methods are various, and each investor to judge different habits, there is a huge difference between the methods. The statistical analysis method is mainly constructed by least squares regression, such as mixed regression model, auto regression model to forecast the stock price trend forecast of this kind of model accuracy than prediction method to high. The first two shortcomings: the regression model is usually premise too much, and the ability to deal with the nonlinear problems. The factors affecting the prediction of stock price trend of large and highly nonlinear. The prediction model based on artificial neural network has a high degree of self-organization, self adjustment and self-learning ability, is a very complex nonlinear system, the prediction results are better than the traditional method. Disadvantages: neural network stock forecasting the model is easy to fall into the local minimum problem based on multilayer neural network and the complicated description of things, tend to increase the number of hidden layer, it will Lead to gradient diffusion problems, thus affecting the accuracy. This article is from the disadvantages of prediction model based on artificial neural network, such as the gradient diffusion and local minimum problem, and put forward the construction of deep learning of stock price trend forecast model based on the Boltzmann machine model is limited. Deep learning is developed based on neural network based on not only inherits the advantages of neural network method, and good overcomes the defects of the neural network method. In this paper, a prediction model based on Boltzmann machine limited to construct the deep belief network learning method is the use of K step of Gibbs sampling, comparing with the divergence algorithm, to train the entire depth of belief network. Finally the collection the GREE stock price information to train the prediction model and the prediction accuracy of this model is tested based on BP by God. The stock price prediction model as the prediction model comparison model, and by using examples Kweichow Moutai and BYD's stock price information to test the prediction accuracy of the model is two, the experimental results show that the deep learning of stock price trend forecast model based on good effect, and the accuracy is better than BP neural network prediction model. The innovation of this paper: (1) the prediction model of Boltzmann limited mechanism built deep belief networks of stock price trend based on learning method is adopted after comparing K step of Gibbs sampling after the divergence algorithm (CD algorithm) to train the prediction model. Finally, an instance is given. (2) the prediction model with the prediction of BP neural network prediction model of stock price trend based on the accuracy of the comparison.

【学位授予单位】：云南财经大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F832.51;TP18

【参考文献】