【作者】 刘娜；

【导师】 冯宏伟；

【作者基本信息】 西北大学 ， 软件工程， 2022， 硕士

【摘要】 股票投资分析作为投资过程中的重要一环,是成功地进行股票投资的重要基础。投资者常常通过分析大量股票指标数据(如收盘价、成交量等),探索市场变化规律,进而在投资中做出决策。由于个人对复杂数据的分析能力有限,如何以智能化技术赋能投资分析,已经成为股票投资领域的主流研究课题。本文将基于机器学习、深度学习智能算法对股票投资分析过程中的三个重要阶段:指数预测、选股、个股预测展开研究。主要工作如下:(1)针对因股票指标多周期波动混叠而导致的指数预测困难的问题,提出了一种融合股票指标多周期波动特征的行业指数涨跌预测方法(Sector index prediction that integrates the multi-cycle volatility features of stock indicators,IMCV)。该方法首先通过自适应噪声完备集合经验模态分解算法拆分股票指标中混叠的多周期波动得到各个指标对应的多周期波动特征;然后,设计了一个并行多分支神经网络模型,有侧重地融合多个股票指标的多周期波动特征,并实现行业指数涨跌预测。通过在申万一级行业指数的“农林牧渔”、“采掘”和“食品饮料”3个数据集上实验证明,IMCV具有比基线方法更优的涨跌类别预测性能;从并行多分支网络的参数中总结出的一般性规律可以为相关人员分析市场提供一定的指导。(2)针对因个股关联程度高导致的股票投资组合风险高的问题,提出了一种基于属性关联网络聚类的风险分散型选股方法(Stock selection based on attribute association network clustering,AANC＿SS)。该方法首先提取股票指标的核主成分特征构建股票属性关联网络,在消除了股票指标冗余的同时,将个股间隐式存在的连接关系显式化表示;其次,采用Louvain算法对属性关联网络中个股聚类,得到了个股连接模式上的差异性类簇;最终依据聚类结果进行多样性选股,降低了股票投资组合中的风险。实验证明,本文所述选股方法在震荡市场、连续上升市场以及包含多种状态变化的市场下均可以获得比基准指数及对比方法更高的收益风险比。(3)针对个股价格预测中股票间联动关系的表示及使用不充分的问题,提出了一种基于收益联动图的个股价格预测方法(Stock price prediction based on co-movement graph for stock return,CGSR)。该方法以个股价格联动的结果为导向构建收益联动图,避免了对股票价格联动的“原因”挖掘不足导致的联动关系表示不充分问题;进而,使用以扩散卷积门控循环单元为基本模块的Seq2Seq模型,有效地融合股票价格数据的联动特征与时序特征,实现收益联动图中个股的未来走势预测。实验证明,CGSR方法在连续多步股价预测中都具有比基线方法更小的误差;通过可视化发现,CGSR方法的未来一步股价预测曲线对真实价格曲线具有更小的滞后性。如上所述,本文提出的IMCV指数预测方法可以获得更精确的市场变动趋势,提出的AANC＿SS选股方法可以使股票投资承受更低的风险,提出的CGSR个股价格预测方法可以获得更精准的股票价格走势。综合三个方法,可以使投资者在投资过程中准确把握市场大环境、合理配置资金、精准掌握买卖时机,在获得更高收益的同时承担更低的风险。更多还原

【Abstract】 As an important part of the investment process,stock investment analysis is an important basis for successful stock investment.Investors often explore the law of market changes by analyzing a large number of stock indicator data(such as closing price,trading volume,etc.),and then make decisions in investment.Due to the limited ability of individuals to analyze complex data,how to use intelligent technology to empower investment analysis has become a mainstream research subjects in the field of stock investment.This thesis will conduct research on three important stages in the process of stock investment analysis based on machine learning and deep learning intelligent algorithms: index forecasting,stock selection,and individual stock forecasting.The main work is as follows:(1)Aiming at the difficulty of index forecasting caused by the aliasing about multi-cycle volatility of stock indicators,a method for predicting the rise and fall of sector indexes that integrates the multi-cycle volatility features of stock indicators is proposed(IMCV).The method firstly splits the aliased multi-cycle volatility in the stock indicators through Complete Ensemble Empirical Mode Decomposition with Adaptive Noise algorithm and obtains the multi-cycle volatility features corresponding to each stock indicator;then,a parallel multi-branch neural network model is designed,it integrates the multi-cycle volatility features of multiple stock indicators,and realizes the prediction of the rise and fall of the sector index.Through experiments on the three datasets of "Agriculture,Forestry,Animal Husbandry and Fisheries","Mining" and "Food and Beverage" of Shen Wan Primary Sector Indexes,it is proved that IMCV has better prediction performance of rising and falling categories than the baseline method;The general rules summed up in the parameters of the parallel multi-branch network can provide guidance for the relevant personnel to analyze the market.(2)Aiming at the high risk of the stock portfolio caused by the high degree correlation of individual stock,a risk-diversified stock selection method based on attribute association network clustering(AANC＿SS)is proposed.The method firstly extracts the kernel principal component features of stock indicators to build a stock attribute correlation network,which eliminates the redundancy of stock indicators and expresses the implicit connection relationship between individual stocks.By clustering stocks in attribute correlation network through the Louvain algorithm,the different clusters in the connection pattern of the stocks are obtained;finally,according to the clustering results diversely select stock,which reduces the risk in the stock portfolio.Experiments have shown that the stock selection method described in this thesis can achieve a higher return to risk ratio than the benchmark index and the compared method in the sideways market,the continuous rising market and the market with multiple state changes.(3)Aiming at the problem of insufficient representation and use of the co-movement relationship between stocks in individual stock price prediction,a stock price prediction method based on co-movement graph for stock return(CGSR)is proposed.This method builds the co-movement graph of stock return based on the results of the co-movement of individual stock prices,which avoids the problem of insufficient co-movement relationship representation caused by insufficient mining of the "cause" of the co-movement between stock prices.The Seq2 Seq model,based on the Diffusion Convolution Gate Recurrent Unit as the basic module,effectively integrates the co-movement features and temporal features of stock prices,and realizes the future price prediction of individual stocks in the co-movement graph.Experiments show that the CGSR method has a smaller error than the baseline method in the continuous multi-step stock price prediction;through visualization,it is found that the one-step stock price prediction curve of the CGSR method has a smaller lag to the real price curve.As mentioned above,the index forecasting method(IMCV)proposed in this thesis can obtain more accurate market trends,the proposed stock selection method(AANC＿SS)can make stock investment bear lower risks,and the proposed individual stock price forecasting method(CGSR)can obtain more accurate individual stock price trends.Combining the three methods can enable investors to accurately grasp the market environment,rationally allocate funds,and accurately grasp the timing of buying and selling during the investment process,so as to obtain higher returns while taking lower risks.更多还原