【作者】 谭华；

【导师】 谢赤；

【作者基本信息】 湖南大学 ， 管理科学与工程， 2008， 博士

【摘要】 随着金融全球化与自由化的推进,金融行业的运行效率与水平在很大程度上决定了一个国家的经济竞争力,而信息化技术越来越成为影响金融行业创新能力与发展水平的重要因素。近年来,许多金融机构开始运用先进的信息技术与智能决策支持技术对业务系统中积累的海量数据进行深入分析,以发现各种有价值的规律。数据挖掘技术作为一种新兴的智能决策支持技术,已经开始在金融行业的部分领域得到应用。在此背景下,研究如何从金融数据中挖掘出各种信息,更好地认识、掌握并利用其规律,无疑具有重要意义。与此同时,金融市场的信息具有不确定性众多、非线性和信息数据的模糊性及非结构性等特点。金融市场中的不确定性包含时间的不确定。这些问题都值得研究者们进行深入研究。不确定性方法与数据挖掘技术有一定的重叠性,两者在单独使用时都具有一定局限性。数据挖掘中存在不确定性问题,金融时间序列分析中也存在不确定性问题,并且传统数理统计方法不适用于从大量的数据中主动发现各种潜在规则,而不确定性方法在单独进行预测时会遇到小数据量等问题。本文根据具体选定的证券市场行情预测对象,将不确定性方法和数据挖掘技术的各自优势结合起来,得到一种基于不确定性方法和数据挖掘技术的不确定时态数据挖掘方法(UTDM)。该方法能更好地发挥不确定性方法和数据挖掘技术的优势,为证券市场的预测提供更好的技术分析方法,从而为投资决策者提供更为精确的定量分析结果。在对不确定性方法和数据挖掘技术的相关理论、研究发展现状及不足进行深入讨论的基础上,选取不确定性方法和数据挖掘技术中几种具有代表性的方法:在不确定性方法中选取模糊集方法、模糊相似关系下的模糊粗糙集及灰色理论;在数据挖掘中选取关联规则、神经网络等方法,用于构建证券市场的有效分析方法。在此基础上分别得到模糊相似关系下的模糊粗糙集挖掘预测方法、趋势特征挖掘预测方法、时间序列模糊关联规则挖掘预测方法及多灰色神经网络预测方法,用这些方法分别解决证券市场中短期的个股及股指的预测问题。论文后续部分则以此为基础进行展开。本文通过模糊相似关系下的模糊粗糙集和数据挖掘技术对股票价格进行预测研究,从证券市场的大量数据中得到强规则。利用模糊集和粗糙集方法将股票价格进行预分类,并按时间属性进行分组,通过给出的模糊相似关系下的模糊粗糙集计算每组的真值,利用数据挖掘技术获得候选属性,最终得到相应时间段内的有用规则,根据所得规则预测某一具体时间段内股票价格的变化趋势。将模糊粗糙集扩展到模糊相似关系下的模糊粗糙集,并应用到对股票价格的预测中,能较原模糊粗糙集方法得到更多的有用规则,准确率更高。将股票中的时间序列转换为以价格变动率为变量的时间序列进行分析,并对趋势特征提取、聚类算法进行改进,将时间序列的预测问题转化为频繁和有效特征集来发现问题,进而对趋势特征模式进行挖掘预测,根据连续一段时间内的涨跌情况判断市场的发展趋势。将时间序列模糊关联规则应用于证券市场的交易规则抽取。选用聚类方法对模糊集属性进行离散化,构造模糊集和隶属函数,引入时间维度,提出适合股票交易规则抽取的时间序列模糊关联规则算法,对一定时间段内股票间及行业间的关联关系进行最大限度的挖掘预测。提出将3种灰色预测模型,即残差GM(1, 1),无偏GM(1, 1)和pGM(1, 1)与神经网络预测模型有机组合起来,建立一种新的多灰色神经网络组合预测方法,并通过对中国证券市场综合指数进行模拟预测进行验证,对证券市场综合指数的预测及比较说明了该组合预测精度的有效性。更多还原

【Abstract】 With the development of financial globalization and liberalization, the efficiency of the financial sector determines the level of a country’s economic competitiveness at a large extent, and information technology is one of the important factors to impact the financial industry innovation ability and development level. Many financial structures use more advanced information technology and intelligent decision support technology to find out useful rules by analyzing massive data, which are accumulated in operational systems. As a new intelligent decision support technology, data mining technology has been used in some fields of financial sector. Based on these conditions, how to mining more useful information from financial data to understand, master, and use its own rules and undoubtedly has special significance.The information of financial market has the characteristicses of uncertainty, nonlinear, fuzzy nature of the information data and non-structure. The uncertainty in financial market, not only includes the uncertainty of time, but also the uncertainty of information and technology. All of these are worthy of in depth research.Uncertainty methods and data mining technology have similar usages to some extent, but they have some limitations when we use them alone. As we know, it has uncertainty in data mining, and the analysis of financial time series also has uncertainty problems. Furthermore, traditional mathematical statistics motheds don’t fit to find potential rules from large amounts of data. In this paper, we mainly study uncertain knowledge and data mining technology to propose a new method-uncertainty temporal data mining method (UTDM), we can obtain a series of new methods from UTDM, and use the new methods to solve the problems we meet during the mining and fortcasting processes. We use these methods to study securities markets, primarily for the securities markets’trends, stock price forecasts, and stock index forecasts.We choose some typical methods in uncertainty methods and data mining technology to setup useful methods to analysis securities markets: fuzzy sets, fuzzy rough sets based on fuzzy similar relationship, gray theory, association rules, networks method. We summarize the related theories and methods of uncertain knowledge, data mining technology and financial time series, point out the limitations of the current studies, lay the foundation for uncertainty temporal data mining method, clarify the objections of this study, and give us important informations to make the further theoretical research.We use fuzzy similar relationship through the rough sets and fuzzy data mining techniques to predict stock prices, obtain strong rules from the securities markets and economic data. We use fuzzy rough sets and data mining technology to forecast stock price at certain given time. At first, we use fuzzy sets and rough sets to make the stock price into some groups based on the attribute”time”, and then compute each true value by the fuzzy rough sets, we can obtain the candidate properties by data mining method. In the end, we can get the useful rules and forecast the change trading of stock price in a certain time. The result of the method we pointed out is more exact than other method. In this paper, we share the time series into a variable rate of change in prices of the time series analysis, improve the trend of feature extraction and clustering algorithm, transfore time series prediction into frequent and effective feature sets to discover problems to mining forecast rules, judge the market trends according to the change in market during a period of time.In order to assist stock investors to make reasonable decision, we use fuzzy association rules to mining securities markets’exchange rules in this paper. Firstly, we disperse the fuzzy sets’attributes by clustering method. Secondly, fuzzy sets and their corresponding membership function of the quantitative attributes are generated by means of the mediods. Finally, we give the algorithm a name which is called FARS. At the end of the 6th chapter, we study the fuzzy association rules based on temporal forms. We combine three gray models: residual GM(1, 1), unbiased GM(1, 1), pGM(1, 1) and neural network to propose a new combination forecasting model. And use it to make forecasts on Composite Stock Price Index of the securities markets in Shanghai, China. The results show that this model could gain optimized forecasting value and could be taken as an effective tool to predict Shares Price Composite Index. The model makes full use of grey prediction modeling information which requires less information and the neural network has strong ability and good nonlinear mapping and fault-tolerant, self-organizing and adaptive characteristics. The forecasts on Composite Stock Price Index of the securities markets in China shows that the combination of the effectiveness of prediction accuracy.更多还原