【作者】 沈巍；

【导师】 牛东晓；

【作者基本信息】 华北电力大学（北京） ， 管理科学与工程， 2011， 博士

【摘要】 反映股票市场总体走势的各类股票指数一直以来是经济波动的晴雨表,是政府对股市进行调控和监督的重要依据,同时也是投资者进行投资不可或缺的重要投资指南。因此,如何对股指进行准确预测,使得政府能够对股市进行有效监管与调控、投资人能够有效避免投资风险,就成为金融领域理论界长期以来的研究焦点。但是,由于股指的波动受宏观、微观、内部、外部等诸多因素的影响,因此,对股指的预测存在较大的难度。本文在前人研究的基础上,对股指预测理论方法及模型构建做了如下几个方面的研究：(1)股指波动影响因素及股指预测模型特点研究。由于影响股指波动的因素众多,本文从宏观经济、技术指标、心理因素三个层面对影响股指波动的因素进行了系统的归纳与总结；结合股指波动特点和影响因素,分析并归纳出股指预测模型应该具备的功能和要求；(2)股指波动统计类预测模型与创新类预测模型比较研究。本文首先在理论上将两种预测模型进行了比较,对两种预测方法在建模的理论基础、数据要求与处理、模型稳定性与适用性、预测精度与长度等方面进行了比较与探讨。其次,将两类模型进行了实证方面的比较,进行了单一指标输入和多指标输入的大样本、小样本及组合模型的比较；(3)运用生物进化算法对神经网络股指预测模型进行优化。本文运用被GA, PSO, AFSA三种生物进化算法优化后的RBF神经网络对上证综指进行预测,并将三种优化算法得到的预测结果进行了预测精度的比较；(4)建立基于数据挖掘的RBF+AFSA股指预测模型和GA-BP股指预测模型,运用数据挖掘技术,将影响股指波动的数量性指标逐一带入预测模型,将表现不好的淘汰,表现较好的再进一步进行优化组合,直到寻找到一个预测精度最好的最优组合为止；(5)建立基于知识挖掘的FPBP股指预测模型和REPTree+RBF+AFSA股指预测模型,在数据挖掘的基础上,运用知识挖掘技术,将影响股指波动的文本性因素,包括宏观经济性因素和心理因素等,进行筛选、分级、归类等预处理,然后带入相应的预测模型,使预测结果能够更为接近真实值。论文综合应用股指预测理论、经济学理论、神经网络理论、群智算法理论、数据挖掘理论、知识挖掘理论以及现代数学方法与计算科学技术,对股指预测理论及方法进行了全面系统的研究。具体研究方法包括：(1)理论与实证相结合的研究方法；(2)定量与定性相结合的研究方法；(3)归纳与比较相结合的研究方法等。本文的主要创新点包括：(1)将统计类的股指预测模型与创新类的股指预测模型在理论与实证两方面进行了比较,结果显示,创新类预测模型在股指预测方面优于统计类预测模型。(2)建立了一种新的对RBF神经网络进行修正的模型--鱼群修正的径向基神经网络预测模型(RBF+AFSA),在数据挖掘的基础上,运用该模型对上证综指进行了预测,并将该种模型的预测结果与其它预测模型进行了比较。(3)运用知识挖掘理论,建立了一个FP_Tree+BP混合预测模型——FPBP预测模型,选用FP Tree决策树技术,对影响股指波动的文本因素进行了筛选,寻找出影响股指波动的主要文本因素：经济增长、CPI、货币政策、投资者心理和突发事件。之后将文本因素做适当数量化处理后输入BP模型。实证结果显示,加入文本因素后的FPBP模型的预测效果好于仅输入数量化指标的BP模型。(4)建立了一个基于知识挖掘的REPTree+RBF+AFSA模型。在RBF+AFSA模型数据挖掘的基础之上,将数量化指标的训练误差和文本指标都带入REPTree分类器做决策树IF-Then规则分析,得到一个预测调整率。根据预测调整率调整数据挖掘得到的预测结果。实证结果表明,引入了文本因素之后的REPTree+RBF+AFSA模型预测精度得到了一定程度的提高。论文研究结果表明：创新型智能化预测模型在股指预测方面优于传统的统计类预测模型；通过建立基于数据挖掘的股指预测模型,对影响股指的数量化因素进行数据挖掘,表明无论在何种预测模型中,经过数据挖掘的多因素组合指标的预测效果都好于单一指标的预测效果；通过建立基于知识挖掘的股指预测模型,对影响股指波动的文本因素进行筛选和分类,并将其带入相关预测模型,结果显示,加入文本因素后的预测模型其预测精度会有一定程度的提高。更多还原

【Abstract】 Stock indices reflecting the general trend of stock market are regarded as the parameter for economic movement. They serve as reference for government in market regulation and guidance for investors in buying and selling financial product. Therefore, stock-index-forecast has long been a research focus in financial fields as it is of great importance both to the government and individual investors. However, as the movements of stock indices are subject to influence of many macro, micro, internal and external factors, it is a very difficult task to have relatively accurate prediction. In this paper, by absorbing results from previous literature, I carried out research on theory and method for forecasting stock indices in the following aspects:(1)Analysis on factors with impact on stock index fluctuation and characteristics of forecasting models. Given the fact that stock index movements can be influenced by many factors, I tried to analyze and summarize them at three layers, namely, macroeconomics, technique indicators and psychological factors. On the basis of analysis, I put foreward functions and requirements of stock index forecasting models.(2)Comparative analysis of statistical forecasting methods and intelligent forecasting methods. First, comparison of the theories of the above two methods, which includes comparative study of theoretical bases, data requirement and processing, model stability and applicability, forecasting accuracy and length. Secondly, in empirical study, I compared the large samples, small samples and compound models.(3)Research on optimization of neuro-network models. To address the problem of low speed and local minimum, I applied intelligent algorithms such as GA, PSO and AFSA, to optimize the neuronetwork algorithm, then I used the optimized algorithms to forecast the indices of Shanghai Stock Exchange and compared the results achieved by the above three models.(4)Through using data mining technique, I introduced quantitative indicators into forecasting model, eliminate those with poor performance and further combined the indicators with good performance into optimized groups, till I found a combination with the highest forecasting accuracy. (5)Apart from data mining, knowledge mining method is also used in this paper. Text factors including macroeconomic and psychological factors were pre-processed before being introduced into corresponding forecasting models, to increase accuracy.Innovative points in this paper include:(1) theoretical and empirical study and comparison of statistical and non-statistical forecasting models plus comparison of long-term and short-term forecasting results of the above two methods; (2)building a new modified RBFNN model---BF+AFSA;(3) using RBF+AFSA to forecast stock indices of Shanghai Stock Exchange, and compared the algorithm with other swarm intelligent models;(4)applying data mining technique to select major factors with impact on the movement of stock indices and introduce them into forecasting models; (5) using FP_Tree technique to select factors with impact on stock index fluctuation, for instance, economic growth, monetary policies, CPI, psychological factors and breaking out incidents, and establish a indicator system of factors with influence on stock index movement;(6)using knowledge mining technique in forecasting procedures to rank and regroup the text factors, then introduced then to REPTree model, and by so doing, forecasting accuracy was increased.Research result of this paper indicates:innovative intelligent forecasting methods are superior to traditional statistical forecasting methods in short-term stock index forecasting; through using data mining and knowledge mining technique, that is, when quantitative and non-quantitative factors are introduced into related forecasting models, forecasting accuracy increases to some extent.更多还原