【作者】 杨杰；

【导师】 黄庆华； 司徒剑；

【作者基本信息】 华南理工大学 ， 工程硕士（专业学位）， 2017， 硕士

【摘要】 股票市场不仅是国家经济的晴雨表,更是股民投资和企业融资的重要途径。有效的股价趋势预测,不仅可以指导投资者进行合理的投资决策,更可以为国家制定相关经济政策提供参考。因此,股票预测一直是金融预测领域的重要研究方向,具有重大的现实意义和应用价值。然而,股票市场是一个极其复杂的动态系统,它受众多不确定性因素的影响,使得股票价格的波动表现出较强的非线性特征,极大增加了股票预测的困难度。为实现对股价趋势的预测,需在处理大量数据的基础上研究分析股市内部的复杂规律。数据挖掘技术能够揭示海量数据背后隐藏的信息,为股票预测提供了新的方法和思路。股票预测的关键在于构建合适的预测模型,而模糊推理系统具有较强的非线性映射能力,能够以任意精度逼近复杂的非线性关系,在预测模型构建的合理性以及适用性方面都具有其独特的优势。因此,本文将数据挖掘技术和模糊推理系统应用于股价趋势的预测,提出两种不同的预测方法。针对股价波动的非线性以及模糊推理系统在股票预测上存在规则获取困难和规则不够公正客观等问题,本文提出了一种基于双聚类和模糊推理的交易点预测方法。首先利用双聚类技术挖掘股票历史数据中的规律和信息,将它们作为专家知识构建模糊规则,然后通过模糊推理输出推理结果,并采用粒子群算法构建动态阈值模块,最后通过动态阈值模块将推理结果转换成对股价趋势的预测。另一方面,股价趋势的预测是一个上涨和下跌的二分类问题,但是由于股价变化规律复杂,单个分类器往往不能很好地描述整个数据集的特征,导致分类效果不佳。因此,本文提出一种基于朴素贝叶斯和AdaBoost的混合预测模型。首先利用双聚类算法挖掘股票数据中的股价趋势模式,然后将上涨模式和下跌模式组合成弱分类器,并采用朴素贝叶斯作为分类方法,最后利用AdaBoost算法将多个弱分类器组合成强分类器,以提升分类的准确率。由于每个弱分类器是由不同的趋势模式组合而成,保证了弱分类器的多样性,从而提升了强分类器的泛化能力。为了验证本文方法的合理性和有效性,设计了多个对比实验。实验结果表明,本文所提出的两个方法无论在个股还是股票指数的预测上都能获得比其他方法更高的平均收益率。更多还原

【Abstract】 The stock market is not only the barometer of national economy,but also an important means of individual investment and corporate finance.Effective forecast of stock price trend can guide investors to make reasonable investment decisions,moreover,it provides a reference for the country to develop relevant economic policies.Therefore,stock forecasting is a key research direction in the field of financial forecasting,and has great practical significance and application value.However,the stock market is a complicated dynamic system.It is affected by many uncertain factors,making the fluctuation of stock price presents a strong nonlinear characteristics which greatly increases the difficulty of stock prediction.In order to realize the prediction of stock price trend,we need to investigate the complex rules of stock market based on the processing of a large amount of data information.As it is known,data mining enables the reveal of hidden information behind the massive data,which provides a novel method for stock prediction.The key to the stock prediction is to construct appropriate prediction model.The fuzzy inference system has strong nonlinear mapping ability,which can approximate the complex nonlinear relationship with arbitrary precision,leading to the unique rationality and applicability in constructing a prediction model.Thus,for the purpose of stock price trend prediction,we propose two different forecasting methods based on data mining technique and fuzzy inference system.Due to the volatility of stock price,coupled with the difficulties of fuzzy inference system on stock predicting,such as the difficulty in obtaining rules,the inability to generate rules automatically,and the lack of objectivity and justice of the rules,this paper presents a forecasting method of trading points based on biclustering and fuzzy inference.Firstly,using the biclustering technology to extract the laws and valuable information from historical data of the stock,the extracted information will be treated as an expert knowledge to build fuzzy rules automatically.After that,fuzzy inference is employed to output reasoning results which will be finally transformed to the prediction of stock price trend according to a dynamic threshold module constructed by particle swarm optimization algorithm.On the other hand,to forecast the stock price trend is to classify the trend into rising or falling type.Owing to the complex change rules of stock price,a single classifier performs poorly since it is not able to describe the characteristics of the entire data set.Consequently,this paper proposes a hybrid forecasting model based on naive Bayes and Adaboost.At first,using the biclustering algorithm to extract the pattern of stock price trend,then combining the rising patterns and falling patterns to construct weak classifiers by naive Bayes,finally the Adaboost algorithm is used to combine multiple weak classifiers into a strong classifier,which could improve the accuracy of classification.Moreover,considering that different weak classifier is composed of different trend patterns,the diversity of the weak classifiers can be ensured,resulting in the enhancement to the generalization ability of the strong classifier.In order to verify the rationality and effectiveness of the proposed methods,several comparative experiments have been designed.The experimental results show that the proposed methods obtain higher average return rates both in stock and stock index when compared with other methods更多还原