基于一种改进特征选择方法的股票分类研究

发布时间：2018-03-24 00:18

本文选题：混合式特征选择　切入点：改进遗传算法　出处：《北京交通大学》2017年硕士论文

【摘要】：股票分类一直是金融和投资领域中重要的研究方向,并且是一项具有挑战性的任务。股票分类的关键是股票特征选择及分类模型的确定。近年来许多研究表明,成功的特征选择方法可以提高分类的准确性。那么,问题的关键是如何选择股票分类中具有代表性的特征。本文提出一种改进的混合式特征选择方法。这种新的两阶段特征选择方法结合了三种过滤式方法和改进的遗传算法(IGA),其中,IGA作为封装式方法,用于识别最优特征子集,从而提高分类的准确性和泛化能力。首先,使用三种不同的过滤式方法,实现特征的快速排序,并由此获得每个特征的权重。特征的权重作为重要的先验信息应用在改进的遗传算法中。在下一阶段,首先,先验信息用在遗传算法的初始种群设定上,从而加速种群的收敛;然后,改进交叉和变异算子,降低随机性造成的不利影响,引导搜索最佳特征子集,通过先验信息增加群体多样性;最后,引入多种群遗传算法(MPGA),可以同时结合三种过滤式特征选择方法的优势并做出鲁棒和准确的决策。本文选取从2009年至2013年在上海证券交易所交易的A股1108家公司的4173个真实历史数据,以股票年回报率的变化趋势为输出变量,选取七大类24个具有代表性的财务指标作为输入变量,实证本文提出的混合式特征选择方法的有效性。通过横纵向对比实验,比较分析了单种群改进遗传算法(SPIGA)和传统遗传算法(GA),GA、MPGA和本文提出的IGA。实验结果表明,IGA具有比过滤式方法和传统遗传算法更高的精度水平和更好的泛化能力。IGA可以成功选择出股票分类的重要特征,通过选出的特征重新对股票分类,结果表明,IGA是有效并且适用的特征选择工具,选择的特征是股票分类的良好指标。
[Abstract]:Stock classification has always been an important research direction in the field of finance and investment, and it is a challenging task. The key of stock classification is the selection of stock characteristics and the determination of classification model. Successful feature selection methods can improve the accuracy of classification. The key to the problem is how to select the representative features in stock classification. In this paper, an improved hybrid feature selection method is proposed. This new two-stage feature selection method combines three filtering methods and an improved one. Genetic algorithm (GA), in which IGA is used as encapsulation method, It is used to identify the optimal feature subset, thus improving the accuracy and generalization of classification. Firstly, three different filtering methods are used to realize the rapid sorting of features. The weight of each feature is obtained. As an important priori information, the weight of the feature is applied in the improved genetic algorithm. In the next stage, the prior information is used in the initial population setting of the genetic algorithm to accelerate the convergence of the population. Then, the crossover and mutation operators are improved to reduce the adverse effects caused by randomness, to guide the search for the best feature subset, and to increase population diversity through prior information. By introducing multi-population genetic algorithm (MPGA), we can combine the advantages of three filtered feature selection methods and make robust and accurate decisions. This paper selects 1108 A-share companies traded on the Shanghai Stock Exchange from 2009 to 2013. 4173 real historical data, Taking the change trend of stock annual return as output variable, 24 representative financial indexes of seven categories are selected as input variables, and the effectiveness of hybrid feature selection method proposed in this paper is demonstrated. The comparison between single population improved genetic algorithm (SPIGA) and traditional genetic algorithm (GA) and IGA proposed in this paper shows that IGA has higher precision level and better generalization ability than filter method and traditional genetic algorithm. Successful selection of important characteristics of stock classification, The result shows that IGA is an effective and applicable feature selection tool and the selected feature is a good index of stock classification.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F832.51;TP18

【参考文献】