基于EEMD-MARS-SVR模型的证券市场分析

发布时间：2018-01-31 22:21

本文关键词： 集合经验模态分解多元自适应回归样条法支持向量回归金融时间序列　出处：《上海社会科学院》2017年硕士论文　论文类型：学位论文

【摘要】：本文提出一个由集合经验模态分解(EEMD)、多元自适应回归样条法(MARS)、支持向量回归(SVR)三部分构成的分析预测模型,并将其用于金融时间序列的建模预测。其中,EEMD是集合经验模态分解(EMD)的改进版,将白噪声引入EMD中,有效解决了 EMD中一个IMF存在多种频率的缺陷,一般应用于通信和IT领域的数据处理,其应用于金融领域的优点在于无需事先设定任何基函数。MARS利用样条函数来拟合局部的复杂非线性关系,并能够通过剪枝过程得到解释变量的权重,可用于变量的筛选。SVR是机器学习中常用的预测模型,广泛应用于IT和金融领域,能够针对给定的训练集训练出相适宜的模型,并作出预测。本文首先对现代股票定价理论的发展、常用预测模型和宏观经济变量对证券市场的影响三方面进行了文献上的回顾,确定了以EEMD-MARS-SVR为分析模型的研究框架,并引入了 10个大类,84个小类的公共经济变量指标。本文认为这些公共经济变量指标都是有可能对股票市场产生影响,并留待筛选的。之后,本文对A股市场5个指数时间序列和沪深300的300只成分股时间序列分别进行建模分析,时间区间为2010年6月1日至2016年11月30日,一共66个月,1579个交易日的数据。分析过程中,首先对目标时间序列进行EEMD分解,分解得到了数个IMF子序列,之后将IMF子序列重组成高频序列、低频序列和趋势序列。之后,利用MARS方法,在84个公共经济变量指标中分别筛选出与高频序列和低频序列有关的变量指标。这些变量指标可以分别用来对高频序列和低频序列进行预测分析。最后,将筛得的公共经济变量和目标序列引入到SVM分析框架中,进行预测。为了评价预测效果,本文引入了平均绝对误差(MAPE)和方向对称度量(DS)进行衡量,并与其他预测模型进行比较,结果显示本文预测方法在样本内预测和样本外预测上都具有显著的优势。在模拟选股回测过程中,本文方法在理论上具有投资价值。
[Abstract]:This paper presents an analytical and predictive model consisting of three parts: set empirical mode decomposition (EMD), multivariate adaptive regression spline method (MARSS) and support vector regression (SVR). It is applied to the modeling and prediction of financial time series, in which EMD is an improved version of EMD, which introduces white noise into EMD. It effectively solves the defect of multiple frequencies in a IMF in EMD, which is generally applied to data processing in the field of communication and IT. The advantage of its application in the field of finance is that it does not need to set any basis function. Mars uses spline function to fit local complex nonlinear relations and can get the weight of explanatory variables through pruning process. SVR, which can be used to filter variables, is a commonly used prediction model in machine learning. It is widely used in the fields of IT and finance, and can train suitable models for a given training set. First of all, this paper reviews the development of modern stock pricing theory, the influence of common forecasting models and macroeconomic variables on the securities market. The research framework based on EEMD-MARS-SVR is established, and ten kinds of analysis models are introduced. This paper holds that these public economic variables are likely to have an impact on the stock market and remain to be screened. In this paper, five index time series and 300 component stock time series of Shanghai and Shenzhen 300 in A share market are modeled and analyzed respectively. The time interval is from June 1st 2010 to November 30th 2016. A total of 66 months, 1579 trading days of data. In the analysis process, the target time series is first decomposed by EEMD, and several IMF sub-sequences are obtained. The IMF subsequences are then reorganized into high frequency sequence, low frequency sequence and trend sequence. Then, the MARS method is used. The variables related to high frequency sequence and low frequency sequence were screened out from 84 common economic variables. These variables can be used to predict and analyze the high frequency sequence and low frequency sequence respectively. The selected public economic variables and target sequences are introduced into the SVM analysis framework for prediction. In this paper, mean absolute error (MAPE) and directional symmetry metric (DSs) are introduced and compared with other prediction models. The results show that the prediction method in this paper has significant advantages in both intra-sample prediction and out-of-sample prediction, and in the process of simulating stock selection, the method has investment value in theory.
【学位授予单位】：上海社会科学院
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F224;F832.51

【相似文献】