基于动态时间弯曲的金融时间序列聚类研究

发布时间：2019-03-14 09:05

【摘要】：随着人类进入大数据时代,通过数据挖掘技术将时间序列数据库中隐藏的、有价值的知识挖掘出来得到了愈多的关注,其相关技术己被成功地运用到各个领域。时间序列相似性度量可以衡量时间序列之间相似程度的方法,其度量结果可用于分类、聚类、相似性搜索等数据挖掘任务中。时间序列聚类是时间序列数据挖掘领域中重要的挖掘任务之一,不同的时间序列聚类方法,可以挖掘出不同的隐含信息。本文以时间序列为研究对象,探讨时间序列的相似性度量方法和聚类方法,促使方法可以充分与灵活地应用到时间序列数据挖掘中,然后撷取潜在珍贵的信息和知识。本文的主要研究内容如下:(1)以数值分布特性和趋势波动特征为出发点,提出基于数值符号和形态特征的相似性度量方法。新方法能够充分反映时间序列数值分布和形态特征,有效地提高了时间序列相似性的度量效果。(2)针对传统聚类方法通常需要确定具体聚类数目,及未能充分反映时间序列整体空间结构和相互影响关系的问题,提出一种基于中心度的标签传播时间序列聚类方法。该方法无需指定具体聚类数目即可实现自动聚类,并且根据不同参数构建不同的网络空间结构,聚类数目能够对此进行相应地调整,提高其在时间序列聚类的性能。(3)动态时间弯曲和时间序列聚类在金融领域的应用。一方面,以动态时间弯曲和经典时间序列聚类方法为基础,在金融领域进行进一步探索。针对股票联动性的研究,挖掘股票的隐含信息,对监管部门和投资者决策起着一定帮助作用。另一方面,以沪深300指数为标的指数,利用新的相似性度量方法和聚类方法对现货股票进行聚类分析,选定追踪成分股,并建立优化模型来获得成分股在投资组合中的优化权重,并使得新方法确定的成分股更能准确地模拟标的指数,且能够满足不同投资喜好的投资者投资要求。研究内容通过数值实验分析,并且通过比较研究领域的相关方法,检验了新方法的性能,进一步完善时间序列相似性度量和聚类的研究,同时在一定程度上扩展了时间序列数据挖掘相关理论和提升了方法在金融时间序列数据领域中的应用性能。
[Abstract]:With the entry of big data era, the more attention has been paid to mining the valuable knowledge hidden in time series database through data mining technology, the more attention has been paid to it, and its related technology has been successfully applied to various fields. The similarity measurement of time series can be used to measure the degree of similarity among time series, and the results can be used in data mining tasks such as classification, clustering, similarity search and so on. Time series clustering is one of the important mining tasks in the field of time series data mining. Different time series clustering methods can mine different hidden information. Taking time series as the research object, this paper discusses the similarity measurement method and clustering method of time series, so that the method can be fully and flexibly applied to time series data mining, and then extract potentially precious information and knowledge. The main contents of this paper are as follows: (1) the similarity measurement method based on numerical symbols and morphological features is proposed based on numerical distribution characteristics and trend fluctuation characteristics as the starting point. The new method can fully reflect the numerical distribution and morphological characteristics of time series and effectively improve the effect of measuring the similarity of time series. (2) in view of the traditional clustering methods, it is usually necessary to determine the number of specific clusters. In this paper, a clustering method of label propagation time series based on centrality is proposed, which fails to fully reflect the global spatial structure and interaction relationship of time series. This method can realize automatic clustering without specifying the number of clusters, and construct different spatial structure of network according to different parameters, which can be adjusted accordingly. Improve its performance in time series clustering. (3) the application of dynamic time bending and time series clustering in financial field. On the one hand, based on dynamic time bending and classical time series clustering methods, further exploration is carried out in the field of finance. In view of the research of stock association, mining the implicit information of stock plays a certain role in the decision-making of regulators and investors. On the other hand, taking the Shanghai-Shenzhen 300 index as the target index, we use the new similarity measure method and the clustering method to cluster the spot stock and select the tracking component stock. The optimization model is established to obtain the optimal weight of the component stocks in the portfolio, and make the component stocks determined by the new method can more accurately simulate the underlying index, and can meet the investment requirements of investors with different investment preferences. The research contents are analyzed by numerical experiments, and the performance of the new method is tested by comparing the related methods in the field of research, and the research on similarity measurement and clustering of time series is further improved. At the same time, the theory of time series data mining is extended to a certain extent and the application performance of the method in the field of financial time series data is improved.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.13

【参考文献】