Research On Multivariate Time Series Clustering Algorithm And Application Of Stock Selection Strategy

Time series data is ubiquitous in production and life,and many scholars have studied time series mining technology.With the deepening of research,the mining technology of unary time series has gradually matured,and it has been widely used in the field of financial securities.For example,using time series to analyze the characteristics of securities markets and their relationships,financial time series analysis has become an important direction in financial research.Multivariate time series data uses multiple indicators to describe objects at the same time,and the information that it expresses and the hidden knowledge is far more than that of univariate time series.However,due to its high dimensionality and complexity,the current research on multivariate time series is relatively inadequate,and the analysis based on multivariate data is also lacking in financial applications.Aiming at the different angles of clustering technology two new multivariate time series clustering models are proposed:(1)A feature weighted multivariate time series clustering method is proposed by introducing the at feature weight calculation and the construction of fuzzy membership matrixes.The method assigns different weights to different attributes according to the degree of data dispersion in different dimensions.It can transform the original multivariate time series into a fuzzy membership matrix using distances between multivariate time series objects.Then new method combines the feature weight and signal dimension distance to construct the synthetic fuzzy membership matrix.Finally,the clustering result can be obtained by fuzzy c-means algorithm.This method can effectively improve the clustering quality of multivariate time series and does not need to introduce more parameters.(2)Propose a nonparametric clustering method based on community detection from the perspective of complex networks.The new method avoids parameter selection in the clustering process by comprehensively considering several numbers of nearest neighbors.It can adaptively determine the number of communities in the dataset and realize nonparametric clustering of the multivariate time series.The experiments confirm that the new clustering method reduces the parameters while ensuring the quality of clustering results and clustering efficiency,and is suitable for clustering data sets without prior knowledge.In addition,this paper combines new clustering models with the financial stock market.First,the nonparametric clustering method is used to find the community structure that exists in the multi-stock dataset,and the appropriate number of clusters is determined.Then use the feature weighted clustering method gets the final clustering result.Based on the clustering results,two stock selection strategies with different objectives are proposed.The validity of the stock selection strategy is tested in combination with the Markowitz model.The experiment proves that the portfolio results using the new clustering model and stock selection strategy are of reference value to investors.