Research On Similarity And Evaluation Index Of Stock Clustering

With the rapid development of IT technology in the financial industry,companies have accumulated huge amounts of financial data.How to explore valuable information from the internal structure and trends of financial time series is a challenging and practical task.Faced with huge financial data,in addition to classical time series analysis techniques,data mining and machine learning methods can also be used.Data-driven machine learning and data mining are good at mining patterns and predicting trends.Multi-level and multi-angle analysis based on machine learning technologys enables managers to predict future trends based on past corporate conditions,thereby providing scientific and rational decisions.Clustering stocks based on correlations is not only the key to capturing market information and developing investment strategies,but also helping to understand the operating mechanisms of the market.Based on the most common stocks data,this thesis explores the useful information hidden in stocks data by clustering algorithm.In this thesis,the similarity of stocks clustering and the evaluation index of clustering results are studied.The main research contents are divided into the following two aspects.Firstly,in the aspect of similarity measure of stocks clustering,the linear fusion Linear_SLU and the nonlinear fusion NonLinear_SLU based on Copula function are proposed.The correlation is the most common used measure of stocks time series,but the classical correlation has limitations whether it is Pearson correlation or rank correlation.In order to make full use of the information of different correlations,the fusion correlation based on Copula function are discussed in detail.The linear fusion coefficient and the nonlinear fusion correlation based on Copula function are proposed,on this basis we measure the similarity between stocks and use clustering methods to get clustering results.Clustering experiments are done on 50 SSE stocks inthree different periods of stability,rise and fall to prove the validity of the fusion correlation.The results show that the clustering method based the nonlinear fusion correlation can make the clustering results more reasonable in the period of rising,falling or stationary;clustering results can also be obtained effectively in the stationary period based on Linear_SLU correlation.Secondly,in the aspect of evaluation index of stocks clustering,a stocks clustering evaluation index BEF based on the effective frontier of portfolio is proposed.Starting from the effective frontier curve of portfolio,this method evaluates the quality of clustering by the weighted distance between the effective frontier curve of all stocks and the effective frontier curve of stocks selected from each cluster.Then,we discuss the properties of BEF based on the clustering experiments of stocks data.We study the trend of BEF with the number of clusters,the stability of BEF under different similarities and the optimal number of stocks clustering based on BEF.The effectiveness of BEF is verified by experiments.The innovation work of this thesis can be summarized as the following two points.(1)Linear fusion correlation Linear_SLU and nonlinear fusion correlation NonLinear_SLU are proposed based on Copula function.The experimental data are used to analyze the effective of these two correlation coefficients for stocks clustering.(2)Evaluation index BEF for stocks clustering based on effective frontier theory of portfolio is proposed.The characteristics of the BEF are analyzed through experiments.At the end of the paper,the work done is summarized,and the future research directions are put forward.