Research On Method Of Analysis On Stock Comovement Based On Time Series Data Processing

With the continuous development of Chinese financial market,the stock market has reached the scale of more than 3,800 listed companies.Facing with such a large-scale and rapidly changing stock market,investors need to construct a reasonable investment portfolio to avoid risks and obtain high returns which is difficult.It is an important investment strategy to use the stock comovement effect to build a reasonable portfolio.At present,there is not any clear definition and measurement criteria for the stock comovement effect in the industry,and the judgment of the stock comovement effect mainly refers to the manual classification and empirical method.The relevant academic research is just based on the macro level or some special cases.In recent years,the rise of artificial intelligence technology has enabled more and more machine learning algorithms to be used in various industries.How to use machine learning algorithms to study stock comovement based on the characteristics of stock time series data has also become a research hotspot.Based on the stock timing data,this paper studies from the two perspectives of the comovement between the individual and among the stock group.The main research content of this paper are as follow:The construction of stock time series dataset and the realization of data visualization service.In the paper,this paper designs and implement a set of automatic methods to obtain multi-source data on the Internet and integrate data to construct high-quality stock time series dataset.At the same time,this paper designs and implements a multiperspective visualization system of comparison of stock trend for assisting the research of stock comovement and other related topics.The design of similarity indicator for stock individual comovement analysis.We designs three kinds of indicators from different perspectives,including the correlation coefficient indicator of stock time series data based on the perspective of statistical correlation,the time-weighted soft DTW(tw-s DTW)indicator based on the similarity of stock time series data,and the vector similarity indicator of stock vector which is learned from stock sequences based on the perspective of representation learning.The superiority and validity of these indicators is verified by the experiment in this paper.The design of clustering method for stock group comovement analysis.Based on the previously designed similarity indicators,this paper further designs clustering methods suitable for stock group comovement analysis,including spectral clustering based on relationship matrix constructed from similarity indicators,and improved K-means clustering based on time-weighted soft DTW,K-means clustering combined with global alignment kernel function and self-organizing map neural network clustering algorithm based on stock vector.In addition to using traditional clustering evaluation indexes,this paper also designs stock clustering evaluation indexes that describe the comovement of stock groups and verifies the better performance of different clustering methods designed in this paper compared with manual classification method and traditional clustering methods by comparative experiments.Meanwhile,the applicability of these clustering algorithms is also explored.