The Research On Pattern Mining Of Financial Time Series

A large number of business activities of financial institutions are increasingly dependent on large amounts of historical data analysis, mining the valuable information from the financial data has become the necessary means and core work of making financial management decisions intelligent. Time series is a very important data type in financial field, and the traditional methods of financial time series analysis cannot effectively deal with large-scale data sets and cannot take the initiative from the large amounts of data to find underlying rules. Data mining is an effective way for pattern mining of financial time series.With the characteristics of financial time series analysis and financial needs, this paper uses data mining techniques to research the financial time series pattern mining methods, including financial time series segmentation and representation, financial time series similarity measure, financial time series rule discovery and clustering, and so. These studies have the very important significance for finding the implied rules and predicting the trends in the financial field. The results of this study include the following:(1)The trend and form of financial time series contains a lot of information, however, many time series segmentation methods destroy the form, or smooth out the key points. To solve this problem, this paper proposes Multi-level Extreme Point Segmentation method (MEPS), which determines the importance level of the point according to the neighbor points and divides the sequences on different important levels. This method keeps the key points of information on different layers and can fully capture and express the trends and patterns of time series.(2)For the financial time series similarity measures, this paper proposes a Hierarchical Dynamic Time Warping similarity measurement method (HDTW), time series are segmented at different levels using MEPS algorithm and then computed the corresponding level of similarity using the DTW algorithm. And on this basis, this paper proposes the further improved algorithm (IHDTW), which can improve the accuracy and efficiency of the similarity measurement. Finally, Customer favorites are also considered during the process, event-based time series similarity measures (SMBE) is designed to make the result of similarity measures meet the customer requirement better.(3) Multiple time series, cross-transaction association rules mining has very important significance for predicting the trend of financial time series. This paper proposes an optimized O-Apriori algorithm, which defines the frequency status matrix to store frequent status of item sets and the processing of seeking frequent item sets is reduced according to the definition of cross-transaction association rules, the efficiency of the algorithm is greatly improved. Then this paper proposes the variable-based support of the O-Apriori algorithm (VSO-Apriori), which sets the changeable minimum support threshold for corresponding to different levels of frequent item sets and more useful rules are mined. The paper also creates a dynamic association rules mining algorithm, which can mine the frequent item sets online and track and represent the changing trend of the frequent item sets in different mining windows.(4) Clustering analysis also plays an important role in financial time series mining, which normally prepares initial analysis for different data mining tasks. This paper proposes a clustering algorithm based on IHDTW, which builds the similarity count matrix using hared nearest neighbor similarity (SNN) to find the cluster center sequences and greatly enhances the clustering effects. Then a hierarchical clustering based on SMBE algorithm is proposed for clustering the event similarity meeting the users'needs. The algorithm uses two parameters to judge the distance between classes and greatly enhances the clustering effect. Finally, a real-time for multiple time-series data stream clustering algorithm is proposed, which uses the important points as the feature of the subsequence and designs the dynamic sliding window to ensure the synchronization between multiple data streams. The algorithm can achieve the clustering results at any time and track the real-time evolution process of the clustering.(5)This paper proposes an integrated financial time series forecasting methods, which synthesizes the association rules mining algorithm and clustering algorithm introduced before and validate the method using the actual Shanghai and Shenzhen A trading market's data. The method can predict the stock price change within 3 trading days and the stock tendency change within 60 trading days.