Rough Clustering Of Financial Time Series In Data Mining

Based on strict mathematical conduction and then to conduct parameters estimation and inference, traditional statistics and modern financial econometrics, in which theory frameworks have been built up for years, are to establish statistical models. However, such methods seem unfit due to its dependence on strict hypothesis and importuning all data of series to meet modeling requirements. Data mining techniques overcome this kind of shortage in a way of establishing models motivated by data.Time series data mining is popular today, and many achievements have been made. Whereas, appropriate solution of measuring similarity still lacks of attention, which lays the foundation of several methods in series mining. Apparently, similarity measurement in time series does affect mining results. This dissertation aims at such pivotal issue as well as its applications in series mining, particularly, clustering analysis. Instead of hard clustering, this dissertation introduces a soft clustering method——Rough Clustering method, which can reflect the practicability of the new method on measuring similarity of time series. Main works and innovations of this dissertation are summarized as:Firstly, a method to measure similarity of time series based on multi-scale wavelet transformation is presented with the idea of wavelets analysis. And financial time series cases study is also conducted to show that this method considers all the factors affecting the measuring similarity of series and effectively overcomes the shortage of existent methods that fail to balance between outline and detail differences of series.Secondly, discusses rough clustering of sequences and shows its advantages through financial cases study. Furthermore, analysis on three issues as follow is considered: (1) to discuss the impact of threshold parameters on clustering results by establishing the quality indicators for rough clustering; (2) to integrate the rough clustering and hierarchical clustering so that we can make most of their advantages; (3) to transfer soft clustering into hard clustering, to condense the results of rough clustering by the iteratively-removed-method, and to show its feasibility by comparing with original results.Finally, we also discuss the algorithms used in these methods, and share programming code in form of Matlab. Results from empirical research are convincible.