Studies On Segmentation Methods And Turning Points Identification Of Multivariate Time Series

With the coming of data era,data size and data complexity are continuously increasing.Comparing with univariate time series,it is more difficult to analyze multivariate time series which has more complicated structure,while multivariate time series can supply more meaning-ful information for decision-makers.Therefore,studies on multivariate time series have received more and more attentions.Detection and analysis of the structural variation of multivariate time series on both temporal and dimensional sides is a widely studied theme,and for instance,re-search on multivariate time series segmentation,research on turning points identification of busi-ness cycles in macroeconomic field and research on contrarian strategy based on turning points in financial market all can be distributed to this theme.Based on the previous research work on structural variation of multivariate time series,this thesis obtains some new research ideas and results.The main contributions of this thesis include the following aspects:(1)At present,the theoretical system of univariate time series segmentation is relatively well-developed.In contrast,there are fewer segmentation methods to satisfy the demand for multivariate time series analysis because of the more complicated modelling and computation of multivariate time series segmentation.It has become a common practice to develop more segmentation methods for multivariate time series by extending segmentation methods of uni-variate time series.But on the contrary,this thesis tries to segment multivariate time series by transforming them into a univariate common factor sequence to adapt to the segmentation meth-ods of univariate time series.First,a common factor sequence is extracted from the multivariate time series as a composite index by a dynamic factor model.Then,three typical search methods including binary segmentation,segment neighborhoods and the pruned exact linear time are ap-plied to the common factor sequence to detect the change points and the segmentation results are considered as the final segmentation results of multivariate time series.The case studies validate the applicability and robustness of the proposed approach in multivariate hydrometeorological time series segmentation.(2)Composite indicators of business cycles play a paramount role in the analysis of macroe-conomy,which are helpful for decision-makers to assess economic conditions and make further policies.This thesis develops a novel constructing method of the business cycle composite in-dicator based on information granulation and dynamic time warping,which can not only take the most important indicator real Gross Domestic Product(GDP)into account but avoid the complex estimation process of dynamic factor models on data with different frequencies.First,the quarterly real GDP sequence is divided into information granules by the principle of jus-tifiable granularity.Next,monthly coincident indicators are split into corresponding segments based on the information granules of real GDP,and dynamic time warping is applied to measure the simil:arity between monthly segments and quarterly segments.The weights are derived by normalizing reciprocals of the above distance values.Finally,the monthly composite indicator of business cycles is obtained by taking a weighted cross-section average of those monthly co-incident indicators.The numerical experiment reveals that the composite indicator established by the proposed method can reflect the dynamics of business cycles and accurately identity the locations of turning points in business cycles.(3)As one kind of typical time series,futures price time series are usually highly fluctuant and difficult to predict.Traditional time series models including auto-regressive(AR)model and auto-regressive moving average(ARMA)model cannot effectively predict the futures price in future days.In view of this,this thesis establishes a set of trading strategies for futures automatic trading by use of Bollinger Bands technical indicators and association analysis based on turning points analysis.First,we build Bollinger Bands technical indicators according to futures price time series,and then use fuzzy methods to extract the association rules between Bollinger Bands indexes and price change indexes,which have sound semantic interpretation.Then,on the assumption that history can be repeated,association rules based trading strategies are built for program trading on trading software Tradeblazer.Finally,trading results show that the extracted rules can accurately predict the turning points where inversion of futures price trend happens,which can guide traders to earn profits by trading appropriately in futures markets.