The Research On Hidden Pattern Mining Of Financial Time Series & Its Application

Financial time series is always studied as a whole in existent financial econometrics based on statistical models, some rigorous hypothesis are often indispensable. Especially, it is unbefitting for importuning all data of series to meet a fixed mathematics model. Hence, they are frequently failed to deal with some realistic problems. This dissertation indicates a new idea, which acquires features of financial series not based on statistical model but some hidden local patterns. Time series data mining is a hotspot in many fields and some exciting achievements has been made recently. Whereas, how to effectively discover patterns from financial time series is still a challenging problem. This dissertation aims at several key problems on mining hidden patterns from time series and their application in finance. Some new ideas and methods are proposed. The main work and novel parts of this dissertation are:(1)A feasibility analysis of mining predictive hidden patterns from financial time series is made. Though the validity of technical analysis is doubtable by lack of solid theory, many people adopt it enthusiastically. We can not simply explain it as "No alternative". Actually, efficient market is also doubtable. Since we can't negate the predictability of history data, the idea of mining predictive hidden patterns from financial time series is triable.(2)Taking into account the specialties of financial time series, we select wavelet transform method to de-noise origin series to improve the effect of mining. Some problems about how to determine several key parameters such as wavelet function, threshold rule and decomposition level are discussed.(3)Measuring similarity of different time series is a crucial problem when analyzing or mining time series. Existent methods do not consider individual subjective preference about similarity, whereas it may play a significant role while converting experiences of investor into rules. This dissertation proposes a new similarity model, which measures similarity from three aspects: offset, amplitude and correlation. To get the model parameters, we construct a non-linear optimization model via visual operations. Solving this non-linear optimization problem using genetic algorithms, we can get a similarity measure model which is consistent with individual interest.(4)To get some predictive local patterns from financial time series, a new pattern mining method-TSEOPM (Time Series Event Omen Pattern Mining) is broughtforward and interrelated concepts, principle and procedure are introduced in detail. Simulation results indicate this mining method is validity to discover event omen patterns from whether regular or random time series.(5)The association features of different time series are also interesting. This dissertation provides a new algorithm for association patterns discovery, which is based on the idea of common mechanism Association patterns mining is made up of three procedures: linear segmentation, clustering and association patterns searching, all are introduced in detail. To show its validity, two experiments based on simulative data are introduced.(6)To show the practical utilities of the above-mentioned methods, three applications on actual stock market data are illustrated. Firstly, TSEOPM is applied to Shenzhen and Shanghai stock markets analysis. The result obviously reveals the stock markets of China is not weak efficient. So technical analysis is feasible, Omen patterns discovered by TSEOPM are predictive, which offers a new way of technical analysis. Secondly, association analysis on different stocks is illustrated, which shows two unrelated series may exist interior correlation. Moreover, utilizing some technologies in this dissertation, a technical auto-trade decision-making system is designed, which can timely send out trade information. Especially it can meet investors the requirements of individuation and self-help. Hopefully, it may avail to exalting their experiences into rules.The work of this dissertation is supported by National Nature Science Foundation of...