With competition of financial trade being increasingly serious, the dependence and sensitivity on data when making decisions are higher. Under such backgrounds, many financial institutions begin to use more advanced information technology and intelligent decision support technology to make deep analysis on huge data accumulated by business systems in order to find various kinds of valuable rules. As a new kind of intelligent decision support technology, data mining is being used in financial trade step by step.In financial field time seires is very important and time series mining is an important task of financial data mining. Based on the analysis of data features and actual demands, we make a research on several key issues in financial time series mining in this paper. Such research is important for trend analysis and trend forecast, modeling series more accurate, risk analysis, assets pricing and portfolio research. Generally speaking, the main contents of this paper include:(1) Based on trend and change of trend, a method for dimension reduction and similarity search based on regression coefficient is proposed. It is proved firstly that for any time series the regression coefficient can be computed in lianear time and only use constant space. Then a new dimension reduction method for time series is brought forward which called Piecewise Regression Approximation (PRA). The PRA method is of linear time complexity and is not sensitive to independent noises with stationary means. It is proved the similarity search based on PRA method satisfied the lower bounding lemma, so it is practical and effective. The experiments on real data show we can conduct similarity search based on trend and change of trend by using PRA, and it is very important for data analysis in foreign exchange and futures markets.(2) Based on the features of financial time series of auto correlation and dynamic changing of data generation process, an on-line segmenting algorithm based on ARMA is brought forward. According to the characteristics of least variance forecast of ARMA, a new index called fitness is proposed for judging whether current ARMA model can be used to describe following data. Based on the index an on-line algorithm for segmenting time series is brought forward. This algorithm can judge whether current ARMA model is suitable for following data by computing fitness, and then adjust the ARMA model dynamic or segment the series and