Trend forecasting could be one of the most challenging things in stock market analysis, as the data associated with stock is the time series data characterized in high intensity, noise, uncertainty, etc. In addition, beyond the time series data, from a macroscopic point of view, the influences of the external environment, such as national economy, technical progress, legal and political events, social and demographic factors, even the natural environment, are significant as well regarding the future trend forecasting.

The literatures show the successes of the machine learning methods in stock analysis, but at the same time, the literatures also show that, from the historical data of a stock observed simply extracting some rules that the stock follows, and thereby forecasting the stock movements within a certain future time period, is rather difficult and exhibits lower accuracy. However, while studying a stock with its graphical representation (zigzag curve), the researchers often found the similarities of the stock movements. Thus, a geometric method for extracting the movement similarity from the past data is developed this this research, the key improvement of the correlation extraction method proposed is the graphical trend similarity approximation, using an arc to represent the trend of a portion of the whole time series, the evaluation of the trend similarities is turned to calculate the distances of the two arcs of the two time series to the arc of the time series observed. This means, the new method extracts correlations from general variation trends, avoids extracting correlation form the time series directly, the mismatch problem is thus solved.

To verify the proposed method, based on SVR an experiment is conducted with the real world stock data. Based on the analysis results, the paper concludes that our method improves the performance of stock price prediction. Some future work, which is capable of promoting our method, has been discussed in the thesis.