TST: An Effective Approach to Extract Trend Feature in Stock Time Series

Abstract

Trend representation of stock time series in linear fashion is a fundamental task for the neural network to learn the trend features from stock price fluctuations. Since the threshold error as initial parameters of Piecewise Linear Approximation is designated based on experience, in order to effectively extract the trend information from the subsequence of time series of stock price which is suitable for training, we have proposed a systematic approach combining Trend Manual Selection, Segmentation Algorithm and Training Error Feedback as a series of techniques (TST). Firstly, the ideal turning points in stock time series are manually selected by expert knowledge and experience, then the trend feature representation is generated via segmentation algorithm utilizing PLA and then using RNN to extract the trend information from prepared representation. The results show that this approach will assist segmentation algorithms to find a reasonable value of threshold error with ideal trend features. Through applying TST, it can be targeted to improve effectiveness of the piecewise linear representation of segmentation algorithms on stock time series, as well as to improve the overall ability of neural networks to identify the trend feature from selected time series of stock prices. Furthermore, the threshold error of PLA can be reevaluated by determining the minimum training error so as to obtain proper piecewise linear segmentation of trend representation in stock time series more suitable for learning.