A hybrid system integrating a piecewise linear representation and a neural network for stock prediction

Abstract

Stock price prediction is a challenging problem in the machine learning area due to a great noise in the stock market. In this paper, we propose a novel stock prediction method based on a multilayer feedforward neural network. To reduce the effect of noisy trends, it employs the piecewise linear representation which transforms the original time series of the stock price and the trading volume into a set of time segments. The transformed information is served as the input variables in the neural network for prediction. We tested the proposed method annually from 2001 to 2009. It showed a good performance of about 55% accuracy on average in predicting the price direction. It was also successful in making more profit than the buy-and-hold trading strategy.