A statistical learning-based prediction system of korea composite stock price index

Abstract

This paper proposes a statistical learning-based prediction system which predicts the variation of Korea Composite Stock Price Index (KOSPI). The proposed system is based on the idea of technical analysis. A simple machine learning algorithm using two-dimensional matrix is used as an underlying technique. In the learning stage, Two Index-Volume Pair Appearance (IVPA) matrices are constructed with instances of peaks and bottoms, respectively. The Market Combination Appearance Probability (MCAP) matrix is also constructed by those two IVPA matrices. When the data of a certain day is given, the system can yield the peak probability using the MCAP matrix. By considering the peak probability, investors may have some decisional supports. In the experiments, KOSPI data from 1998 to 2003 is used as training set, and data from 2004 to 2007 is used as test set. The system shows some positive performances, especially for bottom predictions. Although more improvements are needed, the proposed system may be useful for investment of derivatives.