Finding Predictive Patterns in Historical Stock Data Using Self-Organizing Maps and Particle Swarm Optimization

Abstract

The purpose of this thesis is to find a new methodology for finding predictive patterns in candlestick charts without any predefining of how these might look. An algorithm combining particle swarm optimization and self-organizing map has been implemented and evaluated. Non-transformed daily open, high, low and close data has been used as input. The algorithm found predictive patterns that statistically significant outperformed random trading. Moreover, interesting properties such as the optimal length of the pattern, target length and similarity of input to found pattern are discussed.