Prediction of the stock exchange of Thailand index using an evolutionary algorithm portfolio

Abstract

This paper proposes an evolutionary algorithm portfolio aiming to decrease errors of long-term prediction of the Stock Exchange of Thailand (SET) index. We adopt the formula proposed by P. Sutheebanjard and W. Premchaiswadi (2009), which is able to forecast the SET index during 2005 – 2009 with mean absolute percentage error (MAPE) less than 2%. However, the MAPE becomes relatively high when the formula is used to predict the SET index of 2010 and 2011 (i.e., 4.04 and 6.26, respectively). To modify the prediction function to fit better to the recent and future SET index, this paper proposes collaborative learning of evolutionary algorithms (consisting of a genetic algorithm, an evolution strategy and a differential evolution) to reduce the prediction errors. The proposed method forms a portfolio of two or three algorithms and allocates a certain amount of computational time to each of the constituent algorithms. We encourage consistent interactions among the algorithms so that the algorithms, which produce good results in certain years, share the knowledge with the algorithms that yield inferior solutions. The experiments show that using combination of the genetic algorithm and the differential evolution yields better prediction both for the SET index between 2005 and 2011 (i.e., MAPE equal to 1.11) and for the SET index of every year (i.e., MAPE less than 1.5).