DEHypGpOls: A Genetic Programming with Evolutionary Hyper-Parameter Optimization and its Application for Stock Market Trend Prediction

Abstract

Stock markets are one of the most popular financial markets since they can bring high revenues to their investors. In order to reduce the risk factor for investors, intelligent and automated stock market trend prediction tools are developed by using computational intelligence methods. This study presents a hyper-parameter optimal Genetic Programming (GP) model generation framework for a day-ahead prediction of stock market index trends. In order to obtain the best trend prediction model from the stock market data, a Differential Evolution (DE) algorithm is implemented to optimize hyper-parameters of a GP algorithm with Orthogonal Least Square (GpOls). Evolutions of GpOls agents within the hyper-parameter search space allow adaptation of GpOls algorithm for a given training dataset by enabling optimal autotuning of user-defined parameters. Authors demonstrated that this structure can significantly improve modeling performances of GpOls algorithms for data-driven modeling applications. In the study, the proposed DE based hyper GpOls (DEHypGpOls) algorithm is implemented for a day-ahead forecast problem of the Istanbul Stock Exchange 100 Index (ISE100) trends. In the experimental study, real trend data from ISE100 and seven other international stock markets are used to extract a day-ahead forecast model of daily ISE100 trends. The forecast model of the DEHypGpOls algorithm can provide 68% prediction accuracy for the test dataset. In ISE100 index market simulations, daily exchanges of ISE100 index according to buy or sell signal of the forecast model of DEHypGpOls provided about 273% more income compared to the income of a long-term investment strategy.