In this work I develop a novel framework for stock market simulation.It incor-porates support vector classification for prediction of stock price movement into the classical Markowitz mean-variance framework.Support vector classification uses various technical indicators to predict stock price direction and then the predictions are used to select stocks for construction of an efficient portfolio.In addition,I propose a novel heuristic algorithm to solve the efficient portfolio opti-mization problem with assets' weights upper bounds constraints for a given level of investor's risk preference.Then I apply these techniques to a data set of 247 companies listed on Shanghai stock exchange in the period 2005—2015.Experimental simulations consist of two parts.The first part examines how different choices of parameters for efficient portfolio optimizer affect trading strategies' performance,while the second part examines the effect how different parameters of machine learning algorithm affect performance of trading strategies.Results show that support vector machine classification is generally able to outperform the benchmark when the markets are relatively stable.Machine learning is not able to beat the market during periods of high volatility such as turbulent times of 2008—2009 and 2013—2015.In all cases,it is able to provide lower level of risk for a given level of return as compared to the benchmark.Overall,the new framework allows to minimize risk while providing returns on par with the benchmark.There also exist some market inefficiencies that allow to outperform the market in an ideal scenario with no transaction costs.However,in presence of realistic transaction costs returns are not superior to the benchmark while the level of risk is lower. |