Stock prices forecasting with artificial neural networks and portfolio optimization

In the modern Portfolio Theory, minimization of investment risks through diversification of investment tools is a widely adopted notion. Investors seek to obtain maximum profit against the risk they encounter. Investments inherently involve uncertainties and risks as they are dependent on future. Retroactive returns data have been widely used in portfolio optimization studies. The aim of this study was to analyze the performance of Markowitz's mean-variance model. Accordingly, monthly returns values of 26 stock exchange securities traded in ISE-30 within 2008-2014 period were used. Artificial Neural Networks (ANN) and Time-Series Analysis Methods were compared to find the best forecasting tool. Determination of the best parameter values of ANN and the most suitable time period for forecasting was aimed. Different time periods, different number of neurons, input variables and learning algorithms were used to build the models. MAE, MSE and MAPE error measurement methods were used to compare the results. Afterwards, Simple Average Method, Moving Averages Method, Exponential Smoothing and Trend Analysis Methods were used to obtain annual returns forecasting for 2014 on the basis of monthly average returns for 2008-2013, and the results were compared with ANN results. ANN and Exponential Smoothing methods yielded the most accurate results. These two methods displayed no certain advantage over each other. Finally, by use of ANN, Exponential Smoothing and real historical data, the best variance values were compared using real variance values of 2014. Accordingly, the best portfolio optimization outcome was obtained by combined use of ANN-Exponential Smoothing method and Markowitz's Mean-Variance model. Keywords: Portfolio Optimization, Forecasting, Artificial Neural Networks, Time-Series Analysis.