Comparative Analysis of Predictive Power between Neural Networks and Econometric Models on Basis of Changing in Stock Prices

This approach deals with the construction of artificial neural network in Matlab and ARIMA models in R to predict values of stock prices and to compare predictive power of these two source of prediction. The paper presents a theoretical basis for understanding artificial neural networks. Also a wide list of literature on the topic of predictions using neural networks is given and comparison with classical models. This paper analyzes the predictive power of a neural network and ARIMA models on a sample containing the prices of 20 large Russian companies that are used in calculating the RTS index. The sample was about 12 months. This sample was divided into 2 main groups, the first training group consists of 9 months, on the basis of this group the network training will be conducted, and the second group 3 months, on which the neural network test will be conducted. The first group from 01/05/2016 to 31/01/2017, and the second from 01/02/2017 to 02/05/2017. ARIMA models use only second group. The methodology of constructing an artificial neural network is described in detail as well as methodology of constructive ARIMA models in R. The created artificial neural network is a multi-layer perceptron with one hidden layer of neurons. When learning a neural network, the algorithm for back propagation of the error is used. Due to the fact that 20 large Russian companies were selected (the list presented in approach), prediction charts for only one company (Aeroflot) will be given as an example. In addition, the MAPE(Mean absolute error in percent) indicators will be presented for all companies in the sample. The mean absolute error rate in percent for all companies was approximately 3.8 percent for neural network and 6,3 percent for ARIMA models. These values confirm the consistency of our artificial neural network and proof that ARIMA models show poor performance in comparison with neural network.