FORECASTING STOCK PRICES USING THE RECURRENT NEURAL NETWORK LSTM

Abstract

Modern financial processes exhibit complex nonlinear structure and non-stationary behavior over short and long time intervals. Such behavior can be explained by influence of multiple external stochastic disturbances and complicated nature of the price-forming processes at stock exchange. The traditional approach to modeling and forecasting behavior of such processes is based upon application of regression analysis and respective mathematical models. However, this approach to modeling is not always successful due to complicated structure of the time series under study and limited possibilities of the regression models to describe adequately nonlinear and non-stationary behavior of the processes select ed for analysis. That is why it is more appropriate to use as alternative approach neural networks that exhibit very often superior characteristics of modeling and forecasting in comparison to regression models. This study was conducted to get familiar with the structure and principle of operation of the recurrent neural network LSTM (Long short-term memory) and analyze the possibility for its use to forecast the stock prices of one of the major world known technology companies. The theoretical material concerning recurrent neural networks and LSTM network was described in the paper in a volume providing the possibility for understanding basic principles of operation and practical application of the networks. Using as example the Apple stocks statistical data, the functioning of the chosen modeling method was demonstrated and the estimates of the quality of estimated forecasts such as RMSE, MAE, MAPE were calculated; and the accuracy of the forecast was estimated. The results of the simulation achieved showed that recurrent neural networks can be used to predict time series behavior and obtain the results with necessary high accuracy. Further research will suggest application of other types of neural networks, static and dynamic Bayesian networks, and evaluation of their performance on financial data exhibiting nonlinear and non-stationary behavior. Also it is planned to construct specialized decision support system based upon system analysis principles that would contain all necessary functions for further enhancement of modeling and forecasting results