LSTM NETWORK WITH REINFORCED LEARNING IN SHORT AND MEDIUM TERM WARSAW STOCK MARKET INDEX FORECAST

摘要

The aim was to explore the possibility of using deep learning methods to predict and forecast the value of selected stock market indicators based on historical data in a given future horizon. The study takes into account the daily closing values for the period from 4.01.2013 to 20.03.2020 of the WIG, WIG20 and WIG30 indices and COVID19 flash crach influence has also been investigated. The LSTM network is being developed in order to forecast future index values based on historical data. The reinforced learning technique was used to see if it improves prediction results relative to the classic deep learning technique. Results indicate good prediction of the index value in the 10-day horizon, which translates into two weeks of quotations in the front (the final wave of index index collapse deviates from the general nature of the last 500 days, the refraction pattern is missing in the sample, which translates into erroneous predictions of later values). Also, observation update (reinforced learning) improved the prediction result. In addition, the prediction remains quite accurate over the horizon of 25 future quotations, i.e. five weeks.