Forward forecast of stock prices using LSTM neural networks with statistical analysis of published messages

Abstract

The article presents a new approach that combines two separate fields of stock exchange analysis. The aim of proposed solution is to support investors in their decisions and recommend to buy the assets which provide the greatest profits. To achieve this goal, decisive algorithms have been developed using artificial neural networks and technical analysis, which were used along with statistics that refer to the occurrence of single words in the fundamental analysis. Based on this, a model was prepared that in response gives a recommendation for future increases. The system consists of two algorithms. The first of them uses the LSTM (Long Short-Term Memory) artificial neural network. As inputs, information about the current closing price as well as technical analysis indicators along with the value of the current volume were used. The output has been specified as the closing price on the following day. In order to improve the response from the ANN (Artificial Neural Network), statistics of the occurrence of words in publications from last week were used. Subsequent signals gained much more importance if the volume of all transactions was much larger than the moving average of the last 15 periods and if the words that appeared in the last publication caused earlier increases. Additional information for the system are also data that come from Google Trends. This allows to verify the trend of interest and whether the published messages are important.