Investigating the performance of two stock price forecasting models, based on long short-term memory neural network and with two different approaches of feature selection and time series analysis

In the financial literature, the capital market plays an essential role in economic growth through the financing of enterprises, the optimal allocation of resources, the improvement of liquidity of assets, the improvement of company management and the increase of transparency in the economy. One of the most important challenges that shareholders always face in the market is to make the right decision and be in the right position with buying and selling stocks. Stock forecasting that predicts future stock movements and benefits shareholders has been an attractive research area for financial studies and researchers since the past. In this research, it presents two models of neural networks that receive inputs and predict price and movement trends, and finally, the performance of these two models  is compared .The data studied in this research includes the price data of the 5 largest shares of the New York market during the years 2000 to 2020, , in which 80 percent of the initial samples are used as training data and the remaining 20 percent are used as test data.

In the proposed VMD-LSTM model, first, the stock price time series is decomposed using the variational mode decomposition algorithm (VMD) to the intrinsic mode functions (IMF), and then each of these IMFs is predicted by the LSTM model and After interpreting the results. In the second proposed method, available features including price and some of the most important technical indicators are used to predict stock prices. In the GA-LSTM model, genetic algorithm is first used to select the best features from the entire set of features. Then, the time series of the stock closing price was predicted by LSTM network using the selected features. The results of the research showed that because both models are very good in price prediction, the proposed GA-LSTM model, which is developed based on feature selection, has less error and more accurate performance.