Deep Learning Approach for Short-Term Forecasting Trend Movement of Stock Indeces

This paper is discussed the problems of the short-term forecasting of trend movement stock indices using supervised Machine (Deep) Learning approach. For this purpose, we used Recurrent Neural Networks (RNNs) based on cells of Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNNs), and stacking CNN- RNN. The main advantage of proposed approach is that CNN allows to automatically extracting features and hidden patterns from the data, which are pass to the input of the block LSTM for making prediction. For evaluation models we used daily data of five stock indices from 01/01/2015 to 1/06/2021. To assess the forecasting performance, we used the confusion matrix and Accuracy metrics and compared our results with Multilayer Perceptron (MLP) as a baseline. According to obtained results, stacking CNN-RNN outperformed MLP for all time series, but the LSTM model shown better results in flat conditions.