A new attention-based LSTM model for closing stock price prediction

Abstract

Financial time-series prediction has been a demanding and popular subject in many fields. Latest progress in the deep learning technique, especially the deep neural network, shows great potentials in accomplishing this difficult task. This study explores the possible neural networks to improve the accuracy of the financial time-series prediction, while the main focus is to predict the closing price for next trading day. In this paper, we propose a new attention-based LSTM model (AT-LSTM) by combining the Long Short-Term Memory (LSTM) networks with the attention mechanism. Six stock markets indices with four features were used as the input to the model. We evaluate the model performance in terms of MSE, RMSE and MAE. The results for these three metrics are 0.4537, 0.6736 and 0.4858, respectively. The results suggest that our model is skillful in capturing financial time series, and the predictions are robust and stable. Furthermore, we compared our results with the previous work. As a result, our proposed AT-LSTM exhibits a significant performance improvement and outperforms other methods.