Stock Prediction with Stacked-LSTM Neural Networks

This paper explores a stacked long-teen and short-term memory (LSTM) model for non-stationary financial time series in stock price prediction. The proposed LSTM is designed to overcome gradient explosion, gradient vanishing, and save long-term memory. Firstly, build time series with different days for network input, and then add early-stopping, rectified linear units (Relu) activation function to avoid over-fitting during the training stage. Finally, save trained parameters state and new batch size for testing. The results suggest that the developed stacked LSTM produces better predictive power and generalization.