Weight-Dropped Long Short Term Memory Network for Stock Prediction with Integrated Historical and Textual Data.

Thayogo; Antoni Wibowo

Investors and traders need an accurate stock prediction model to help them make decisions. They can use deep learning models such as Long Short-Term Memory Network (LSTM). However, a weakness of LSTM is that it tends to overfit to the training data and have unstable results. To overcome this weakness, this paper proposes using Averaged Stochastic Gradient Descent and Weight-Dropping on an LSTM network (AWD-LSTM). The proposed model regularizes the network by weight-dropping with DropConnect and optimizes the training process using a Non-Monotically Triggered Averaged Stochastic Gradient Descent (NT-ASGD). Additionally, this paper tested with integrating historical data with textual data which was shown to be valuable by other studies. This paper evaluated six variants of the model with different regulizers, optimizers, and data. The results show that (1) DropConnect regulizer performed better than DropOut or No Drop; (2) Adam optimizer is better for stock prediction than NT-ASGD; (3) Adding textual data slightly increases performance; (4) The models were able to gain consistent profits in a market simulation; (5) A variant of the model outperformed a previous study in 4 out of 5 indexes.

LONG-term memory; SHORT-term memory; FORECASTING; INVESTOR confidence; DEEP learning; PREDICTION models

IAENG International Journal of Computer Science, 2020, Vol 47, Issue 3, p96

1819-656X

Academic Journal