A Deep Learning-Based Approach for Stock Price Prediction Using Bidirectional Gated Recurrent Unit and Bidirectional Long Short Term Memory Model

Abstract

Stock market investments make up a significant portion of any country’s growing economy. The rise or decline in the share price has a significant impact on the investor’s profit. Accurate prediction of the dynamic and volatile stock price movements can enhance the confidence of investors in the share market and hence can increase the volume of investment. Due to the chaotic, non-linear, and complex nature of the share price pattern, traditional time-series analysis exhibits poor performance. In this paper, we propose a new deep learning-based forecasting framework, Bidirectional Gated Recurrent Unit (BiGRU) incorporated with an external activation layer that utilizes both the forward and backward propagation feature of the Recurrent Neural Network (RNN). We compare our proposed model with the prevalent bidirectional RNN model, known as Bidirectional Long Short Term Memory (BiLSTM). We implemented both models on three different datasets collected from the NIFTY-50 index. We considered five different evaluation metrics to evaluate the model performance. For the different combinations of hidden layers and data, our model shows better performance in comparison to BiLSTM. The BiGRU model also shows better stability in comparison to BiLSTM, with smaller trainable parameters. It can also accurately forecast stock price for a greater prediction window, in comparison to BiLSTM. Our model can also predict all the sudden spikes on 1000 days ahead prediction precisely.