A Comparative Study of Recurrent Neural Networks for Predicting Indian Bank Stocks

Deep Learning (DL) techniques have been shown to outperform traditional time series methods in prediction in recent years due to their ability to handle large datasets and long-term dependencies. Various Deep learning (DL) models, which include Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), to name a few, are gaining popularity, constantly being revised, and over time have demonstrated excellent performance in learning difficult time-series patterns. Furthermore, these learning methods have variants such as Stacked and Bidirectional models that address some of the limitations of the initial model. The goal of this analytic study is to compare the prediction outcome of various RNNs and validate the best model through a series of experiments conducted on two stocks obtained from India's National Stock Exchange (NSE) utilizing RNN and its variants, Bidirectional RNN (BRNN) and Stacked RNN (SRNN), LSTM and its variants, Bidirectional LSTM (BLSTM) and Stacked LSTM (SLSTM), and GRU with its variant network architectures, Bidirectional GRU (BGRU) and Stacked GRU (SGRU). The efficiency of these models was examined over three different sliding window sizes: three days, seven days, and thirty days. The results show that the BGRU significantly outperforms the other variants, and it is evident that the models performed better when using the 7-day sliding window size. The Wilcoxon signed rank test method validated the best models' results.