Compositional RNN Approach to Accurate Stock Price Forecasting

Abstract

Proper stock price forecasting is crucial for making decisions during volatile and uncertain financial conditions. In this regard, the current paper provides a compositional deep learning approach for time-series multivariate forecasting by employing three types of Recurrent Neural Networks (RNNs): Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Simple Recurrent Unit (SRU). For further enhancement of the predictive ability of the models, the researchers employed the Grey Wolf Optimizer (GWO) and Random Search (RS) methods to fine-tune and optimize models' hyperparameters, leading to the evaluation of 54 model configurations. Among all considered architectures, the LSTM model optimized with GWO demonstrated the highest forecasting efficiency in terms of accuracy, minimized error rate, perfect consistency between predicted and actual stock prices, and low level of bias in stock price forecasting. Besides, the GRU-GWO model and SRU-GWO model outperformed their RS-based counterparts in terms of enhanced stability and reliability. Overall, the results proved that the use of compositional RNN architecture along with metaheuristic-based fine-tuning improved forecasting performance.