Enhancing Stock Price Forecasting Accuracy Through Compositional Learning of Recurrent Architectures: A Multi-Variant RNN Approach

Abstract

Accurate stock price forecasting is essential for profitable and risk-aware decision-making in volatile financial markets. Accordingly, this study proposes a forecasting method that integrates a robust composition of deep learning algorithms applied to well-structured financial data to enhance the predictive accuracy. A compositional learning approach was developed to systematically design and assess Recurrent Neural Network (RNN) architectures, specifically Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Simple Recurrent Unit (SRU), for effective multivariate time-series prediction. A total of 54 models (18 for each type) were created and tested, with hyperparameter tuning using both Random Search (RS) and Grey Wolf Optimizer (GWO). The results show that LSTM-GWO (1-1-0-1) performed best, with R-squared = 99.2427%, MAPE = 1.1721%, RMSPE = 1.6221%, RMSE = 339.3902, WI = 0.9981, NSE = 0.9924, and PBIAS = 0.0523, demonstrating high accuracy, low error, excellent model agreement, efficiency, and minimal bias. The GRU-GWO (2-1-1-1) and SRU-GWO (2-2-0-0) also showed improved results over RS optimization, confirming that GWO consistently boosts model accuracy and stability across all RNN variants. These findings confirm that a systematic architecture design combined with metaheuristic optimization significantly enhances stock forecasting performance. The proposed compositional RNN framework offers a reliable foundation for developing robust and accurate stock prediction systems that aid strategic investment decisions in dynamic and nonlinear financial markets.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint