Enhancing Stock Price Forecasting Accuracy Using Compositional Rnn

Abstract

Predicting Stock Prices Accurately Is Essential For Making Well-informed Decisions In Erratic Financial Markets. This Paper Introduces A Compositional Deep Learning Framework For Multivariate Time-series Forecasting That Integrates Three RNN Variants: LSTM, GRU, And SRU. Grey Wolf Optimizer (GWO) And Random Search (RS) Were Used To Develop And Optimize A Total Of 54 Model Architectures. The Best Results Were Obtained By LSTM-GWO (1-1-0-1), With R2 = 99.2427%, MAPE = 1.1721%, RMSE = 339.3902, WI = 0.9981, NSE = 0.9924, And Minimal Bias (PBIAS =0.0523). Additionally, GRU-GWO And SRU-GWO Performed Better Than RS-based Models, Demonstrating The Efficacy Of Metaheuristic Optimization. The Results Show That GWO And Systematic Architectural Design Greatly Improve Forecasting Accuracy And Model Stability For Reliable Financial Prediction Systems.

Keywords