PERAMALAN TIME SERIES DARI PERGERAKAN HARGA SAHAM DENGAN RECURRENT NEURAL NETWORK = TIME SERIES FORECASTING OF STOCK PRICE MOVEMENT WITH RECURRENT NEURAL NETWORK

Background. Precise and stable stock price predictions are vital for supporting informed investment decisions, especially given the dynamic and nonlinear nature of financial markets. Objective. This study aims to develop a deep learning-based forecasting framework by combining architectural design and metaheuristic optimization strategies. The method involves creating a stacked LSTM structure with a parallel multivariable input system and building composite models from different types of Recurrent Neural Networks (RNN), specifically Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Simple Recurrent Unit (SRU), enhanced with dual-dropout mechanisms and feature segmentation. Methods. Hyperparameter tuning was performed using Random Search (RS) and the Grey Wolf Optimizer (GWO), with experiments conducted on stock price data from Google, Apple, Microsoft, and the Hang Seng Index (HSI). The model performance was assessed using the R-squared, RMSE, MAPE, RMSPE, WI, NSE, and PBIAS metrics. Results. The findings indicate that the parallel input architecture improved the accuracy by up to 12% and reduced the PBIAS by over 50%. The LSTM-GWO model with a 1-1-0-1 configuration provided the best overall prediction results, followed by the computationally efficient GRU-GWO and SRU-GWO, which were less stable but faster. The GWO strategy consistently outperformed the RS in generating more accurate and stable model configurations. Conclusion. These results demonstrate that combining architectural design with metaheuristic optimization significantly enhances the effectiveness of multivariable RNN-based forecasting systems in the financial sector.

Keyword : Deep learning; Modular architecture; Multivariate data; GRU; Grey Wolf Optimizer; LSTM; Random Search; RNN; SRU; Stock price prediction.