Dedy Hartama, Mesran, Agus Perdana Windarto, Putrama Alkhairi, Weni Rosdiana
This study aims to improve the performance of stock market forecasting models by conducting a comparative analysis of CNN architectures using two primary activation functions: Rectified Linear Unit (ReLU) and Sigmoid. By evaluating metrics such as Mean Squared Error (MSE) and Mean Absolute Error (MAE) across several models, including IDCNN, LSTM, DCN, ResNet, as well as the Ensemble IDCNN + LSTM and Proposed Ensemble models, it was found that ReLU consistently outperforms Sigmoid. The results show that models utilizing ReLU, particularly the Ensemble IDCNN + LSTM, achieve the lowest error rates (MSE: 0.0023 and MAE: 0.0361), demonstrating its ability to capture complex nonlinear patterns in stock market data. In contrast, models using Sigmoid exhibited higher error rates, indicating that Sigmoid is less capable of handling the generalization challenges in volatile financial data. This study provides important insights into the impact of activation functions on deep learning model performance and recommends ReLU as the primary activation function for stock market forecasting tasks. With ReLU, models can deliver more accurate and reliable predictions in the dynamic stock market environment. © (2025), (International Association of Engineers). All rights reserved.
STIKOM Tunas Bangsa, North Sumatra, Pematangsiantar, Indonesia; Sekolah Tinggi Ilmu Manajemen Sukma, North Sumatra, Medan City, Indonesia; Public Administration, Universitas Negeri Surabaya, East Java, Indonesia
