<![CDATA[Accurate stock price forecasting remains challenging because financial time-series exhibit complex local patterns, long-term temporal dependencies, and rapidly changing market dynamics. Although hybrid deep learning models have demonstrated promising predictive capabilities, limited evidence is available on how different architectural integration strategies influence forecasting performance. This study compares sequential and parallel hybrid CNN–LSTM–Transformer architectures enhanced with attention mechanisms and Bayesian optimization for predicting BBNI stock prices. Historical market data were processed through convolutional, recurrent, and attention-based learning components, while Bayesian optimization was employed to identify optimal hyperparameter configurations. Both hybrid architectures consistently outperformed the baseline CNN–LSTM model, demonstrating substantial improvements in prediction accuracy. The parallel architecture achieved superior performance in minimizing large prediction errors and explaining variance, whereas the sequential architecture produced lower absolute prediction errors with greater computational efficiency. These findings indicate that predictive performance depends not only on model complexity but also on the way deep learning components are structurally integrated. The study provides practical guidance for designing hybrid forecasting architectures that balance predictive accuracy and computational efficiency for financial time-series prediction in emerging markets.]]>
Copyrights © 2026
