<p>Sentiment analysis plays quite a major role, particularly in understanding user feedback across various domains like business intelligence, where companies interpret consumer emotions to refine marketing strategies. However, the semantic complexity of the Arabic language, at both the sentence and word levels pose significant challenges for standard machine learning and deep learning algorithms, as well as traditional vectorization techniques. To address these limitations, this study proposes a novel hybrid deep learning model that integrates contextual embeddings from Arabic Efficiently Learning an Encoder that Classifies Token Replacements Accurately (AraELECTRA) with a stacked ensemble of Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Unit (BiGRU) networks, each enhanced by an attention mechanism to highlight the most relevant information generated by the hidden layers of BiLSTM and BiGRU. The probabilistic outputs of the base models are aggregated using a Multi-Layer Perceptron (MLP) as a meta-classifier. The proposed architecture is evaluated on three balanced Arabic sentiment datasets using multiple metrics, including accuracy, F1-score, precision, recall, ROC-AUC, memory usage, and training time. Experimental results confirm that the model outperforms baseline and state-of-the-art approaches, demonstrating robust generalization and strong performance across datasets of varying sizes. This architecture provides an effective framework for Arabic sentiment classification, especially in business intelligence applications.</p>