Predicting Next-Trading-Session Stock Movements from Arabic Corporate Announcements: A Hybrid Transformer-Boosted Ensemble

Abstract

Predicting next-day stock movements from corporate announcements is a complex task, especially in markets where financial text is published in low-resource languages such as Arabic. In this study, we develop a hybrid, multimodal forecasting framework designed to capture both the semantic depth of Arabic announcements and the quantitative structure of market behavior. The system combines three complementary components: A fine-tuned multilingual Transformer for extracting text representations, gradient-boosted trees for modeling financial features, and a temporal Transformer capable of learning sequential market patterns. Each model contributes regression and classification outputs, which are then integrated through a stacked ensemble. To mitigate noise from small price changes, the framework employs a gray-zone weighting scheme that reduces the impact of near-zero returns. Additionally, we calibrate predicted probabilities using isotonic regression to ensure stable confidence estimates and enable effective threshold tuning. We evaluate the proposed approach on more than 46,000 Arabic corporate announcements from the Saudi Stock Exchange (Tadawul). The ensemble consistently surpasses all individual models, achieving 0.868 AUC, 76.9% accuracy, and an F1-score of 0.834 for directional prediction, representing a substantial gain over standalone text-based, tabular, or temporal models. For magnitude forecasting, the ensemble reduces RMSE from 0.829 to 0.718. To our knowledge, this is the first large-scale study to fuse Arabic financial text, structured market data, and temporal patterns into a stacked ensemble model for event-driven forecasting. The results highlight how combining language signals with market indicators can improve prediction quality in Arabic-language markets, where reliable automated tools remain limited.