Stock Market Prediction with Machine Learning: A Regime-Aware Hybrid Transformer Framework

Stock market prediction is uniquely challenging due to the interplay of price trends, fundamentals, sentiment, and volatile regime switching. Existing models such as knowledgedriven event embeddings (66.93% accuracy) and numerical attention mechanisms (6.96% MSE reduction) often fail to adapt to abrupt regime changes. We propose a hybrid framework emphasizing interpretability and phase-adaptability. It integrates a Tab Transformer for fundamentals, domain-tuned FinBERT sentiment encoders, and Temporal Convolutional Networks (TCN) with attention. A dynamic attention fusion layer enhances stability, building on prior risk-aware TCNs (Price MAE of 1.23). We review recent seminal works, detail the methodology, and establish future research directions.