A deep learning framework for forecasting complex, non-stationary time series using attention-based models such as Transformers, Temporal Fusion Transformer, and Informer. Includes multivariate data generation, windowing, hyperparameter tuning, and advanced metrics.