Transformer-based model for accurate and scalable panel time series forecasting. Builds on Temporal Fusion Transformer with innovations like segment-wise attention, multi-scale decomposition, and cross-entity attention.