Why time series models struggle with mixed units—and how to fix it

Most time series models treat each variable independently or mix them in raw space, losing important relationships between temperature and pressure, stock price and volume, etc. Falcon-X solves this by translating all variables into a unified latent space, then uses a new attention mechanism that captures both positive reinforcement (when rising temperature boosts demand) and negative feedback (when high prices suppress it). The model transfers to new datasets without retraining. Code released.