COSMOS: Dynamic Latent-State Discovery under Multivariate Drivers

Measuring how multivariate drivers affect multivariate outcomes in dynamic systems requires compact state representations that remain interpretable. Existing latent-state approaches often prioritize likelihood fit rather than explicitly controlling the geometry and interpretability of the learned states. We introduce COSMOS, a latent-state discovery framework that jointly learns representative outcome regimes and driver-conditioned transition models through an objective balancing regime coherence and transition consistency. We characterize when incorporating transition dynamics improves state distinguishability relative to emission-only clustering. Experiments on synthetic data and data from a large multinational retailer show that COSMOS improves latent-state recovery while producing balanced, interpretable regimes and transition effects that are suitable for downstream operational analysis. A newsvendor experiment further illustrates how the learned regimes can serve as a decision-support layer.