RegimeWorld: A Calibrated Multiscale World Model for Long-Horizon Financial Market Simulation Under Distribution Shift

Financial markets are a canonical example of an irregular, multi-scale, regime-switching temporal system: volatility clusters, tail risk is state-dependent, and the joint distribution of returns shifts abruptly across calm, trending, and crisis periods. Most deep temporal forecasters optimize point accuracy on stationary benchmarks and are evaluated with metrics that do not audit calibration under shift. We present RegimeWorld, a lightweight world model that couples (i) a Gaussian hidden Markov regime filter, (ii) an Interacting Multiple Model (IMM) Kalman filter that mixes regime-conditioned local-level dynamics across a macro (daily) and micro (intraday) time scale, and (iii) an Adaptive Conformal Inference (ACI) layer that recalibrates the model's predictive intervals online as regimes shift. We build an open, fully synthetic multiscale regime-switching market simulator that reproduces Cont's stylized facts (heavy tails, volatility clustering, leverage effect) with a known, auditable data-generating process, and use it to stress-test calibration under distribution shift without any risk of benchmark contamination. RegimeWorld's raw predictive intervals overcover on average (coverage 0.97 vs. a nominal 0.90) while single-regime and i.i.d. Gaussian baselines undercover (0.86–0.88); wrapping RegimeWorld in ACI recovers near-nominal coverage (0.895) and tracks the target through a held-out crisis episode where the uncalibrated baselines' coverage collapses. We report all metrics, negative findings (imperfect regime identifiability between calm and trending states), and release the simulator, model, and evaluation code.