When hidden regimes shape causal relationships: identifying switching systems

Many real systems—like glucose levels or financial markets—switch between distinct operating modes with different causal relationships. The challenge: can you reliably infer these hidden regimes and their causal structures from observed data? This work proves identifiability conditions for Markov Switching Models with nonlinear dynamics, instantaneous effects, and non-Gaussian noise, extending classical hidden Markov theory. They introduce FlowMSM, a framework pairing regime detection with causal discovery, and validate it on synthetic data and financial time series.