Fit a single distribution to a year of transit times on a busy trade lane and you will get something that fits nothing. The data is not one population. It is two or three: a calm regime where the lane behaves, a stressed regime where congestion feeds on itself, and occasionally a disrupted regime where the normal rules are suspended entirely. The lane does not drift between these states. It switches.
Regime-switching models take this seriously. Instead of one distribution, the model carries several, plus the probabilities of transitioning between them. The estimation problem becomes: which regime are we in right now, how long do regimes of this kind persist, and what does each regime imply for the decisions on the table?
Why the switch matters more than the noise
Within a regime, variance is manageable and mostly self-correcting. Across a regime change, everything an operator believes goes stale at once. Transit times, reliability, prices, and capacity availability move together, because they share a cause. A model that averages across regimes misses this correlation structure and therefore underestimates exactly the compound events that hurt: the week when the delay, the rate spike, and the capacity shortage all arrive together, because they were never independent.
Detection is therefore the operational prize. The early statistical signature of a regime shift, dwell times fattening in the tail, rate volatility clustering, schedule reliability decaying, is visible in the data days before it is visible in the P&L. A planning layer that watches for the signature can reposition while options are still cheap.
Two rulebooks, one network.
Tune how often the world flips and how long it stays flipped.
The long-run share follows from the two probabilities alone. The pain concentrates in the runs.
Planning across regimes
Once regimes are explicit, planning improves in a specific way: the optimizer can hold different policies for different states and price the transition risk into today's commitments. Contracts, buffers, and routing rules stop being one-size-fits-all and become state-contingent. The network behaves less like a fixed machine and more like an organism with reflexes. In markets where the stressed regime arrives a few times a year, those reflexes are worth more than any single-digit efficiency gain in the calm one.