The model works. The strategy does not.
A hidden Markov model sorts each session into three volatility states, refit out of sample on every market independently. It succeeds at what it was built for and fails at what it was hoped to do.
Across 11 markets and 2,571 out-of-sample sessions, volatility rises monotonically through the three states on 11 of 11. Return does not. It carries no ordering at all on 8 of 11, and where an ordering does appear it points both ways: ascending on 2, backwards on 1. The states rank risk. They do not rank return, and a strategy that trades them as though they did loses to a static sixty-forty.
Eleven markets, three years, one plate
Each row is one market: its price path above, its regime rail below. Crisis is the only state that washes the whole row, so simultaneity reads as a column instead of as eleven separate rails.
BullBearCrisisEleven models were fitted independently, one per market, with no shared state variable. The alignment describes one event. It is not a contagion claim, and nothing here estimates a link between markets.
What the plate shows
Between 10 Apr 2025 and 7 May 2025, 9 of 11 markets sat in Crisis at once, and held there for 28 days. Eleven models that share no state variable agreed on one month.
Then look again at the price paths inside that band. Several of them are rising. That is the finding: these states rank violence, not direction, and de-risking into the darkest one sells the rebound along with the crash.
The same eleven markets, twice
Both plates are built identically and differ in exactly one variable. Left, annualised volatility by state. Right, annualised return by state. Read down each column and the argument makes itself.
ANNUALISED VOLATILITY · RISES ON 11/11
ANNUALISED RETURN · NO ORDERING ON 8/11
Each line runs Bull to Bear to Crisis, left to right; dots carry the state colour. Markets with fewer than 3 crisis episodes are excluded from the return verdict, because an ordering read off two episodes is not an ordering. Backwards on SPY. Ascending on ^CNXIT, NG=F.
Where each market stands today
State, how long it has held, and the position size implied by holding a constant 10% risk budget. That last figure is arithmetic on a measured volatility, not a backtested position.
INDIA
GLOBAL
COMMODITIES
1 of 11 markets are in Crisis as of 24 Sept 2026: Gold. Position size is target volatility divided by the realized volatility measured in that regime, capped at 1.0. Arithmetic on a measured quantity, shown so a risk budget can be read off the regime. It is not a backtested strategy, not a traded position, and not advice., capped at one; the cap is drawn rather than hidden.