Every real regime-detection method is a lagging indicator. A regime is defined by the persistent behaviour of returns, so evidence of a new regime is a run of new-regime days — which means recognition arrives only after some of those days have happened. Anything that claims to flag the shift on day one is either refitting the past (lookahead dressed up as insight) or noise-chasing that flips constantly. The honest question is not "how do I detect regime changes instantly?" but "how much lag am I paying, and what do I get for it?"
Price above or below a long moving average is the oldest regime definition, and it has real virtues: it is transparent, unfoolable in hindsight, and its 150-year track record across markets is well documented. Its costs are equally plain — a 200-day average confirms a bear market months into the decline, and in sideways churn it whipsaws, paying a toll on every false flip. We publish SMA and EMA boards precisely because their crudeness is a feature: they are the baseline any cleverer method must beat.
A hidden Markov model treats regimes as unobserved states, each with its own volatility and return character, and computes — day by day, using only information available that day — the probability of being in each. Two advantages matter. First, graded evidence: the model can say "60% bear" while a trend rule can only shout yes or no; the shift from 20% to 60% is information a binary rule throws away. Second, the state definitions come from the data rather than from a parameter someone chose in 1970. The costs: it is harder to explain, it can be fit badly (our validation page exists to prove ours walk-forward, out of sample), and it still lags fast breaks — probability mass takes days of ugly returns to move.
Both methods above read one market's own prices. A third family reads relationships: equities against bonds, commodities against inflation protection — the growth-and-inflation quadrant map the Compass draws. Its edge is breadth: relative prices across asset classes sometimes rotate before a single market's own trend cracks. Its weakness is indirection — it describes the macro story being priced, not your market's realised state, and the two can disagree for months.
One model flipping is a data point. Three unrelated models flipping in the same season is a regime change.
Trend rules, state probabilities and cross-asset maps fail differently — which is exactly why running all three is informative. Concordance is the signal: when the crude rule, the probabilistic model and the macro map disagree, you are in transition and should size like it; when they line up, the environment has genuinely changed. That is the architecture of this site — HMM, trend boards and Compass side by side, with every flip logged and dated so the lag we pay is public record, not marketing.
See every regime flip, logged and dated →