Radar Perene / Archive / science

Regime switching: when the rules change mid-game

◦ Index methodology v2.2 (working papers with DOI). See the methodology.

Science

A market relationship works for ten years with the regularity of a clock. Analysts cite it, models incorporate it, custom promotes it to law. Then, without ceremony, it stops. The first months of failure are treated as an exception; the following ones, as bad luck; a year later, someone finally states the hypothesis nobody wanted: the clock did not break — the time zone changed. The world that generated the regularity ceased to exist, and the series began obeying a different rule.

Regime switching is the name the literature gives to the idea that a series does not obey a single set of statistical rules all the time: it alternates between persistent states — regimes — and the typical behavior of one state can be invalid in the other. The average, the agitation and the relationships with other series can all change value when the regime changes.

In Portuguese, the term appears almost exclusively in quantitative-management material, treated as a modeling technique. This article treats it as something else: a problem of knowledge. If the rules change, what exactly does a historical average describe — and for how long?

From Hamilton onward

The classic formalization has a date: in 1989, James Hamilton proposed a model in which the American economy alternated between two states — expansion and recession — each with its own dynamics, and in which the switch between them was itself part of the model. The internal engineering belongs in the machine room; the conceptual novelty survives any summary: instead of treating breaks as defects of the series, the model treats them as citizens — the state of the world is a variable, not a nuisance.

The uncomfortable consequence comes with it. If a thirty-year series crossed three regimes, whoever computes a single average over the thirty years has measured none of the three worlds — they have measured a blend that may never have existed on any single day. The number comes out precise in appearance and orphaned in reference: it describes an average regime no investor ever inhabited.

When a stylized fact expires

The concept yields the most when it stops being technique and becomes an epistemological question: how much of what is known about a market has an expiry date? This trail has already examined how a recurring pattern becomes a stylized fact — a regularity that survived enough tests to earn the name. Regime switching is the termination clause of that contract: a stylized fact is always a fact of a regime, and the archive that recorded it does not guarantee the next regime.

That does not condemn historical knowledge — it condemns the lazy use of it. The mature response is not to discard the past but to date it: to ask, before each regularity, under what conditions it was observed and whether those conditions still stand. Between "this has always worked" and "this worked while the world was like this", the second sentence is longer, less sellable, and the only defensible one.

Where the house met the rule change

This house's experience with the subject is public in two records. The first is an entire study devoted to the problem: one of the six working papers in the house's series, public on Zenodo under DOI 10.5281/zenodo.21402939, builds a regime score for the Brazilian macroeconomic environment — the study's question, in lay language, is how to recognize which state the environment is in using public data. The document is the credential; the applied method lives in it, not in this article.

The second record is more uncomfortable and, for that reason, more valuable: the house's intramarket-relationships working paper (10.5281/zenodo.21327663) had a second version that openly revises a finding the first one upheld. A relationship that looked firm under one period's data did not survive re-examination — and the revision was published with the same visibility as the original finding. In numbers: both versions remain public on Zenodo, with active DOIs, side by side. That is the protocol working: when the rule changes, the record changes — the alternative would be defending a stopped clock.

Frequently asked questions

Can regimes be predicted?

The literature is far more successful at dating regimes retrospectively than at anticipating switches. Recognizing the current state with a short lag is already hard; predicting the next switch runs into the same limits this trail discussed under the random walk.

How many regimes does a series have?

The number is a modeling choice, not a revealed property of nature. Two- and three-state models dominate the literature out of parsimony — which serves as a reminder that "regime" is a reading lens, not an anatomical organ of the series.

Is regime switching the same as a structural break?

Relatives with different temperaments. The classic structural break is a single, permanent rupture; regime switching describes recurring states between which the series alternates. A change of era calls for the first lens; the alternation between calm and turbulence, for the second.

How does a reader use the concept without running any model?

With one hygiene question: was this number — average, correlation, pattern — measured inside a single regime, or does it straddle different worlds? Studies that declare the period and conditions of their sample answer it; studies that hide the window answer it too, in another way.

There is one kind of regularity that promises to outlive every regime: the calendar's. Every year has a May — and every year someone repeats the proverb about selling in May. What is left of it when tested properly is the next step: Seasonality or folklore: what survives of "sell in May" →

House readings: today's note, in the Daily · the precedents, in the Atlas.

Examining whether a specific regularity belongs to the current regime or the previous one is workbench material — the kind of exercise the house conducts on request.