Radar Perene / Archive / science

Regime changes, analog changes: when comparing with the past stops making sense

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

Science

Every method has a question it would rather not hear. For historical analogy, it is this one: what if the present resembles nothing at all? The method lives by finding neighbors in the archive — but who guarantees the archive contains legitimate neighbors? A market that has changed structure, participants or rules can produce configurations no earlier episode anticipates. In that case the "nearest" neighbor still exists arithmetically — a list of distances always has a minimum — but its proximity is an artifact of the computation, not a fact about the world.

A historical analogy model must know how to declare that the present has no valid parallel in the archive — because the nearest neighbor of an unprecedented moment is merely the least distant of the strangers. That capacity for refusal is, in the design of the house's analog engine, as constitutive as the capacity for answering.

The problem: minimum distance is not similarity

The mechanics of formal analogy carry a silent weakness. Ask for the k nearest neighbors and the method delivers them — always. It was built to order distances and pick the smallest; it was not built to warn that all the distances are large. Without an additional layer of judgment, the model treats "the most similar episode in the archive" and "a similar episode" as synonyms, and they are not: the first is a relative superlative, the second an absolute claim.

An everyday parallel helps: search a summer wardrobe for the coat best suited to a blizzard. Some coat will top the list — and wearing it beats nothing, but presenting it as "the wardrobe's answer to blizzards" would be a fraud of framing. That is exactly what a historical analogy does when it never measures the absolute quality of its own neighbors: it hands over the best of an inadequate set with the same face it would give a genuine twin.

Regimes: when the whole archive ages at once

The deeper reason for this caution has a name in the literature: regime change. As we describe in regime switching, economic series do not evolve under fixed rules — they alternate between states with distinct behaviors, and the borders between states are precisely where regularities break. An analogy computed with neighbors from another regime imports outcomes from a world whose rules no longer apply.

The delicate point is that regime changes do not announce themselves. A currency that moved from managed to floating, a structural interest rate that stepped down from its historical level, a transformed market composition — after ruptures like these, earlier episodes remain in the archive looking exactly as they always did, but part of their information has expired. The archive does not become wrong; it becomes dated in sections, with no label saying which.

Refusal as a first-class answer

In the design of the house's engine, a query that finds no valid parallel returns exactly that: the declaration that the current moment lies too far from everything the archive records for the split of outcomes to deserve trust. What the numerical criterion of that refusal is, and which real cases it has already discarded, is bench content of the draft study. The principle, however, is public and is what matters: the refusal is not a failure of the model — it is one of the correct answers, and among the most informative.

It is informative because it says something no forced neighbor would: we are in territory the mapped past does not cover. For anyone reading risk, that is first-order information — uncertainty without precedent calls for more humility than uncertainty with a repertoire. And it protects the rest of the method: an engine that always answers, including when it should not, contaminates the trust owed to the times the answer was legitimate. The credibility of "the similar episodes ended this way" depends on the system knowing how to say "there are no similar episodes" when it is true.

Lay usage has the mirrored defect

The contrast with market conversation is worth noting, because it errs in both directions at once. Facing a genuinely novel moment, public commentary rarely admits the void — it forces the most dramatic available analogy. And facing an ordinary moment, with dozens of modest precedents, it prefers to declare it "unprecedented", because novelty sells. The formal method, with a refusal criterion declared in advance, errs less at both ends precisely because the decision to compare or refuse does not belong to the narrator's mood.

Frequently asked questions

How does one detect a regime change without waiting years?

With difficulty — and any honest answer starts there. Clues exist: structural breaks in previously stable relationships, dated institutional changes, persistent behavior outside historical bands. None is conclusive in real time; the regime-switching literature deals with exactly that frontier.

Couldn't the model's refusal be excess rigor?

It could, and the cost is asymmetric on purpose: refusing a valid analogy costs one answer; accepting an invalid analogy costs a wrong answer wearing the look of evidence. The design prefers the first error.

Should episodes from old regimes be removed from the archive?

No — removal would be erasing history. The issue is not expelling episodes but qualifying the comparison: a neighbor from another regime may inform less, and the model must reflect that instead of treating it as a full peer.

Has this happened in the house's practice?

The principle of discarding for lack of a valid parallel is part of the engine's design under study. Which queries have already ended in refusal is operational content, not public — the commitment to the refusal's existence is what can be stated.

---

Continue the trail: Why comparing with 2008 and with 2020 gives different answers →

House reading: the borders between regimes, dated case by case, are in the Atlas; the current state, in the Diário.

Assessing whether a specific period has a valid historical analog — or sits in unmapped territory — is an examination the house's bench conducts when sought out.