Radar Perene / Archive / science

\"Looks like 2018\" is a distribution, not a prediction

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

Science

In any tense week, the sentence appears: "this looks like 2018". Or 2015, or 2008 — the year varies, the gesture does not. And the gesture carries a trap almost no one disarms: the listener hears "looks like 2018" and understands "what happened in 2018 is about to happen again". The comparison, which was an observation about the present, becomes a prophecy about the future somewhere between one mouth and one ear.

Saying that the current moment resembles a past episode means opening a fan of possible outcomes with their historical frequencies — never pointing at a single outcome. The honest analogy delivers a distribution; the dishonest one delivers a prediction. That distinction is the foundation of the historical analog engine the house keeps under study, and it stands on its own, needing no detail of the engine.

What resemblance licenses — and what it does not

Suppose today's environment resembles, by defined criteria, five past episodes. What happened after those five? If the answer were always the same, the analogy could become a near-prediction with a clean conscience. But the answer is never the same: of the similar episodes, some preceded recoveries, others preceded further declines, others preceded months of stagnation. Similarity at the starting point does not compress the variety of the arrival points.

That is why the honest output of an analogy is a set: "in episodes comparable by these criteria, the outcomes split this way". Whoever replies "fine, but what is going to happen?" is asking for the prophecy the method just refused to make. What the analogy licenses is less seductive and more useful: knowing which outcomes the past has already produced from similar starting points, and how often — the map of possibilities, not the script of the next scene.

Why the prophecy sells so much better

The prophetic version of the analogy dominates public discourse for market reasons, not methodological ones. A point prediction is memorable, quotable and — decisively — verifiable only later, after the author has already collected the attention. A fan of scenarios demands that the reader live with uncertainty, which is exactly what they wanted to outsource. The commentator who says "2018 again: the market falls" offers cognitive relief; the one who says "the similar episodes ended in three different ways" offers the truth, and the truth is more work.

There is also a statistical error built into the prophetic version, a close relative of a subject we have covered before: if prices follow something close to a random walk, the resemblance of past trajectories does not oblige future trajectories to coincide. Two random paths can be twins for months and part ways the next day without violating anything. The analogy informs about the environment; it does not chain the outcome.

How the house formalizes the intuition

The historical analog engine under study at the house starts from exactly that restriction and turns it into design: given the current environment, described by defined variables, the engine searches the archive for the closest episodes and returns what came after them — all of the "afters", not the most dramatic or the most recent. The answer takes the shape of a split of outcomes, frequencies in plain view, and the question "which one will happen?" goes unanswered by construction, because the method does not know and does not pretend to.

Which variables describe the environment, how proximity is measured, how many neighbors enter the count — that is the study's bench, and it stays there. What this article can state is the principle no technical detail alters: the output of a well-built historical analogy is always plural. When someone presents a single-outcome analogy, the reader knows one of two things was done: either the diverging outcomes were omitted, or the sample of similar episodes held a single case — and neither deserves the name of evidence.

The question left standing

One legitimate question remains: similar how? Every analogy depends on a prior choice — which features of the present count for measuring resemblance, and how many past episodes enter the comparison. That choice is not neutral, and it is where a serious analogy parts ways with a conversational one. It is the subject of the next article in this trail.

Frequently asked questions

So is a historical analogy good for anything?

It is good for what it actually delivers: knowing the repertoire of outcomes that similar environments have already produced, and the frequency of each. It is information about the terrain, not about the next step.

What if all the similar episodes had the same outcome?

Unanimity in a small sample is fragility, not certainty — half a dozen agreeing cases is still half a dozen cases. Agreement raises interest; it does not convert the fan into a forecast.

"Looks like 2018" according to whom?

That is the right question. Without a declared similarity criterion, the analogy is rhetoric: one picks the year that supports the thesis one already had. A criterion declared in advance is what separates method from illustration.

Is the house's analog engine published?

It is at the study stage, not yet deposited. The principle that governs it — returning splits of outcomes, never point predictions — is public, and it is what this article describes.

---

Continue the trail: k nearest neighbors: how the house chooses which crises to compare with today →

House reading: the archive episodes that feed any analogy are in the Atlas; today's reading of the environment, in the Diário.

Running this question — "what does the current moment resemble, and what came after the lookalikes?" — for a specific asset is research on demand, of the kind the house does in conversation.