Radar Perene / Archive / science

Recomputing a signal registry: what remains when the right ruler is applied

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

Science

Every registry of market observations accumulates, over time, an uncomfortable question almost nobody asks out loud: how much of what is here would have been recorded if today's rules had applied from the start? Institutional memory is generous with itself — observations noted with imprecise dates, signals remembered as "obvious at the time", rules that firmed up gradually and whose final version nobody knows when it began to apply. The house asked its own registry that question. And answered it the only way that produces an answer: by recomputing everything.

Recomputing a signal registry means reprocessing one's own history applying, at each past date, only the information that existed up to that date — the right cutoff date — and the declared rules, with no courtesy exceptions. The typical result is a contraction: part of what "counted" as a signal stops counting. The contraction is not a loss; it is the measure of how generous memory had been.

Why registries inflate on their own

No bad intention is needed to inflate a registry. Three banal mechanisms suffice. The first is the imprecise date: an observation remembered as "from March" may have been consolidated in April — and a month's difference is the difference between anticipating and commenting. The second is the moving ruler: rules that evolved over time get applied mentally backwards, and old observations receive the benefit of criteria that did not yet exist. The third is the subtlest: context information. What was known "at the time" tends to include, in recollection, pieces of what was only known later — the silent leakage we describe in Look-ahead bias.

The three mechanisms push in the same direction: upward. Memory-based registries only err on the side of more. That is why recomputation has a predictable result in sign — it shrinks — and an unpredictable one in size: nobody knows, before running it, how much of their own history survives.

The house's exercise

The house applied this ruler to its own signal registry, in a dated and recorded exercise: the entire history was reprocessed with the correct cutoff dates and the declared rules, the past held to the same standard as the present. Part of what was on record stopped being on record — observations that, held to the information actually available at their date, did not meet the birth criterion the registry demands.

How much it shrank, which signals fell and from which period: that stays on the bench, and the reason is the same as in the other texts of this trail — the registry's content is working material; what is offered to the public is the discipline. What can be said with precision is the principle of the outcome: the post-recomputation registry is smaller and worth more. Every line that remained now has a property that used to be hope: it would survive an audit with the dates in hand.

The kinship with calendar folklore

The reader who followed Statistical seasonality vs. folklore will recognize the structure of the problem. Calendar folklore — the month that "always" falls, the rally that "always" comes — is a collective signal registry, kept by the market's memory, never recomputed. It inflates through the same three mechanisms: imprecise dates, moving rulers and context leaking backwards. The difference between folklore and a defensible statistical regularity is exactly one recomputation: apply the right ruler, with the right dates, and see what remains. What remains tends to be less — and it is the part that matters.

The lesson transfers to any archive: the question "how much of this would survive today's ruler applied with yesterday's dates?" is devastating and fair. Archives that never faced it carry an unmeasured liability of generosity.

What the contraction teaches

One effect of the exercise deserves its own record: recomputation does not only correct the past — it recalibrates the present. Knowing the registry will be reprocessed with dates in hand changes the care with which each new observation is noted today: the consolidation date gains rigor, the applied rule becomes explicit, the available context is separated from what has not yet arrived. The auditable past disciplines the present that has not yet become past. It is the same mechanism the reproducibility audit produces on code: the habit of re-executing improves what is yet to be executed.

Frequently asked questions

What is recomputing, in one sentence?

It is reprocessing a history applying at each past date only the information available up to that date and the declared rules — the audit that separates what was recorded from what was remembered.

How much did the house's registry shrink?

The number stays on the bench, along with the registry's content. What is public is the direction — it shrank — and the methodological reason: memory-based registries only err on the side of more, and the contraction is the measure of that generosity.

Was an observation cut in recomputation a mistake?

Not necessarily. It may have been a reasonable reading in its time — it merely did not meet, with that date's information, the formal birth criterion of a signal. The cut does not judge the observation's intelligence; it judges its documentation.

How often should one recompute?

The ruler matters more than the calendar: whenever rules or sources change, the earlier history needs reprocessing under the new version — or the registry starts mixing periods with different criteria, which is the defect it exists to avoid.

---

Continue the trail: What "existing" means for a market signal: deadline, not intuition →

House reading: the record of the present, made with the dates in hand, is the Diário; the patterns that survived auditing, the Atlas.

Running a third-party signal history through the same ruler — right cutoff dates, declared rules, measured contraction — is a service the house performs on request.

Read also: Look-ahead bias: the error that lets a model \"predict the past\"