Radar Perene / Archive / science

The signal that only exists after waiting: the discipline of fixed windows

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

Science

The market produces interesting observations all day long — an extreme here, a divergence there, a configuration that resembles another. The problem was never finding them; it is deciding which ones deserved the name of signal. And the market's most common answer to that question is the worst possible one: the decision is made afterwards. The observation followed by a rally becomes "the signal that anticipated the rally"; the one that led nowhere becomes silence. Retroactive baptism is painless, invisible and fatal to any statistic built on top of it.

In the discipline this house practices, a signal is only recorded as valid after maturing inside a time window defined before it existed. Between the observation and the signal there is a mandatory wait, of pre-fixed duration — and whatever the market does during the wait has no vote in the baptism.

What the wait buys

The fixed window solves the retroactive baptism problem by construction. If the rule says an observation becomes a signal only upon completing its prescribed maturation — and if that duration was fixed beforehand, for all observations equally —, then the silent choice of promoting the ones that aged well and forgetting the rest no longer exists. All observations enter the same funnel, wait the same time and are judged by the same ruler. The ones that do not cross stay recorded as not having crossed — which is information, not waste.

The cost is evident and deliberate: the discipline gives up reacting in the instant. A matured signal arrives later than the headline. The methodological bet is that a registry of signals that arrive late and mean something is worth more than a stream of instant alerts whose hit rate nobody can measure — because they were baptized after the outcome.

The architecture, without the filling

The house maintains a signal registry of its own operating under this discipline, and a study about it in preparation. What can be described is the architecture: observations enter with a date; the maturation window for each type of observation was defined beforehand; the clock runs; at the end of the wait, the observation becomes a recorded signal or a record of a non-signal — and neither state can be edited afterwards. Which observations sit in the funnel today, which windows apply to each type and what the registry has accumulated: that is the registry's content, not its architecture, and it stays on the bench.

The distinction between architecture and content is not evasive — it is the point. The credibility of a signal registry does not come from the signals it contains; it comes from the rules that prevent it from lying. A registry with loose rules and brilliant signals is folklore with a stamp.

The kinship with honest testing

Whoever read Honest backtesting, step by step will recognize the logic: there, the rule is that the strategy may only use information available at the simulated date; here, the rule is that the signal may only be baptized with information available at the end of its own window. The two disciplines attack the same enemy — the leakage of the future into the past, which we treat in detail in Look-ahead bias —, one in simulation, the other in live production.

The live version is harder, and it is worth saying why. In a backtest, the future sits on disk and the temptation is technical. In production, the future arrives every day in the form of news, and the temptation is human: last week's observation seems confirmed by today's session, and promoting it to signal "because now it is clear" feels like common sense. The fixed window exists precisely to make that common sense impossible. The clock decides; today's clarity does not.

How to recognize the discipline from outside

The reader needs no access to anyone's registry to test whether this discipline exists. Three questions suffice. Do this source's signals have a birth date prior to the outcome that celebrates them? Is the baptism rule prior to the signals — published, dated, immutable? And do the non-signals appear — is there a record of what was observed and did not mature? Sources that only display winning signals, baptized at uncertain dates, under rules nobody has seen, do not keep a registry: they keep a scrapbook, edited by the memory of what worked.

Frequently asked questions

What is signal maturation, in one sentence?

It is the mandatory wait, of duration fixed in advance, between a market observation and its recording as a valid signal — during which the market's outcome has no vote in the baptism.

Doesn't the wait make the signal arrive too late?

It makes the signal arrive after the headline, by choice. The right question is not "does it arrive in time to react?", but "does it mean something when it arrives?" — and the second property only exists at the price of the first.

What window length does the house use?

The durations are part of the registry's content and stay on the bench. What is public is the rule: fixed before the observations, equal for observations of the same type, immune to what the market does during the wait.

Isn't this just pre-registration under another name?

It is a direct relative — the pre-registration of hypotheses applied to a signal's life cycle. A text further along this trail covers exactly that bridge.

---

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

House reading: what the indices record each day is in the Diário; the patterns with a track record, in the Atlas.

The same window discipline can be designed for another operation's signal registry — from the draft of the rules to the first maturation — and the house designs it alongside, when sought out.

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