Radar Perene / Archive / science

How many windows fit a signal: why seven and not one

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

Science

Ask whoever promotes a "signal that works" what its horizon is, and watch the answer slide. Does it work over a month? Three? A year? The usual reply — "it depends on the investor's profile" — sounds like prudence, but it is the opposite: it is a confession that the horizon will be chosen later, once it is known which one made the signal look good. A signal without a declared horizon is a target painted around the arrow.

In the house's signal registry, every signal is evaluated simultaneously across seven fixed time windows, all defined before any result is known — because a signal that only works over a horizon chosen after the fact is not a signal, it is a selection. Which seven windows, and their sizes, is bench detail; what matters to the reader is the architecture, and the architecture explains itself in full.

The single-window problem

Evaluating a signal in a single window creates two temptations, both silent. The first is retroactive choice: with the result in hand, there is always some horizon over which the performance looks respectable — the honest researcher who "just checked the other windows to be safe" has already contaminated the test without noticing. The second is invisible fragility: a signal can look consistent over one horizon and invert its behavior in the neighboring one, and the single window would never reveal it. The reader of a single-window result has no way of knowing whether they are looking at a pattern or at a cropping.

The finance literature knows this vice up close — it is a cousin of the problem we describe in sample size in financial series: when the researcher is free to choose the cut after seeing the data, the effective quantity of evidence collapses, even when the number of observations looks large.

What multiple fixed windows buy

Fixing several windows before the test buys three things a single window cannot offer.

The first is structural honesty: with the horizons locked in advance, there is no room to pick the flattering cut — the signal will be seen in all of them, including the ones that contradict it. The second is a portrait of coherence: a genuine pattern tends to show up recognizably across neighboring horizons, while chance tends to produce disconnected results from one window to the next. Coherence across windows proves nothing by itself, but incoherence warns of a great deal. The third is a demarcation of reach: real signals have expiry dates — some show up fast and dissipate, others only appear over months. The span of windows marks out where the signal lives, and a signal with a declared address is both more useful and more falsifiable than a ubiquitous one.

Why not one, and why not fifty

If seven windows beat one, would fifty beat seven? No — and the reason is the other face of the same vice. Multiplying windows indefinitely multiplies the odds that one of them shows a striking result by pure chance; whoever tests fifty horizons and highlights the best three has reproduced, with an air of rigor, the very retroactive selection fixed windows exist to prevent. The number of windows is a design decision: few enough that each one means something, many enough to cover the horizons where the question makes sense — and, above all, locked before the test.

That is why the house's answer is "seven and not one", rather than "the more the better". The exact count matters less than the two properties holding it up: enough plurality to expose fragility, and prior fixation to eliminate choice. A battery of windows chosen after the results is just the single window with more ceremony.

The discipline outlasts the finding

The signal registry is a house study still in draft, and the part of it this article describes — the fixed-window architecture — is the part least dependent on any result. That inversion is deliberate: in market folklore, the value lies in the signal and the method is a footnote; in a research operation, the method is the durable product and each individual signal is a temporary guest. Signals are born, age and die; the ruler that evaluates them is what remains — and it is the ruler, not the guest of the moment, that a house can show in public without handing over what is still under test.

Frequently asked questions

Why exactly seven windows, and not six or ten?

The specific count is an internal design decision, calibrated to the horizons where the house's questions make sense. What this article defends is not the seven — it is the pair of properties: multiple windows, fixed before the test.

Does a signal need to work in all seven to count as valid?

No. Demanding universal approval would be another form of arbitrariness. What the span of windows produces is a portrait of coherence: where the signal shows up, where it does not, and whether that shape is stable — the assessment is of the whole, not of a scoreboard.

Doesn't this make research slower?

It does, and that is part of the argument. A signal evaluated over multiple fixed horizons requires waiting for the longest one to mature. The slowness is the cost of the honest answer; the fast alternative answers a different question.

Do fixed windows eliminate the risk of self-deception?

They do not — they close one specific avenue of it, the retroactive choice of horizon. Other avenues remain open and demand other defenses, such as pre-registering hypotheses and retesting on new data.

---

Continue the trail: The signal registry that also knows when to quit →

House reading: the state of the indicators, window by window, is in the Diário; the patterns that survived time, in the Atlas.

Designing a window scheme for a specific strategy — how many, what sizes, fixed how — is bench conversation, and the house has it with those who seek it out.