Radar Perene / Archive / science

Pre-analysis registration: the hypothesis declared first

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

Science

Every quantitative research project has an invisible drawer. In it sit the tests that came to nothing: the thirty-nine variations that failed before the fortieth worked, the windows that spoiled the result, the indicators that refused to cooperate. The reader of the final study never sees the drawer — only the winning test, presented as if it had been the first and only question. Pre-analysis registration exists to lock that drawer before it can be filled.

A pre-analysis registration is a dated document, filed before any test is run, that declares the hypothesis, the data to be used, the method and the criterion for success. When the result comes out, it is compared with what was promised — not with what it would have been convenient to promise afterwards.

Where did the idea come from?

Medicine got there first, for the gravest reason: clinical trials that switched outcomes midstream, reporting the effect that showed up instead of the effect they had set out to measure. The answer was to make prior registration a condition for publishing. Economics followed — the American Economic Association's trial registry has operated since 2013 — and the logic migrated naturally to any field where data are abundant and attempts are cheap.

Market research is the extreme case of that abundance. Dozens of series, hundreds of possible windows, thousands of parameter combinations. Whoever tests enough finds something — statistics guarantees it by chance alone, with no merit from the researcher. The difference between a discovery and a search artifact is, very often, a single piece of information: was the question asked before or after looking at the answers?

What registration changes in practice

The prior document reverses the burden. Without it, the author must convince the reader that they did not rummage until something turned up; with it, the doubt dissolves on its own, because the hypothesis is dated before the test. Three practical effects follow.

There is a subtlety the methodological literature nicknamed the "garden of forking paths," in Gelman and Loken's phrase: the researcher does not need to test forty variations in bad faith to fool themselves. It is enough to take, at each fork in the analysis — excluding or keeping that atypical year, using mean or median, a twelve- or twenty-four-month window — the decision that looks reasonable in light of the data already seen. Each choice is defensible; their sum is a disguised search. Prior registration cuts the problem at the root because the forks are decided before there is any landscape to influence them.

The first effect is disciplining the author. Writing "I expect X to happen because Y" before running the model is embarrassing in exactly the right dose — it forces a reason, not merely a hope. The second is giving negative results a status: if the registered hypothesis fails, the failure is a finding, not an embarrassment to hide. The third is enabling outside audit: anyone can compare what was promised with what was delivered.

How this shows up in the house's work

In Radar Perene's internal protocol, no test runs without a written hypothesis — the format is short, almost telegraphic: what is expected, in which direction, and why. The applied protocol stays outside this article; what can be shown is its public consequence.

The third working paper in the house's series, on the tactical version of the Ânima Index, tested a specific hypothesis: that extreme readings of the indicator would carry an exploitable contrarian edge. The result did not confirm the hypothesis — and the study was published anyway, with the refutation in its own conclusion, in an open repository with a permanent identifier. In numbers: the text sits on Zenodo under DOI 10.5281/zenodo.21327608, with the tested hypothesis and the negative outcome in the same document. Publishing the test that failed costs some pride; not publishing it would cost the only proof that the tests which succeeded went through the same sieve.

That is the point that usually goes unnoticed: the value of prior registration lies not in the studies that confirm, but in the ones that deny. An archive containing only victories is indistinguishable from a well-hidden drawer.

What registration is not

Pre-analysis registration is not a straitjacket. Free exploration of the data remains legitimate — it is where the questions come from. What changes is the label: what was explored presents itself as exploration, and only what was declared before the test presents itself as a test. Nor is it armor against error: a registered hypothesis can be badly framed, measure the wrong thing, or fail through sampling bad luck. Registration documents the honesty of the process, not the truth of the result.

Frequently asked questions

Doesn't registering in advance stifle research?

No, if the division is clear: exploring is free; testing requires prior declaration. Many registrations even provide for additional analyses labeled as exploratory.

Where does a researcher register a pre-analysis plan?

In public field registries (such as the American Economic Association's, for economics), in open repositories with a timestamp, or in any file whose creation date can be verified by third parties.

How does pre-analysis registration differ from p-hacking?

They are opposites. P-hacking is testing many variations and reporting the one that passed; prior registration makes that maneuver visible, because the reported test must match the declared one.

Is testing several hypotheses always wrong?

No — as long as all of them appear in the results, with the appropriate statistical corrections, and not just the winner.

---

Continue the track: How a paper is submitted to a Brazilian economics journal →

House readings: the daily record, dated and anchored, is in the Diário; the precedents that survived testing, in the Atlas.

Turning a loose question into a registrable hypothesis is an exercise the house knows from the inside — the kind that yields more in conversation than in text.