Pre-Registration
The practice of recording a hypothesis, its test design, and its success criteria — timestamped — before examining the data that will judge it, so that a confirmation cannot have been shaped by the answer.
Pre-registration is a discipline quantitative research borrowed from medicine. Clinical trials began registering their designs in advance because the alternative had a name and a body count: outcomes switched after the data arrived, negative results quietly shelved, analyses re-run until something cleared significance. Markets research has the same disease with fewer regulators.
The mechanism it defends against is not fraud but flexibility. Every choice made after seeing the data — which instruments to include, which window to test, which metric to headline, which “obviously broken” period to exclude — is a researcher degree of freedom, and each one consumed silently converts an out-of-sample test into an in-sample fit. A hypothesis written down before the test binds those choices while they are still cheap to bind: universe, period, parameters, evaluation metric, and the result that will count as failure, all fixed before the data can argue back.
The binding is only as strong as its provability. A registration that could have been backdated proves nothing, so the timestamp must be independently verifiable — which in practice means registering into something append-only, and recording when the data itself arrived relative to the registration. The ordering is the evidence: a hypothesis registered before the data existed is a prediction; the same hypothesis written afterwards is a description wearing a prediction’s clothes. Our own Burned Door post-mortem documents what it looks like when that ordering is destroyed rather than preserved.
Pre-registration pairs with the kill ledger: registration records what a trial was supposed to be, the ledger records what became of it, and together they make the trial count behind a surviving strategy auditable. Neither makes a strategy good. They make the evidence for it mean what it claims to mean — which is the property data snooping removes, and the property no amount of after-the-fact statistics can restore to an unregistered result.