Skip to reversion research
Mean Reversion StrategyReference / Excursion / Evidence

Rolling threshold selection and untouched holdout

Preserve the history of lookback, stretch, band and horizon choices instead of reporting only their favorable combination.

Every tuned definition has a selection role

Lookback, scale convention, stretch threshold, band and horizon can all affect an observed recovery table. Changing any after inspecting outcomes uses that sample for selection. A chronological label does not restore its untouched role. Save every attempted candidate and why it was accepted, rejected or revised. A simple deterministic rule can have a large search history even with no fitted weights. The site supplies known authored examples to inspect behavior, not a blind threshold tournament or an optimized combination that readers should use in a market.

A rolling label can hide changing candidates

If a reference or threshold is recomputed under a predeclared policy, identify that policy as part of the candidate. If updates are chosen ad hoc after unfavorable results, preserve that different history. A current chart may not identify the version that generated earlier events. Keep raw causal source and transformation scope alongside settings. Genuine fitting and selection need their permitted periods. The original lab lets visitors explore entered settings, but its descriptive output does not automatically become reserved financial evaluation or an independently certified strategy when a favorable number appears.

Horizon changes can select the endpoint

A three-observation non-hit may become a five-observation hit on the same inspected path. That does not validate the original horizon. The bounded workbook case retains later values but excludes them from its chosen metrics. Changing band can likewise create a step-zero success without improving an input forecast. Outcome definition and native economic exit are distinct objects, both requiring frozen ownership. Real entry ranking remains raw and causal before the experiment's one predeclared exit is applied. A later exit-adjusted result does not rewrite which entry candidate was originally selected.

A reserved sample needs preserved access history

Record when each outcome region was first opened and which decisions it informed. A repeated tuning surface cannot be called untouched confirmation for the last revision. Software fixtures have expected answers by design and must not be counted as independent strategy samples. The authoritative full research factory and independent master determine real completion. A valid no-model master can finish the process, while a compact threshold grid alone cannot. The website creates no additional paid capacity or model promotion authority simply because a candidate deserves further genuine testing.

Authored reference and excursion case

An authored investigator tries band one over three observations, then changes both band and horizon after seeing a later touch. The second result may be correctly computed but belongs to a revised selected question. It cannot be described as untouched confirmation of the first. No real price history or profitable threshold is supplied. The exercise asks for a candidate registry and sample-access record before a table of favorable recoveries is interpreted as independent evidence.

ObjectRecordBounded finding
OriginalBand 1 / horizon 3Frozen first question
Inspected later hitOutside original horizonSelection information
Revised settingsNew band or horizonNew candidate scope
Reserved evidenceActual access record neededNot restored by renaming
Authored gradual and overshoot paths around frozen reference 100
Authored EDU paths, not genuine market prices, return observations or calibrated odds.

Primary references and scope

Source review: 14 October 2026. Official mechanism descriptions are not this desk's paid-account test, independently captured fills or vendor-return audit.

Related reversion questions