Skip to reversion research
Mean Reversion StrategyReference / Excursion / Evidence

Stationarity, regime changes and falsifiable reversion

Design a bounded stationarity experiment without turning a unit-root test, fitted spread or chosen chart into proof of trading profit.

Original editorial revision: 9 October 2026. Expanded edition: 14 October 2026.

Educational diagnostics, not a signal service. A standardized price displacement is not a calibrated probability; invented examples and user-supplied observations are not verified market performance.

State which process is supposed to revert

A price level, log price, return and fitted pair spread are not interchangeable test inputs. Name the exact series and transformation before choosing a stationarity test. For a spread, record both instruments, synchronization, adjustment policy and how the hedge coefficient was estimated. A fitted relationship chosen with the whole history can make its training picture look stable while using information unavailable earlier. Preserve the estimation boundary and the later application rule. A visual line oscillating around zero is an invitation to investigate, not proof that future observations will remain governed by the same relationship.

A test result answers its specified hypothesis

The Augmented Dickey-Fuller documentation describes a unit-root null and exposes choices including deterministic terms and lag treatment. Record those choices, sample dates, missing-value handling and the reported statistic. Rejection of the specified unit-root null does not establish a profitable trading rule; failure to reject does not prove that every bounded-horizon recovery idea is impossible. A p-value is not the probability a trade will win. Repeatedly changing trends, lags or transformations after seeing the output creates another selection problem. Report the complete procedure rather than advertising the smallest p-value found during exploration.

Keep fitting and application apart

Estimate transformations, hedge weights and reference parameters using the declared training region. Apply that frozen definition to the validation and reserved regions. If an operational rule refits periodically, specify what it can know at each refit and preserve the version that produced each event. Do not silently recompute all old residuals with today's coefficient. Chronological separation alone cannot cure a field that was revised later or a label that reaches across the split. A practical audit should follow one observation from source timestamp through transformation, event generation and final outcome attribution.

Predeclare what a break looks like

A regime-monitoring rule needs an observable condition and an action, not the retrospective statement that a bad period was unusual. Possible research questions concern reference drift, changed dispersion, persistent residual sign or data-quality failures. The thresholds and response should be part of the specification before a reserved sample is opened. A pause, re-estimation or retirement changes the tradable population and must remain in the record. Do not erase difficult events because a later explanation calls them a different regime. Preserve whether the interruption arose from a model condition, exchange calendar or operator decision.

Add economics after the diagnostic contract

A stable statistical process can still be expensive or inaccessible to trade. A pairs implementation needs both legs, short availability where relevant, financing, spread, execution sequencing and handling of a partial position. The notebook result is not a broker fill. Use a separately labelled execution study with the declared entry and exit rules, costs and account constraints. Keep rejected or unfilled opportunities in the population report. Do not substitute a convenient continuous series for a particular executable contract without changing the label and admitting that the reference study answers a narrower question.

Make rejection a first-class output

Write a compact experiment record: process identity, hypothesis, source version, causal transformations, test settings, independent evaluation boundary and stop conditions. Retain observations that fail the reference rule, not only those that recover. If a source cannot establish its historical availability, leave the question unresolved rather than inventing timing. If a chosen relationship fails under the frozen evaluation, save that result. The next hypothesis may be worthwhile, but the inspected sample is no longer untouched evidence. A credible reversion publication teaches readers how to discard a fragile idea as clearly as it teaches the attractive formula.

A predeclared failure sheet

Before obtaining observations, write two different failures: the historical source cannot establish its availability clock, or the frozen process fails the specified reserved diagnostic. The first is unresolved data identity; the second is a research result under a valid contract. Do not collapse them into the same verdict. Add a third case where the statistical diagnostic passes but execution costs defeat the implementation. This small decision sheet keeps missing evidence, statistical behaviour and trading economics from being confused later.

Source documents

Documentation establishes the stated mechanism, not tested trading performance. Checked 9 October 2026; a later version can differ.

Continue the investigation

AR(1) expectation half-life worksheet

A destination revision creates new label ownership

The changed-reference card proposes replacing an event's reference with a later value. That is a different question, not confirmation that the old process was stable. A structural change and a revised data source need their actual evidence; neither is diagnosed by the seven authored paths. The tool fits no ADF or KPSS test and estimates no equilibrium. Genuine research must freeze its process and allowed revision policy before selection, then preserve the independently required master rather than treating a descriptive path as a stationarity certificate.