Skip to reversion research
Mean Reversion StrategyReference / Excursion / Evidence

Mean reversion versus momentum

Compare frozen contraction and continuation rules on the same opportunity population without relabelling a chart after the outcome.

Original editorial revision: 10 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.

Define contraction and continuation at a named horizon

A mean-reversion hypothesis asks whether a declared displacement contracts. A momentum hypothesis asks whether a declared directional feature predicts continuation. Each needs its own input window, reference and future horizon. The same path can continue over a short interval and contract later, so labelling an entire chart as one regime hides the actual experiment. Begin with two specifications that could both fail.

Do not define the reversion entry as whatever the momentum rule declined, or define momentum as the winning remainder. Save separate eligibility and abstention rules before evaluation. A model residual, absolute price difference and past return are different features even when their signs coincide. Comparison is meaningful only when the reader can identify what each rule knew at the same decision boundary and what later observation was used to judge it.

Follow one fictional path through two definitions

Take invented prior closes 98, 100 and 102, with mean 100. The decision close is 106, six points above that earlier mean, and the next close is 108. A strictly-closer reversion label fails because distance expands from six to eight. A positive-direction continuation label over that one interval succeeds because price rises by two. If the following close is 101, distance is then one, and a two-interval contraction question has a different answer.

None of those labels is a trade return. A long or short order needs its own executable price, quantity, exit and cost treatment. The reference remains 100 in this example; updating it would create another test. The purpose of the path is to reveal how horizon and reference change the label, not to demonstrate that a rule should reverse after an adverse bar or that one family beats the other.

Build a fair comparison population

Record every opportunity at the shared decision clock, including cases where either rule abstains. Different eligibility rates affect capital use and implementation; silently comparing only each rule's wins is not a common population. Preserve unknown labels and source outages separately. If one rule requires a longer history, explain the resulting eligibility rather than granting the other rule access to its unavailable warmup observations.

Choose compatible units for each comparison. Prediction quality, per-event payoffs, exposure-adjusted behaviour and account returns answer different questions. A mean-reversion rule with many overlapping positions cannot be compared with a single-position continuation rule using a naive sum of independent trade returns. Freeze funding and position-combination conventions before the economic comparison. Reporting both rule-specific populations and the common eligible subset helps readers see what was gained or lost by aligning them.

Do not infer a trading regime from a convenient coefficient

An autoregressive coefficient describes a declared process and fitting sample. A positive coefficient below one can describe decaying expected displacement under an unchanged equilibrium without forecasting a particular first-passage time. A coefficient estimated on transformed data is not automatically a property of raw prices. Save the transformation, deterministic terms and uncertainty instead of presenting one fitted number as a permanent market label.

Likewise, a rising moving average does not establish an independently validated momentum edge. If a regime filter is used, document its observations and availability time before it selects between rules. A later crisis label or retrospective narrative cannot be inserted into an earlier decision. Keep a stable no-switch baseline so any claimed benefit of switching can be inspected. This guide contains no fitted coefficient or empirically selected regime boundary.

Reserve evaluation and preserve entry identity

Train feature transformations and choose authorised parameters within the declared research region. Chronological splitting is useful, but the documented TimeSeriesSplit mechanism does not inspect a financial label's complete future span. Check whether labels or warmup histories overlap forbidden regions and whether an imposed gap matches the actual information horizon. Irregular event samples also require a declared clock rather than assuming equal-duration folds.

Freeze the entry objective and identity before applying one predeclared exit for economic validation. An exit can reject or constrain an implementation without retroactively changing which entry was ranked first. If a comparison adds a discretionary rescue or different holding rule after seeing the reserved result, save it as a new experiment. The previous reserved outcome is now known; repeatedly retesting on it does not create independent evidence for the revision.

Report differences a reader can inspect

A comparison packet includes both specifications, common and rule-specific counts, source availability, candidate history and evaluation outputs. Show adverse intervals and unresolved exposure alongside favourable examples. If a cost or funding convention differs between implementations, explain why and retain a like-for-like sensitivity case where possible. Avoid a universal winner when the evidence concerns only one population and horizon.

The blank comparison record captures the definitions rather than imaginary performance rows. Reproduce the one- and two-interval labels from the fictional path before running a real study. A valid conclusion may be that the two rules measure different phenomena or that available data cannot support execution claims. This article neither recommends a trading family nor reports a completed head-to-head backtest; it helps a consumer recognise the evidence needed before accepting one.

Review checkpoints

Evidence requests, not completed study results
ObjectPreserveDo not infer
DefinitionReference, direction and horizonOne chart has one permanent strategy label
ComparisonCommon opportunities and abstentionsSelected closed winners form a fair test
SelectionFrozen rules and available regime inputsA retrospective rescue remains systematic

Download blank review record (CSV)

Headers only. No trades, subscriber outcomes or operating receipts are supplied.

Source documents

Documentation establishes the stated mechanism, not tested trading performance. Checked 10 October 2026; a later version can differ.

Detailed research

Process and horizon change the comparison

A price-level reference, return autocorrelation and fitted residual represent different objects. A directional continuation can coexist with temporary movement toward a destination, and an overshoot can cross while distance grows. A meaningful comparison states the actual process, horizon and financial action before interpreting those labels. This edition supplies no head-to-head market backtest or favorable algorithm winner. Its path and card specimens expose reference choices while leaving actual fitted models and independent completion absent. Retain the original discussion and request compatible genuine sources for any stronger comparison.