Causal z-scores and the overshoot problem
Define a prior-window reference, measure absolute recovery and avoid calling every correctly directed move a successful reversion.
Define the strategy, reproduce conditional expectancy and distinguish statistical recovery from executable account economics.
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.
Mean reversion names a family of hypotheses, not a single investment product. Specify what is expected to contract: price displacement, a spread, a relative value measure or a model residual. Name the earlier reference, qualifying event, decision clock and holding horizon. The question becomes whether this frozen specification has a reproducible economic result in an identified population, not whether the strategy label is generally good.
Write the implementation before evaluating outcomes. Identify instruments, units, entry and exit conventions, permitted exposure and treatment of unresolved opportunities. Different horizons can turn the same price path into different trades. A historical reference study may be useful without supporting actual orders. Keep its scope in the title so a buyer does not mistake a statistical chart for a verified subscriber account. No strategy family receives a universal profitable verdict in this publication.
Consider invented closed outcomes with a 0.60 win fraction, an average gross winning outcome of +1R and an average gross loss of -2R. The conditional gross expectancy is 0.60 times 1 minus 0.40 times 2, or -0.20R per closed outcome. If additional costs average 0.10R, the conditional net figure is -0.30R. A sixty-percent winning fraction is therefore compatible with a negative expectancy under these assumed payoffs.
For a separate fictional rule with wins of +2R and losses of -1R, the same fraction and extra cost yield +0.70R. That difference follows from the invented payoff contract; it is not a recommendation or an empirical comparison. Define R's monetary denominator and whether the payoff figures already include costs. A gap loss can exceed the intended one-R budget. The arithmetic does not supply an executable account return, capital requirement or probability of future profitability.
Stationarity and reference recovery answer statistical questions. The ADF source documents a unit-root test with choices for deterministic terms, lag structure and sample. A rejection under a particular specification is not a probability that a trade reaches a target. It does not choose a broker, spread, position size or economic holding horizon. A non-rejection also needs interpretation within the test's assumptions and the available sample.
If studying a rolling reference, do not treat the reference's movement as price recovery automatically. Measure distance against the declared unchanged earlier reference when that is the chosen label. A distribution assumed for a z-score is another modelling commitment, not an observed normal-tail probability. Keep test outputs, recovery counts and executed payoffs in separate tables so a statistical diagnostic cannot quietly become a marketing profitability claim.
A favourable conditional average can coexist with concentrated losses, long unresolved holdings or a funding constraint. Report chronological exposure, overlapping positions and the maximum-holding policy. A rule that adds positions as displacement grows must declare the resulting quantity path; its initial stop budget does not describe total later exposure. Averaging down is a changed action contract, not a statistical consequence of mean reversion.
Preserve adverse observations and still-open outcomes at the source cutoff. Do not keep extending a holding clock until an eventual recovery appears. Reference expectation half-life is not an executable deadline or a promise of first passage. Use clearly labelled scenarios for gaps and costs where actual evidence is unavailable. A scenario can show that the arithmetic is sensitive to an assumption, but it cannot establish the real frequency of that assumption.
Reference windows, thresholds, instruments and exits create a family of candidate experiments. Save rejected variants and the authorised selection objective. Freeze entry identity before applying a predeclared exit for economic validation; do not select entries using later exit-adjusted return and then call the resulting record independent. A reserved region inspected repeatedly becomes part of the research process rather than a fresh test.
Causal source availability matters even when timestamps are chronological. Corrected prices, corporate-action fields and later event labels can cross a decision boundary invisibly. Retain fitting scope and availability evidence along with the code. Evaluate a compatible baseline using the same opportunity population and accounting convention. There is no supplied empirical population in this guide, so no p-value, profit factor or strategy return is invented to fill the evaluation section.
A defensible assessment identifies which layers pass: software accounting, causal reference construction, population reconciliation, reserved economic evaluation and actual execution. Failure at one layer should not be disguised by strength at another. A flawless formula can be applied to incomplete history, while complete instructions can still produce an unattractive net result. State the unresolved layer and the evidence request that would address it.
Use the blank record to capture the actual specification and payoff convention before considering a seller's headline win fraction. Check the two fictional expectancy values separately from any market research. For a purchasing decision, inspect the exact offer, access conditions and record scope rather than relying on the generic method name. This page offers a way to ask a narrower, answerable question; it does not endorse a provider, allocate personal capital or certify that mean reversion will remain profitable.
| Object | Preserve | Do not infer |
|---|---|---|
| Frequency | Outcome definition and eligible count | A high win fraction guarantees profit |
| Economics | Payoffs, cost scope and capital convention | A recovery count is a net return |
| Validation | Frozen candidates and reserved region | Software fixtures validate a trading edge |
Download blank review record (CSV)
Headers only. No trades, subscriber outcomes or operating receipts are supplied.
Documentation establishes the stated mechanism, not tested trading performance. Checked 10 October 2026; a later version can differ.
Define a prior-window reference, measure absolute recovery and avoid calling every correctly directed move a successful reversion.
Understand positive AR(1) decay, observation clocks, coefficient sensitivity and why half-life is not a trade deadline.
Design a bounded stationarity experiment without turning a unit-root test, fitted spread or chosen chart into proof of trading profit.
The terminal recovery percentage divides a distance change by initial distance. It has no position quantity, cash base or exit policy. A band hit followed by a rebound can create different intermediate and terminal descriptions without any executed trade. The complete default path demonstrates this using invented prices only. Do not compound that percentage or treat it as a real win rate. Profitability requires the full native financial policy and independent research evidence, not a stronger adjective attached to a mathematically correct distance statistic.