Residual Stationarity Check¶
Diagnostic test — instantiates Trend Detection and Removal
Checks whether residuals after trend handling are stable enough for the intended analysis.
A Residual Stationarity Check is the acceptance test that runs after trend removal: it asks whether what is left behind — the residual, the detrended series, the seasonally-adjusted line — is stable enough to carry the downstream analysis. "Stable enough" means its statistical character no longer drifts: roughly constant mean, roughly constant variance, no leftover systematic direction. Its defining move is that it never looks at the raw series and never proposes a break or a cut; it takes the output of some other mechanism and issues a pass/fail verdict on the residual. If the residual still trends, the check fails it and sends the analysis back — because a forecast, a control limit, or a regression built on a still-drifting residual will quietly mislead. It is the gate, not the surgery.
Example¶
An epidemiology team monitors weekly counts of a notifiable infection. A long, slow decline runs through the record as vaccination coverage rises, so before they can flag genuine outbreaks, they detrend — subtracting the fitted decline to leave residual deviations. The question is whether those residuals are now safe to feed into an outbreak-detection control chart. The Residual Stationarity Check applies a formal test: an Augmented Dickey–Fuller test to probe for a remaining unit-root drift, cross-read against a KPSS test whose null is the opposite.[1] On an illustrative run, ADF rejects the drift and KPSS fails to reject stationarity — the two agree the residual is stable — so the residual is cleared for the control chart. Had the residual still trended, the chart would have thrown a rash of false outbreak alarms every time the baseline wandered past a threshold, and the check would have blocked it.
How it works¶
What distinguishes it from every mechanism upstream of it is that it tests an output rather than producing one:
- Take the residual as given — the detrended or adjusted series from whatever handled the trend.
- Specify what "stable enough" means — constant mean, constant variance, no autocorrelated drift, at the tolerance the downstream use demands.
- Apply a formal test — a unit-root or stationarity statistic, ideally a pair with opposing nulls so agreement is informative and disagreement is a flag.
- Return a verdict, not a fix — pass, fail, or "borderline; re-handle the trend"; the remedy belongs to the upstream mechanism.
Tuning parameters¶
- Test choice — unit-root (ADF, Phillips–Perron) versus stationarity (KPSS); pairing them guards against either one's blind spot.
- Significance level — how much residual drift is tolerated before the verdict flips to fail.
- Which moments to check — mean-stability only, or variance and autocorrelation structure too.
- Segment vs. whole — test the full residual or test within regimes when a break is suspected upstream.
- Sample-length sensitivity — how the verdict is stress-tested against the window it was computed on, since short samples under-power the test.
When it helps, and when it misleads¶
Its strength is that it converts "the detrending looks fine" into a checkable claim, catching the case where a trend was under-removed and left a residual that will corrupt whatever consumes it.
Its failure modes are the well-known ones for stationarity testing. These tests are low-powered on short series — they often fail to reject a unit root simply for lack of data, so a "non-stationary" verdict can be an artifact of a short window rather than a real drift. They are also blind to structure they do not test for: a residual can pass a mean-stationarity test while hiding a changing variance or a seasonal cycle. The classic misuse is to treat a single passing test as a certificate of a clean decomposition. The discipline is to pair opposing-null tests, report the verdict with its window length, and read a pass as "no drift detected at this power," not as proof of stationarity.
How it implements the components¶
Residual Stationarity Check fills the validation slice of the archetype's machinery — the check, not the cut or the fit:
stationarity_diagnostic— its primary act: the formal test of whether the adjusted residual is stable enough for its intended use.residual_pattern_target— it operates on the designated residual as the object under test, certifying (or rejecting) it as the representation the downstream analysis will consume.
It does not locate structural breaks (change_point_marker) or frame a piecewise trend_hypothesis — flagging regimes that should be segmented rather than smoothed is Change-Point Detection Test, its nearest twin, which scrutinizes the raw series for breaks whereas this check scrutinizes the residual for drift; nor does it estimate or remove any trend itself (Differencing Transform, Rolling-Window Trend Estimate).
Related¶
- Instantiates: Trend Detection and Removal — the gate that certifies a detrended residual before downstream inference relies on it.
- Consumes: the residual output of whichever mechanism handled the trend — e.g. Differencing Transform or Rolling-Window Trend Estimate.
- Sibling mechanisms: Change-Point Detection Test · Differencing Transform · Rolling-Window Trend Estimate · Decomposition Plot · Seasonal Adjustment Procedure
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Residual Stationarity Check operates by computes whether an already-derived residual series satisfies defined mean, variance, and dependence stability conditions. That concrete deployed or enacted form is Analysis, Modeling & Optimization under the frozen taxonomy.
Nearest alternative: Assessment, Review & Assurance — Although Assessment, Review & Assurance can support this mechanism, the frozen evidence makes its operative form the act that computes whether an already-derived residual series satisfies defined mean, variance, and dependence stability conditions; the alternative is therefore secondary rather than defining.
Review outcome: Adjudicated after independent review; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Checking post-detrending residual stationarity is a canonical time-series statistical diagnostic.
Review resolution: Both blind reviewers agree that statistics_experimental_design is the primary historical origin. Explicit reconciliation of alternate origin disagreement, domain reach disagreement adopts reviewer_a's evidence: Checking post-detrending residual stationarity is a canonical time-series statistical diagnostic. The selected record uses alternates=none, origin_mode=single_lineage, and domain_reach=multi_domain; the other review proposed alternates=data_science, mathematics, origin_mode=single_lineage, and domain_reach=specialized. The selected combination better preserves the mechanism-specific formative lineages and calibrated scope; broader present-day use is not treated as proof of additional historical origin.
Review outcome: Reconciled after independent review; high confidence.
References¶
[1] Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. “Testing the Null Hypothesis of Stationarity against the Alternative of a Unit Root: How Sure Are We That Economic Time Series Have a Unit Root?”. Journal of Econometrics 54(1–3), 159–178 (1992). Introduces a formal stationarity-null test and interprets it alongside Dickey–Fuller unit-root tests, whose null hypothesis runs in the opposite direction. registry ↩