Stationarity Check¶
Diagnostic test — instantiates Stochastic Process Envelope Modeling
Tests whether a process's statistical properties are holding still or shifting over time, delivering a verdict on the stationarity assumptions a model rests on.
Most time-series models quietly assume the process is stationary — its mean, variance, and dependence do not change with time. Stationarity Check is the diagnostic that puts that assumption on trial. It first registers which stationarity conditions the downstream model actually needs, then runs formal tests — unit-root tests for trend, variance-break tests for shifting spread, distributional comparisons across windows — and returns a verdict: does the assumption hold, and if not, which part fails? Its defining idea is that it judges, it does not fix: the output is a diagnosis and a named violation, never a recalibrated model. That restraint is deliberate — a check that also repaired the model would blur the moment of decision, and the value here lies precisely in producing a clean verdict that a separate mechanism can then act on.
Example¶
Before fitting a volatility model to a series of daily asset returns, an analyst runs a Stationarity Check. She registers the conditions the model needs: mean near zero and stable, variance regime constant, autocovariance time-invariant. She runs an Augmented Dickey-Fuller test for a unit root, then compares variance across pre- and post-crisis sub-periods. The mean comes back stationary, but variance clearly clusters — quiet stretches punctuated by turbulent ones — and there is a sharp break at a known crisis date. The verdict is unambiguous: the series is not covariance-stationary, and the specific offender is the stable-variance assumption. That single finding redirects the modeling from a plain fit toward a regime-aware volatility model, with the violated assumption documented rather than discovered later as a mystery forecasting failure.
How it works¶
- Register the required assumptions. Enumerate exactly which stationarity conditions the downstream model depends on: constant mean, stable variance, time-invariant autocovariance, or full distributional stationarity.
- Test each condition. Unit-root tests for trend and mean; variance-break or CUSUM tests for shifting spread; two-sample distribution tests comparing early and late windows.
- Classify the failure. Distinguish trend-stationary from difference-stationary from an abrupt regime break — they call for different remedies.
- Return a verdict, not a repair. Output holds/fails plus the named violated assumption, and hand that verdict to whatever mechanism acts on it.
Tuning parameters¶
- Segmentation scheme — how the timeline is split to compare windows; finer splits catch local shifts but lose power per segment.
- Test battery — which tests to run (e.g., ADF, KPSS, a break test); complementary tests guard against any single test's blind spot.
- Significance level — the false-positive versus missed-nonstationarity trade; a strict level rarely cries wolf but lets slow drift pass.
- Assumption scope — strict stationarity versus weak (covariance) stationarity; testing more than the model needs wastes power.
- Break sensitivity — tuned toward abrupt structural breaks or toward slow continuous drift.
When it helps, and when it misleads¶
Its strength is catching the assumption violation that silently invalidates a forecast before the forecast is trusted, and leaving behind an explicit ledger of what the model requires. That ledger is what lets a later reviewer know exactly which condition to re-test as data accumulates.
Its failure mode is low power: in short samples these tests can fail to reject nonstationarity that is really present, or spuriously flag it when it is not. The associated classic trap is spurious regression[1] — two independent trending series can look strongly related, which is precisely why stationarity must be tested before any relationship is modeled. A further misuse is running one test once, reading "stationary," and never rechecking as the process evolves. The guarding discipline is to run several complementary tests, register the assumptions explicitly, and re-run the check on a cadence rather than treating one verdict as permanent.
How it implements the components¶
drift_and_shift_test— the battery of unit-root, variance-break, and distribution-shift tests that constitutes the check.stationarity_assumption_register— the explicit ledger of which stationarity conditions the downstream model depends on, recorded so violations are named rather than vague.
It stops at the verdict: update_filtering_rule and model_validity_boundary — recalibrating or retiring the model in response — belong to Drift Recalibration Loop, the actuator this check feeds.
Related¶
- Instantiates: Stochastic Process Envelope Modeling — this check tests the stationarity assumptions the whole envelope model leans on.
- Sibling mechanisms: Drift Recalibration Loop · Innovation Residual Monitor · Sequential Filter Update · Markov Chain Model · State-Transition Kernel · Poisson Event Model · Prediction-Interval Fan Chart · Trajectory Ensemble Simulation · Stochastic-Process Diagram
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Stationarity Check operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it tests whether a process's statistical properties are holding still or shifting over time, delivering a verdict on the stationarity assumptions a model rests on.
Independent corroboration: The frozen evidence defines Stationarity Check as 'Tests whether a process's statistical properties are holding still or shifting over time, delivering a verdict on the stationarity assumptions a model rests on', so its operative form is Analysis, Modeling & Optimization.
Nearest alternative: Assessment, Review & Assurance — Stationarity Check includes features of a bounded evaluation of existing evidence or work that produces a finding or disposition, but its defining operation is an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Testing time-invariance of process properties is time-series statistics.
Related originating lineages:
- Data Science & Analytics — Drift invalidates models.
- Economics & Finance — Economic series often violate stationarity.
- Mathematics — Mathematical modeling, proof, and abstract-structure practice supplies a parallel or contributing lineage for the mechanism's defining operation: tests whether a process's statistical properties are holding still or shifting over time, delivering a verdict on the stationarity assumptions a model rests on.
Review resolution: The blind reviewers agree that statistics_experimental_design is the primary origin and differ only on alternate origin disagreement, domain reach disagreement. I preserve every independently explained alternate from both records rather than imposing a numeric cap. I retain single_lineage because the combined evidence shows one traceable formative lineage. The broader reach of multi_domain records portability separately from historical provenance; encyclopedia_synthesis=false preserves the affirmative synthesis judgment where either reviewer identified one.
Review outcome: Reconciled after independent review; high confidence.
References¶
[1] Granger, C. W. J., & Newbold, P. "Spurious Regressions in Econometrics". Journal of Econometrics 2(2), 111–120 (1974). Granger and Newbold show that independent autocorrelated series can yield apparently strong, significant regressions. registry ↩