Skip to content

Stationarity

Version
v1 · 2026-08-24 · History
Prime #
53
Domain group
Formal Sciences
Origin domain
Experimental Design & Statistics
Also from
Physics, Economics & Finance
Related primes
Invariance, Variability, Randomness, Equilibrium

Core Idea

A property of a process whose statistical characteristics (mean, variance, autocorrelation) do not change over time or space.

How would you explain it like I'm…

Same pattern over time

Imagine you flip the same coin every day for a year. Some days you get more heads, some days more tails — but the coin itself doesn't change. It's still a 50-50 coin. Stationarity means the rules of the game stay the same, even though each round looks a little different.

Rules-don't-change

Picture rolling a six-sided die over and over. Each roll is different, but the chances of getting each number stay the same every time you roll. That sameness — the rules don't shift — is called stationarity. The weather in a city across many years might be roughly stationary if average temperatures and rainfall patterns stay about the same, but if the climate is warming, it stops being stationary because the rules are changing.

Stable statistical rules

Stationarity is a property of a random or time-varying process where the statistical rules generating the data stay the same over time. Individual measurements still fluctuate — temperatures vary day to day, stock prices wobble — but the underlying probability distribution (the average, the spread, how values relate to each other across time) doesn't drift. If you measured the mean and variance in one decade versus another, you'd get roughly the same numbers. Stationarity matters because most statistical and forecasting tools assume it; when it fails — say, during a financial crisis or climate shift — those tools give misleading answers and you need different methods.

 

Stationarity is the property of a stochastic process or time-varying system whose statistical characteristics — mean, variance, autocorrelation, and higher moments — remain invariant over time or across translations along the relevant dimension. The essential commitment is that while individual realizations fluctuate, the generating rules do not drift: the distribution governing outcomes this year is the same as last year, in one region the same as another. Every stationarity claim specifies (1) the process or quantity whose statistics are being assessed, (2) the notion of stationarity being invoked — strict (the full joint distribution is time-invariant), wide-sense (only mean and autocovariance are time-invariant), or cyclostationary (periodic invariance), (3) the temporal or spatial window over which the claim holds, and (4) the tests or evidence supporting or challenging it (augmented Dickey-Fuller, KPSS, structural-break tests). Stationarity is almost always an approximation valid on some scale and invalidated by regime change on another — which is why it is one of the most consequential and frequently violated working assumptions in time-series analysis, signal processing, and econometrics.

Broad Use

  • Climatology: Checking if historical climate data patterns are stable enough for forecasting.

  • Economics: Testing if market returns are stable over time to apply certain forecasting models.

  • Signal Processing: Assuming constant signal properties for simpler analysis and filtering.

  • Ecology: Population dynamics might be treated as stationary for certain intervals if conditions remain unchanged.

Clarity

Distinguishes between systems that can be analyzed with steady parameters and those that undergo shifting baselines.

Manages Complexity

Allows for simpler modeling and statistical methods when stationarity holds, avoiding dynamic re-calibration.

Abstract Reasoning

Encourages identifying breaks from stationarity (regime shifts) and rethinking models accordingly.

Knowledge Transfer

Applicable to time-series analysis across domains, from hydrology to financial risk.

Example

River Flow Data: Assumed stationary for water resource planning, though climate change may break this assumption.

Relationships to Other Abstractions

Local relationship map for StationarityParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.StationarityPRIMEPrime abstraction: Probability — presupposesProbabilityPRIMEPrime abstraction: Time — presupposesTimePRIMEPrime abstraction: Invariance — is a kind ofInvariancePRIMEDomain-specific abstraction: Restoration Baseline Error — includesRestorationBaseline ErrorDOMAINDomain-specific abstraction: Autoregressive Integrated Moving Average — presupposesAutoregressive …DOMAINDomain-specific abstraction: Stationary process — is a kind ofStationaryprocessDOMAINDomain-specific abstraction: Trend-stationary process — is a kind ofTrend-stationaryprocessDOMAIN

Current abstraction Stationarity Prime

Parents (3) — outgoing typed relations

  • Stationarity is a kind of Invariance Prime

    Stationarity is a specialization of invariance whose preserved feature is a process's statistical distribution under time (or spatial) translation.

  • Stationarity presupposes Probability Prime

    Stationarity presupposes probability because the invariance claim is about the joint distribution of the process under temporal translation.

  • Stationarity presupposes Time Prime

    Stationarity presupposes time because its claim is precisely that statistical characteristics remain invariant across translations along the temporal dimension.

Children (4) — abstractions linked here by typed relations

  • Stationary process Domain-specific is a kind of Stationarity

    It remains its own entry because its identity is fixed by the index set and process, strict or weak convention, finite-dimensional distributions or moments, allowed shifts, mean and covariance existence and any ergodicity claim.

  • Trend-stationary process Domain-specific is a kind of Stationarity

    What makes it its own entry: the domain-specific identity determined by the trend class, residual stationarity notion, structural-break treatment, and shock-reversion claim are explicit and unit-root behavior is rejected by evidence.

  • Autoregressive Integrated Moving Average Domain-specific presupposes Stationarity

    The ARMA component requires stationary behavior of the transformed time series.

  • Restoration Baseline Error Domain-specific includes Stationarity

    The error contains an assumption that the condition-generating process remains stable enough for the historical reference to stay viable.

Hierarchy paths (3) — routes to 3 parentless roots

Not to Be Confused With

  • Stationarity is not Statistical Inference because Stationarity is the property that a stochastic process's distribution does not change over time, while Statistical Inference addresses methods for drawing conclusions about populations from samples.
  • Stationarity is not State and State Transition because Stationarity concerns the temporal stability of a process's statistical properties, while State and State Transition concerns the specification of system conditions and rules governing changes.
  • Stationarity is not Equilibrium because Stationarity means the statistical distribution does not change but the underlying values can vary widely, while Equilibrium means the system settles at a fixed point or cycle and stays there.