Stationarity describes a time series whose statistical behavior remains stable under a specified definition. In financial research, the most common practical version is weak or covariance stationarity: the mean is constant, the variance is finite and constant, and covariance between observations depends on the lag between them rather than the calendar date.
Strict and weak stationarity are not the same
Strict stationarity requires the joint distribution of the process to be unchanged by a shift in time. Weak stationarity imposes conditions on the first two moments and autocovariance structure. A series can satisfy one practical modeling requirement without satisfying every stronger definition.
That distinction matters because saying simply “the series is stationary” can hide what property was actually tested.
Why stationarity matters to a model
Many time-series methods estimate relationships from historical observations and assume those relationships remain sufficiently stable for inference or forecasting. If the mean, variance or dependence structure changes materially, parameters estimated from one period may not describe another.
Stationarity is therefore an assumption about the data-generating process, not a label that makes a strategy profitable.
Price levels and returns can behave differently
Asset price levels can contain trends or stochastic persistence while transformed return series can have different statistical properties. Researchers should test the variable actually used in the model rather than making a blanket statement about “the market.”
Simple Returns vs Log Returns explains one common transformation from prices to returns.
Visual stability is not enough
A chart that looks flat can still contain dependence or changing variance, and a series that looks volatile can still satisfy a particular weak-stationarity model. Plots are useful diagnostics, but formal tests and model residual checks provide additional evidence.
Unit-root tests answer a narrower question
Tests such as the Augmented Dickey-Fuller family are commonly used to investigate a unit-root null hypothesis. Rejecting or failing to reject that null is not a complete test of every form of stationarity.
Test results depend on specification choices such as deterministic terms, lag length, sample period and structural breaks. A failure to reject is not proof that a unit root must be the true data-generating process.
Variance can change even when the mean is stable
Financial return series may display periods of calm and turbulence. That changing conditional variance is one reason Volatility Modelling and GARCH Models exist.
A model can treat the unconditional process and the conditional variance structure differently; the exact assumptions should be stated rather than compressed into one stationarity label.
Structural breaks can make a historical test misleading
Policy changes, market structure, crises, new participants or methodology changes can alter a series. A test over the full sample may average across periods that behave differently.
Market Regime Detection studies model-dependent ways to describe changing states, while the broader idea of market regimes describes how market behaviour can differ across conditions.
Differencing can change the question
Taking differences is a common way to remove certain forms of trend or persistence, but each transformation changes what the variable means. First differences of prices are not the same as percentage returns, and repeated differencing can remove economically useful long-run information.
Validate stability out of sample
Even if a model passes historical diagnostics, the relevant question is whether its assumptions remain useful on later data. Time-Series Cross-Validation and Walk-Forward Analysis provide ordered evaluation frameworks.
Stationarity is best treated as a stated modeling assumption that must be checked against the exact variable, definition and sample. It is not a universal property of prices, returns or financial markets.