Can a stationary series have correlated observations?

Yes. The constructed autoregressive process has positive lag correlations.

Does stationarity require a zero mean?

No. Any finite constant mean is compatible with weak stationarity.

Will differencing always solve nonstationarity?

No. It addresses some forms of changing level, but other changes can remain and unnecessary differencing can introduce avoidable dependence.

AI Glossary / Time series and signal processing / Advanced

Stationarity

Definition

Stationarity is a property describing statistical stability over time. Strict stationarity means that shifting all observation times by the same amount leaves every finite joint distribution unchanged. Weak, or covariance, stationarity requires a constant mean, finite constant variance and covariance that depends only on time separation. The intended definition should be stated because these conditions are not interchangeable in every setting.

What it means

Stationarity does not mean that observed values remain constant or that successive observations are independent. A process can fluctuate substantially and still have stable statistical behavior. Conversely, a line that looks roughly horizontal can hide changing variability or dependence. The property concerns the process generating observations, while an actual dataset supplies only limited evidence about that process.

How it works

For weak stationarity, write the mean as E[X_t] = μ and the covariance as Cov(X_t, X_(t+h)) = γ(h). The lag h matters, but the calendar location t does not. For strict stationarity, the same invariance applies to entire joint distributions rather than just these moments.
Analysts examine time plots, variation, dependence and relevant domain changes. Transformations, detrending or differencing may address particular nonstationarities, but they are not universal repairs. Statistical tests assess specified hypotheses under assumptions; they do not certify that every future period will follow the same process. Stationarity assumptions must also match the model and the period over which it is used.

Three practical examples

A monitoring team checks whether a sensor's residual variation remains stable after accounting for operating conditions. A growing noise level matters even if the average reading stays unchanged.
A forecasting analyst examines whether sales differences behave more consistently than sales levels. The transformation changes the modeling target and must be reversed appropriately when forecasting levels.
An AI team evaluates a model across successive periods rather than relying only on a random split. A stable historical relationship is an assumption to investigate, especially after operational changes. These are illustrative evaluation choices, not findings from a deployed system.

Think of it this way

Imagine a game whose probabilistic rules remain the same while each play produces a different result. Stationarity concerns that stability of the rules under time shifts, not repetition of the same result. If the game quietly changes its variability or how one play influences the next, an unchanged average alone can conceal the change.

Worked example

Construct the process X_t = 0.5 X_(t-1) + ε_t, with independent Gaussian innovations of mean zero and variance three. For its stationary version, the mean is zero and the variance V satisfies V = 0.25V + 3. Solving gives V = four. Start with X_0 normally distributed with mean zero and variance four, independent of future innovations, so this initialization matches the stationary distribution.
The lag-one covariance is 0.5 times four = two, and the lag-two covariance is 0.25 times four = one. Corresponding correlations are 0.5 and 0.25. The process is stationary despite this dependence. Gaussianity and the time-invariant mean and covariance structure also give strict stationarity in this construction.
Now use the same transition equation but fix X_0 = 0. Variance at time one is three; at time two it is 0.25 times three plus three = 3.75. At time three it is 3.9375. The variances approach four but are not equal at these starting times. Stable transition coefficients alone therefore do not make every initialized sequence stationary from the outset.
For comparison, construct a random walk Z_t = Z_(t-1) + ε_t with Z_0 = 0 and the same independent innovations. Its mean remains zero, but variance is 3t: three, six and nine at times one, two and three. It is not weakly stationary. Its first differences equal ε_t and are stationary under the stipulated innovation assumptions.
Finally, add ten to every value of the stationary X process. The new mean is ten, variance remains four and lag covariances are unchanged. A nonzero constant mean does not violate stationarity. These are calculated properties of invented processes, not inferences from a short observed chart.

Illustrative case study

Fictional Harbor Forecasting uses the constructions above in a model-review workshop. A participant argues that the random walk is stationary because its expected value stays at zero. The analyst calculates its expanding variance and contrasts it with the initialized autoregressive process.
The team records which definition of stationarity each modeling step requires. For a real dataset, it plans to examine several periods, transformations and operational explanations before selecting a specification. It does not treat the constructed example as proof that differencing is always necessary or that an autoregressive model always starts in equilibrium. This scenario demonstrates careful reasoning rather than validated forecasting improvement.

Common misconceptions and mistakes

A constant mean and variance are insufficient for weak stationarity if lag covariances change with time. Stationarity does not imply independence, causality or perfect predictability. Strict stationarity need not provide finite second moments; with finite second moments it implies weak stationarity. The converse is not generally true outside settings such as Gaussian processes. A failed rejection in a test can reflect limited information rather than strong evidence that a process is stationary.

Business use

State the stability assumptions behind a forecast, alert threshold or learned relationship. Monitor whether location, variability or dependence changes after a new policy, product or operating regime. Compare candidate transformations using both diagnostics and the intended prediction task. Preserve time order in evaluation and explain what future conditions are assumed. Stationarity can make historical information more reusable, but a mathematical assumption is not a guarantee against later distribution changes.

Check your understanding

Why does the zero-start autoregressive example fail exact stationarity during its initial periods?

Answer: Its variance changes from zero to three, then 3.75 and onward toward four. The transition rule stays fixed, but the initialized distribution is not yet the invariant distribution.

Related terms

Stationary process
Autocorrelation
Random walk
Differencing
autocovariance
unit root
distribution shift
ergodicity

Learn next

Change the stationary autoregressive coefficient to 0.8 while keeping innovation variance three. The stationary variance becomes 3/(1 - 0.64) = 25/3, approximately 8.3333. Lag-one covariance becomes 20/3, while lag-one correlation is 0.8. Greater persistence can coexist with stationarity when the initialization is appropriate.

Sources

Hyndman and Athanasopoulos, Forecasting: Principles and Practice, Stationarity and differencing: https://otexts.com/fpp3/arima.html NIST/SEMATECH e-Handbook, Stationarity: https://www.itl.nist.gov/div898/handbook/pmc/section4/pmc442.htm The autoregressive and random-walk constructions, arithmetic and Harbor Forecasting scenario are original teaching examples. Their distributional assumptions are explicit.

Previous: Static quality scores Next: Stationary distribution