Can stationary observations be correlated?

Yes. The worked autoregressive example has positive serial correlation.

Does strict stationarity require finite variance?

No. Finite moments must be added when needed.

Does differencing always solve nonstationarity?

No. Its usefulness depends on the type of change and the resulting model.

AI Glossary / Statistics and probability / Intermediate

Stationary sequence

Definition

A stationary sequence is a sequence of random variables whose relevant statistical properties remain unchanged when the time index is shifted. Strict stationarity requires every finite collection to have the same joint distribution after a common time shift. Weak, or covariance, stationarity requires finite second moments, a constant mean, and covariances that depend only on the lag between observations. These definitions concern a probability model, not a requirement that the observed values stay constant or that successive observations be independent.

What it means

Stationarity allows information from different parts of a sequence to be combined under a stable model. For example, the relationship between observations one hour apart can be treated as the same whether those observations occur early or late in the modeled period.

This stability is stronger than merely having the same average. A series can have a constant mean while its variance changes or its dependence pattern evolves. Conversely, a stationary process can show runs, cycles, and persistent dependence in a realized sample. Its observations need not look like unrelated random scatter.

How it works

Begin by stating which definition is needed. Strict stationarity is expressed through joint distributions for all finite sets of time points. Weak stationarity focuses on the first two moments: E(X_t) = mu and Cov(X_t, X_(t+h)) = gamma(h), with finite variance gamma(0).

Strict stationarity with finite second moments implies weak stationarity. Without those moment conditions, a strictly stationary process need not have a finite variance. Weak stationarity alone does not generally imply strict stationarity, although it does for a jointly Gaussian process because means and covariances determine its finite-dimensional distributions.

For observed data, inspect plots, changing spread, seasonality, structural breaks, and dependence patterns. Model-based tests can provide evidence about particular departures, but no finite dataset proves unrestricted stationarity. Transformations, detrending, or differencing may help with some forms of nonstationarity, but should be justified by the process and checked afterward. Distinguish stationarity of the modeled residual sequence from stationarity of the original measurements.

Three practical examples

1. A forecasting analyst models demand fluctuations around a justified stable baseline while checking whether promotions have changed the underlying process.

2. A sensor-monitoring team examines whether residual variability and autocorrelation remain similar across operating periods.

3. A simulation researcher initializes a process in its stationary regime to study dependence without an avoidable startup transient.

Think of it this way

Think of a game whose rules stay the same from round to round. The outcomes can vary and influence subsequent outcomes, but moving the observation window does not change the probabilistic rules governing it. Seeing a lucky streak does not by itself prove that the rules changed.

Worked example

Consider an original autoregressive model X_t = 0.5 X_(t-1) + e_t, where the innovations are independent with mean zero and variance 3. Assume the process is initialized in its stationary regime, with the innovations independent of the past.

Its mean is zero. If V is the stationary variance, then V = 0.5 squared times V + 3. Thus 0.75V = 3 and V = 4. The lag-one covariance is 0.5 times 4 = 2, and the lag-two covariance is 0.5 squared times 4 = 1. More generally, the autocovariance is 4 times 0.5 to the power |h|, depending on the lag h rather than the calendar time.

The observations are therefore correlated despite weak stationarity. If instead X_0 is fixed at zero, the variance of X_1 is 3 and the variance of X_2 is 0.25 times 3 + 3 = 3.75. This startup sequence is not stationary from the outset, even though its variance approaches 4.

For comparison, a random walk Y_t = Y_(t-1) + e_t starting at zero has variance 3t. Its constant zero mean does not make it weakly stationary because its variance grows with time. These are constructed models, not empirical findings about a business series.

Illustrative case study

In a fictional online retailer, analysts fit a model to weekly residual demand after accounting for known seasonal effects. They initially assume that the residual process is stable throughout the year.

A major delivery-policy change produces a lasting shift in both average residual demand and variability. The team separates the operating regimes and reassesses its forecasting assumptions rather than relying on one pooled variance estimate. It also distinguishes a short-lived fluctuation from a persistent structural change by combining statistical evidence with the documented business event.

Common misconceptions and mistakes

Do not equate stationarity with independence, a flat observed graph, or a constant mean alone. A series can have the same marginal distribution at each time while its joint dependence changes, so equal one-time distributions are insufficient for strict stationarity.

Stationarity is also different from ergodicity. For example, X_t = Z for every t, with one random Z drawn for the entire sequence, is strictly stationary. Yet a single long trajectory reveals only that one value, so its time average need not equal the population mean. A stationary model can therefore require additional assumptions before time averages reliably estimate population features. Repeated differencing without justification can create unnecessary complexity and distort interpretation.

Business use

Stationarity assumptions support time-series forecasting, monitoring, and uncertainty estimation. Their practical value is enabling stable relationships to be learned from history. Businesses should specify the operating regime and time horizon over which that stability is plausible, monitor changes, and revise models when policies or environments shift. Passing one diagnostic should not override clear evidence that the data-generating process has changed.

Check your understanding

A sequence has mean zero at every time, but its variance equals t. Is it weakly stationary?

Answer: No. Weak stationarity requires constant finite variance as well as a constant mean and lag-dependent covariance.

Related terms

Stochastic process
Autocorrelation
Random walk
Structural break
strict stationarity
covariance stationarity
autocovariance
ergodicity

Learn next

Learn autocovariance and autoregressive models, then compare stationarity with ergodicity. Examine how initialization, deterministic trends, and unit roots affect time-series assumptions.

Time series

Sources

Roger Peng, A Very Short Course on Time Series Analysis, Stationarity: https://bookdown.org/rdpeng/timeseriesbook/stationarity.html

Hyndman and Athanasopoulos, Forecasting: Principles and Practice, ARIMA models: https://otexts.com/fpp3/arima.html

The primary author material was consulted. Moment conditions are stated explicitly. The autoregressive, random-walk, and constant-random-sequence constructions are original teaching examples.

Previous: Stationary process Next: Statistic