Stationarity Assumption

Weak stationarity constrains selected features, specifically the mean, variance, and lag-dependent covariance. Strict stationarity imposes a stronger requirement: the entire joint distribution must remain unchanged when the time origin shifts. This distinction matters because a model may satisfy stable first- and second-moment behavior without preserving every aspect of its distribution, whereas strict stationarity demands broader temporal consistency.

Lag-based covariance makes relationships comparable across different portions of a time series. Observations separated by the same interval should exhibit the same covariance pattern, whether they occur early or late in the record. That consistency allows historical dependence structures to inform forecasting and modeling, while position-specific covariance signals that the process may behave differently as time progresses.

These features introduce systematic changes that make historical behavior less representative of later observations. A trend shifts the process level, seasonality creates recurring time-dependent patterns, and changing variance alters the scale of fluctuations. Each violation can undermine analyses based on stable statistical behavior, so the selected model must account for the specific form of instability present.

Analysts can compare whether the process maintains a consistent mean and variance and whether covariance is determined by lag rather than by the observations' positions in time. For a stronger assessment, they must consider whether the full joint distribution remains unchanged under time shifts. These comparisons connect the diagnostic question directly to the chosen form of stationarity.

The appropriate adjustment depends on the source of instability. Differencing can address changes that accumulate over time, detrending can remove systematic movement in the level, and transformations can help when variability changes. If these adjustments do not provide a suitable representation, analysts may instead select models specifically designed for nonstationary processes.

It is especially useful when analysts want patterns estimated from historical observations to represent future behavior. Stable statistical features provide a basis for carrying dependence and variability learned from the past into forecasting. When trends, seasonal effects, or changing variance violate that premise, preprocessing or a model built for nonstationary processes becomes more appropriate than relying on an unchanged historical pattern.