Strict Stationarity

When you model data that changes over time—like daily sales or hourly sensor readings—you usually want the “rules of the process” to stay the same. Strict stationarity is the strongest way to say exactly that.

What it means (the technical core)
A time series is strictly stationary if its full probabilistic behavior does not change over time. Formally, for any set of time points and any shift \(k\), the joint distribution of \((X_{t1}, X_{t2}, \dots, X_{tn})

How it differs from “weak” stationarity
You’ll often hear about weak (covariance) stationarity, which only requires:

constant mean over time,
constant variance over time,
autocovariance depends only on lag, not on the calendar time.

Strict stationarity implies weak stationarity (when moments exist), but not vice versa. A series can have stable mean/variance yet still have changing tail behavior or changing higher-order dependencies—violating strict stationarity.

Intuition and a concrete example
Think of strict stationarity as “the same dice, rolled the same way, every day.” For example:

Daily exam scores drawn from the same unchanging distribution each semester could be strictly stationary.
House prices that keep the same average growth but occasionally enter “boom/bust” regimes are typically not strictly stationary, because the distribution shifts.

Why it matters in AI/ML
Many time-series tools assume stability over time. If strict stationarity fails, models can misestimate uncertainty and dependencies:

ARMA/ARIMA and many forecasting baselines rely on stationarity ideas.
In ML, non-stationarity shows up as concept drift, hurting generalization for online prediction.
Volatility clustering in finance motivates models like GARCH, which relax strict stationarity assumptions.

Strict Stationarity means a time series’ full joint probability distribution is unchanged by shifting time: any collection of values (e.g., at t1, t2, …) has the same distribution as at (t1+k, t2+k, …) for any k. It is stronger than weak stationarity because it constrains all moments and dependencies, not just mean and autocovariance. This matters for valid inference and model assumptions. Example: i.i.d. noise is strictly stationary.

Imagine a movie set where every scene has the same “feel”: the lighting, background noise, and action patterns don’t change as time passes. That’s the intuition behind strict stationarity.

In time series data (numbers recorded over time, like hourly temperatures or stock prices), strict stationarity means the data’s behavior is completely unchanged by shifting the clock. Not just the average or the typical ups and downs, but the entire pattern of randomness stays the same at all times. If you look at any stretch of the series, it should “look statistically the same” as any other stretch of the same length.