Stationarity

Stationarity describes a time series that behaves the same way, statistically, no matter which stretch of it you look at. Its average does not trend up or down, its spread does not widen or shrink, and the way each point relates to the ones before it stays steady. Slide a window across a stationary series and each window looks like a sample from the same process. The everyday intuition is a series with no trend, no seasonal march, and no swelling volatility — the fluctuations wobble around a fixed level. A non-stationary series, by contrast, drifts. Think of a stock price wandering upward for years, a subscriber count climbing each quarter, or a variance that explodes during a crisis. Strictly, all the statistical moments must be constant; in practice, people usually mean the weaker version where the mean, variance, and autocorrelation hold steady.

Stationarity matters because a great many forecasting and econometric methods assume it. Classic models like ARMA are built on the idea that the relationships they estimate today will still hold tomorrow, and that only works if the series is stationary. Fit such a model to a drifting series and it can produce a spurious result, appearing to find structure that is really just two things trending together. That is why testing for stationarity is a routine first step in time-series work, often with statistical tests such as the augmented Dickey-Fuller test. If a series is not stationary, analysts usually transform it until it is, most commonly by differencing — modeling the change from one period to the next rather than the raw level — or by removing a trend or seasonal component. The stationary residual is what the model then works on.

Stationarity is best understood against the things that break it, trend and seasonality. A trend is a persistent drift in the mean, the series climbing or falling over the long run, which makes the average depend on when you measure it. Seasonality is a repeating pattern tied to the calendar — higher sales every December, more traffic every weekend — which makes the statistics swing on a fixed cycle. Both violate stationarity because the series no longer looks the same across different windows. The remedy is to strip them out. Differencing removes a trend by modeling period-to-period change, and seasonal differencing or seasonal adjustment removes the repeating cycle. What is left, if the work is done properly, is a stationary series whose behavior no longer depends on the calendar.

It helps to separate stationarity from the related idea of ergodicity and from mere smoothness. Stationarity is about the statistical properties staying constant over time; it does not mean the series is flat or unchanging. A stationary series still fluctuates — sometimes violently — it just fluctuates around a stable mean with stable spread. Nor does stationarity require the series to be independent from one point to the next. Stationary series can have strong autocorrelation, as long as that correlation structure itself does not change over time. The single question stationarity asks is whether the process generating the data is stable. If a shock permanently shifts the level, or volatility ratchets up and stays up, the series is non-stationary even though it may look calm in places.

Working with stationarity well starts with checking for it before you model, not after. Plot the series and look for trend, seasonality, and changing spread; then confirm with a formal test rather than eyeballing alone. If the series is non-stationary, transform it deliberately — difference it to remove a trend, take seasonal differences or adjust for the calendar, and consider a log transform when the variance grows with the level. Model the stationary version, then translate the forecasts back to the original scale. Keep a record of every transformation, because a forecast of differenced, logged data means nothing until it is undone. The aim is a series whose statistical behavior is stable enough that the patterns a model learns will still apply to the periods you want to predict.

The failures cluster around ignoring stationarity or over-treating it. Fitting a model to a trending series without checking can produce spurious correlations and forecasts that look precise but rest on a drift that will not continue. Over-differencing is the opposite mistake — differencing a series that was already stationary adds noise and destroys real structure. Assuming stationarity holds forever is risky too, because a regime change — a new pricing model, a pandemic, a policy shift — can break a stationarity that held for years, so the property should be rechecked as new data arrives. Stationarity is not a permanent label; it is a claim about the data-generating process that stays true only while that process does. Treat it as a working assumption to be tested, not a fact to be assumed.

Stationarity comes from stationary, from Latin stationarius (standing still), applied in statistics to a process whose properties stand still over time.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where stationarity is a core concern: