Glosarium
Contents

Data Analysis › Time & Forecasting

Stationarity

When a series' statistical behaviour doesn't drift over time.

Data Analyst · Senior

Also known as: stationarity, stationary series, weak stationarity

A series is stationary when its average level, variability and correlations stay roughly constant over time — no trend, no widening swings, no changing seasons. Most classical forecasting machinery (ARIMA, and the theory behind autocorrelation plots) assumes it, which is why non-stationary data must be transformed first, usually by differencing: model the day-to-day changes instead of the levels.

levels (trending):     / / / / /      → not stationary
day-to-day changes:    ~ ~ ~ ~ ~      → roughly stationary, modellable

Differencing is not free information — it discards the level to stabilise the dynamics. Difference twice and you are modelling noise-of-noise. Transform the minimum needed (one difference, a log for stabilising variance), and check the result rather than assuming.

The classic mistakes:

  • Modelling raw trending levels. Autocorrelations near 1.0 at every lag, regressions with sky-high R² between unrelated trending series (spurious regression), forecasts that extrapolate drift forever. Difference first.
  • Over-differencing. Each difference injects noise and erases long-run information. If one difference stabilises it, stop.
  • Assuming variance is stable. Level-stationary with exploding variance (common in growth data) still violates the machinery. Logs or explicit volatility handling come before modelling.
  • Unit-root tests as oracles. Statistical tests for stationarity have weak power on short series and break under structural shifts. Plot the series and its differences; let the tests confirm, not decide.
  • Forgetting why it matters. Stationarity is not a ritual — it is the condition under which “patterns in the past describe the future” is mathematically licensed. Without it, every fitted pattern is suspect.

The working rule: plot, difference the minimum that stabilises, verify the differences look stable, then model. See time series for why order makes all of this load-bearing.