Statistics Glossary → Stationarity
Stationarityconcept
The property of a time series whose statistical characteristics, its mean, variance and lag-to-lag dependence, do not change over time.
Under weak (covariance) stationarity the mean and variance of the series are constant, and the covariance between two observations depends only on the lag between them, not on their dates. ARMA models, the interpretation of the autocorrelation function and many forecasting methods assume stationarity. Trends, changing variance, structural breaks and especially unit roots break it. In a series with a unit root (e.g. a random walk), shocks have permanent effects and the series does not revert to a mean.
Stationarity is examined with tests. In the ADF and Phillips-Perron tests the null hypothesis is 'there is a unit root', and in the KPSS test it is 'the series is stationary', so the tests are more informative when used together. Non-stationary series are usually made stationary by differencing (yₜ − yₜ₋₁), log transformation or modelling the trend; d and the seasonal D in ARIMA show how many times the series is differenced. A regression between two non-stationary series can give a high R² even without any real relationship (spurious regression). If the series are cointegrated, their relationship is studied with an error correction model.
An everyday example
Daily temperatures in a city, once seasonality is removed, fluctuate around a long-run average; a hot day does not cause a lasting change, and the series is stationary. A share price, by contrast, does not return to a fixed average: a fall today pulls all later prices down. That series is non-stationary, while its daily returns (the differences) usually are stationary.
In MerQur
Time series analysis gives the ADF, KPSS and Phillips-Perron unit-root and stationarity tests, and 'Differencing recommendation (ndiffs / nsdiffs)' gives the required number of differences. In the ARIMA forecast, 'Difference (d)' and 'Seasonal (P, D, Q) m' specify the differencing within the model, and 'Automatic model selection (auto.arima)' chooses d by a KPSS-based rule for the number of differences.
Analyses where this concept appears
Related concepts
Unit root tests (ADF, Phillips-Perron)KPSS stationarity testDifferencing