MerQuR+Statistics Glossary

Statistics Glossary → KPSS stationarity test

KPSS stationarity testconcept

A time-series test whose null hypothesis is that the series is stationary (around a level or a trend) and whose alternative is a unit root.

The test of Kwiatkowski, Phillips, Schmidt and Shin regresses the series on a constant (level stationarity) or a linear trend (trend stationarity) and examines the partial sums of the residuals. If the series is stationary, these cumulative sums stay bounded; with a unit root, they keep growing. The LM statistic is computed from the squares of these sums divided by an estimate of the long-run variance. The long-run variance is estimated with a window (number of lags) that accounts for autocorrelation in the residuals, and this choice can affect the result.

KPSS complements the ADF and Phillips-Perron tests, whose null hypothesis is a unit root. If ADF rejects the unit root and KPSS does not reject stationarity, the series is taken as stationary; the reverse pattern points to a unit root. If both reject or neither does, the conclusion is unclear, and structural breaks, fractional integration or low power should be considered. A non-significant KPSS result does not prove stationarity.

An everyday example

Forty years of a country's annual inflation are examined. The ADF test cannot reject a unit root (p = 0.31), while the KPSS level-stationarity test rejects stationarity (p < 0.01). Because both tests point the same way, the series is treated as having a unit root and is first-differenced before modelling; on the differenced series KPSS no longer rejects.

In MerQur

The 'KPSS specification: level stationarity / trend stationarity' and 'lags: auto / short / long' options in the Time Series analysis; it also appears in ARIMA Forecast when the order of differencing is chosen.

Analyses where this concept appears

Time SeriesARIMA Forecast

Related concepts

StationarityUnit root tests (ADF, Phillips-Perron)