Home › Stats-TimeSeries

Stationarity tests — Lodestar.Stats.TimeSeries

Two tests that answer the question a correlogram raises and cannot settle: may this series be modelled as it stands, or does it need differencing or detrending first? Both at statsmodels 0.15.0 parity.

test	its null	entry point
augmented Dickey-Fuller	the series has a unit root	Stationarity.AugmentedDickeyFuller
KPSS	the series is stationary around a level or a line	Stationarity.Kpss

The nulls are opposite, so a reader runs both. A small ADF p-value rejects a unit root; a small KPSS p-value rejects stationarity. Rejecting one and not the other points the same way from two sides; rejecting both or neither says the series is too short, or its trend is not what the test assumed.

type	carries
Stationarity	the two tests
DickeyFullerOptions	the trend terms, the lag rule and the maximum lag
DickeyFullerResult	the statistic, MacKinnon's p-value and critical values, the lag used
KpssOptions	the null's trend and the lag window rule
KpssResult	the statistic, its tabulated p-value, and whether that p-value was clamped
TrendTerms	constant, linear trend, quadratic trend, or nothing
LagSelection	Akaike, Schwarz, the t statistic, or a fixed lag
KpssLagRule	Hobijn's automatic window, Schwert's legacy one, or a fixed one
PValueBound	which way the truth lies when a p-value is the end of its table

The Python equivalence table maps each statsmodels call to its counterpart here.