Skip to content
ML Academy

Introduced on day 22

Autocorrelation function (ACF)

The correlation of a series with itself at lag k, plotted for k = 1, 2, ...; bars outside the white-noise band hint at memory or seasonality.

In detail

The sample ACF at lag k divides the lagged cross-product of deviations from the mean by the total variation. For white noise about 95% of bars fall within plus or minus 1.96 over the square root of the series length, so one bar in twenty outside the band is expected by chance. ACF is read after the series is made roughly stationary; on a trending series every bar is large.