Autocorrelation measures how much a series remembers its own past.
It is the time series version of correlation, comparing values a fixed number of periods apart.
It drives both forecasting and the dangers of treating time series rows as independent.
The ACF plot shows the autocorrelation at lag 1, lag 2 and onwards as a row of bars.
Bands on the plot mark the range expected from random noise, so bars outside them are worth attention.
A trend produces high autocorrelation that fades slowly across many lags.
Seasonality produces peaks at the seasonal lag and its multiples, such as 12, 24 and 36 for monthly data.
Random noise shows almost no autocorrelation at any lag.
The partial autocorrelation plot, or PACF, shows the link at each lag after removing the effect of shorter lags.
The ACF and PACF together guide the choice of settings in forecasting models.
Autocorrelation in the leftover errors of a model shows that it missed a pattern.
Ignoring autocorrelation makes ordinary standard errors too small and tests too eager to find effects.
This matters in experiments too, since daily metrics from the same users are rarely independent.
In practice, plotting the ACF is a quick way to find the main seasonal period of an unfamiliar series.
The usual mistake is reading high autocorrelation as a sign of a good forecast, when it often just reflects a trend.