Statistics Glossary → Autocorrelation (ACF) and partial autocorrelation (PACF) functions
Autocorrelation (ACF) and partial autocorrelation (PACF) functionsconcept
Functions showing a time series' correlation with its own past values at different lags (ACF), and the direct correlation remaining once the effect of the intermediate lags has been removed (PACF).
The lag-k autocorrelation is the correlation between the value of the series at time t and its value at time t - k; the plot over all lags is called a correlogram. The partial autocorrelation measures the lag-k relationship after removing the effect of the intermediate lags 1 to k - 1; if today's value depends only on yesterday's, for example, an indirect relationship with two days ago appears in the ACF but disappears in the PACF. The bands of roughly ±1.96/√n on the plots show the limits expected for white noise (an uncorrelated series).
The ACF and PACF guide order selection in ARIMA models: in a pure AR(p) process the PACF cuts off after lag p while the ACF dies away gradually; in a pure MA(q) process it is the other way round, with the ACF cutting off after lag q. A very slowly decaying ACF suggests that the series is non-stationary and needs differencing; pronounced spikes at seasonal lags (e.g. 12, 24) point to seasonality. The Ljung-Box test checks whether the autocorrelations at several lags are jointly zero and is used especially to check whether model residuals are white noise. In spatial data the corresponding idea is spatial autocorrelation, measured for example by Moran's I.
An everyday example
If today is hot in a city, tomorrow is probably hot too, so the autocorrelation at a lag of 1 day is high. A relationship with two days ago also appears, but most of it is an indirect link passed on through yesterday. The partial autocorrelation removes that indirect share and shows whether the day before yesterday makes any direct contribution to tomorrow.
In MerQur
In the Time Series analysis, the 'Number of ACF/PACF lags' parameter sets the largest lag shown in the correlogram and the Ljung-Box table. In ARIMA Forecast the 'Residual diagnostics (ACF + normality)' extra output gives the ACF of the residuals, and Mann-Kendall + Sen's Slope also has autocorrelation correction options.
Analyses where this concept appears
Time SeriesARIMA ForecastMann-Kendall + Sen's Slope
Related concepts
Ljung-Box (portmanteau) testARIMA orders (p, d, q) and driftSerial correlation tests (Breusch-Godfrey)