A measure of the linear relationship between two separate instances of the same random variable, as distinct from the population correlation coefficient, ρ, which refers to the linear relationship between two different random variables. As with ρ, the possible values lie between−1 and 1 inclusive, with unrelated instances having a theoretical autocorrelation of 0.
In the case of a time series, autocorrelation measures the extent of the linear relation between values at time points that are a fixed interval (the lag) apart. Similarly, spatial autocorrelation quantifies the linear relationship between values at points in space that are a fixed distance apart (in any direction in the case of an isotropic process; see isotropy). It is usually found that spatial autocorrelation is near 1 for points close together and decays to 0 as the distance increases—thus the daily rainfalls at the Lords and Oval cricket grounds in London will resemble each other closely, but will bear little or no resemblance to the rainfalls at the Kensington Oval in the West Indies. The phrase ‘serial correlation’ was introduced by Yule in 1926, while ‘autocorrelation’ was first used in 1933 as a description of a (related) function used by Wiener in 1926.
For a random variable X at time (or location) t, the population autocorrelation function (ACF) for lag l, ρl , is given by
the autocovariance function for lag l divided by the variance of Xt (which, for a stationary process, is equal to that of Xt+l). At low lags autocorrelation is usually positive. It usually declines towards 0 as the lag increases. See also partial autocorrelation function.
The sample autocorrelation for lag l, rl, is given (for l=1, 2,…, n−1) for the sequence of n values x1, x2,…, xn (ordered in space or time) by
where x¯ is the sample mean.
A plot of the variance of 1/{2(Xt−Xt+l)} against l is called a variogram (or semi-variogram). The related plot of autocorrelation versus lag is called a correlogram, and a plot of the autocovariance against lag may be called a covariogram. See also serial correlation; periodogram.