ABSTRACT
A time series sometimes repeats patterns or has other properties whereby earlier values have some relation to later values. Autocorrelation (sometimes called serial correlation) is a statistic that measures the degree of this affiliation. As an example, a time series of daily measurements might repeat itself every two weeks (i.e. show a lag of 14 days). That means the value for day 15 is similar to that for day I, day 16 is similar to day 2, and so on. The autocorrelation statistic (explained below) is a number that, by its magnitude, shows that similarity. (In practice, using measurements of just one feature, we first compute autocorrelation using a lag of one. That is, we first compare day I to day 2, day 2 to day 3, and so on. Then we do repeat computations using lags of two, three, and so on. Methodically evaluating the computed statistic for successively larger lags reveals any dependence or correspondence among segments of the time series.)
Autocorrelation in a time series can cause problems in certain kinds of analyses. For example, many statistical measures are designed for unrelated or independent data (as opposed to correlated data). Applying measures built for independent data to correlated data can produce false or meaningless results. Secondly, autocorrelation might influence our choice of prediction methods, as least for short-term predictions.
