Notes_3, GEOS 585A, Spring 2007

1
3 Autocorrelation

Autocorrelation refers to the correlation of a time series with its own past and future values.
Autocorrelation is sometimes called serial correlation, which refers to the correlation between
members of a series of numbers arranged in time. Alternative terms are lagged correlation,
and persistence. Geophysical time series are frequently autocorrelated because of inertia or
carryover processes in the physical system. For example, the slow drainage of groundwater
reserves might impart correlation to successive annual flows of a river. Or stored photosynthates
might impart correlation to successive annual values of tree-ring indices. Autocorrelation
complicates the application of statistical tests by reducing the effective sample size.
Autocorrelation can also complicate the identification of significant covariance or correlation
between time series (e.g., precipitation with a tree-ring series). Three tools for assessing the
autocorrelation of a time series are (1) the time series plot, (2) the lagged scatterplot, and (3) the
autocorrelation function.

3.1 Time series plot
Positively autocorrelated series are sometimes referred to as persistent because positive
departures from the mean tend to be followed by positive depatures from the mean, and negative
departures from the mean tend to be followed by negative departures. In constrast, negative
autocorrelation is characterized by a tendency for positive departures to follow negative
departures, and vice versa. Positive autocorrelation might show up in a time series plot as
unusually long runs, or stretches, of several consecutive years above or below the mean.
Negative autocorrelation might show up as an unusually low incidence of such runs. Because the
departures for computing autocorrelation are computed relative the mean, a horizontal line
plotted at the sample mean is useful in evaluating autocorrelation with the time series plot.
Visual assessment of autocorrelation from the time series plot is subjective and depends
considerably on experience. Statistical tests based on the observed number of runs above and
below the mean are available (e.g., Draper and Smith 1981), though none are covered in this
course. It is a good idea, however, to look at the ime series plot as a first step in analysis of
persistence. If nothing else, this inspection might show that the persistence is much more
prevalent in some parts of the series than in others.


3.2 Lagged scatterplot
The simplest graphical summary of autocorrelation in a time series is the lagged scatterplot,
which is a scatterplot of the time series against itself offset in time by one to several years. Let
the time series of length N be , 1, ...,
i
x i N = . The lagged scatterplot for lag k is a scatterplot of
the last N k observations against the first N k observations. For example, for lag-1,
observations
2 3,
, ,
N
x x x are plotted against observations
1 2, 1
, ,
N
x x x

.
A random scattering of points in the lagged scatterplot indicates a lack of autocorrelation.
Such a series is also sometimes called random, meaning that the value at time t is independent
of the value at other times. Alignment from lower left to upper right in the lagged scatterplot
indicates positive autocorrelation. Alignment from upper left to lower right indicates negative
autocorrelation.
An attribute of the lagged scatterplot is that it can display autocorrelation regardless of the
form of the dependence on past values. An assumption of linear dependence is not necessary.
Notes_3, GEOS 585A, Spring 2007

2
An organized curvature in the pattern of dots might suggest nonlinear dependence between time-
separated values. Such nonlinear dependence might not be effectively summarized by other
methods (e.g., the autocorrelation function, which is described later). Another attribute is that the
lagged scatterplot can show if the autocorrelation is driven by one or more outliers in the data.
This again would not be evident from the acf.
Fitted line. In the class MATLAB script for the assignment, lagged scatterplots are drawn for
lags 1-8 years. The straight line that appears on these plots is fit by least squares, and is intended
to aid in judging the preferred orientation of the pattern of points.
Correlation coefficient and 95% significance level. The correlation coefficient for the
scatterplot summarizes the strength of the linear relationship between present and past values.
The correlation coefficient is annotated at the top of the plot, along with the correlation that
would be considered approximately significant at the 95% significance level against the null
hypothesis that true correlation is zero. This approximate 95% confidence level is computed as
2.0 divided by the square root of the sample size, or series length. As described in Chatfield
(1975, p. 25), if a series is completely random, then, for large sample size sample size N, the
lagged-correlation coefficient is approximately normally distributed with mean 0 and variance
1/N. The probability is thus roughly ninety-five percent that the correlation falls within two
standard deviations, or 2.0/sqrt(N) of zero.

3.3 Autocorrelation function (correlogram)
An important guide to the persistence in a time series is given by the series of quantities called
the sample autocorrelation coefficients, which measure the correlation between observations at
different times. The set of autocorrelation coefficients arranged as a function of separation in
time is the sample autocorrelation function, or the acf. The first step to understanding the
autocorrelation function is to understand the product-moment correlation coefficient. Assume N
pairs of observations on two variables x and y. The correlation coefficient between x and y is
given by

( )( )
( ) ( )
1/ 2 1/ 2
2 2
i i
i i
x x y y
r
x x y y

=
( (

(1)



where the summations are over the N observations.
A similar idea can be applied to time series for which successive observations are correlated.
Instead of two different time series, the correlation is computed between one time series and the
same series lagged by one or more time units. For the first-order autocorrelation, the lag is one
time unit. The first-order autocorrelation coefficient is the simple correlation coefficient of the
first 1 N observations, , 1, 2,..., 1
t
x t N = and the next 1 N observations, , 2,3,...,
t
x t N = .
The correlation between
t
x and
1 t
x
+
is given by

( )( )
( ) ( )
1
(1) 1 (2)
1
1 1/ 2 1/ 2
1 1
2 2
(1) (2)
1 1
N
t t
t
N N
t t
t t
x x x x
r
x x x x

+
=

= =

=
( (

( (

(2)
Where
(1)
x is the mean of the first 1 N observations and
(2)
x is the mean of the last
1 N observations. As the correlation coefficient given by (2) measures correlation between
Notes_3, GEOS 585A, Spring 2007

3
successive observations, it is called the autocorrelation coefficient or serial correlation
coefficient.
For N reasonably large, the difference between the sub-period means
(1)
x and
(2)
x can be
ignored and
1
r can be approximated by


( )( )
( )
1
1
1
1
2
1
N
t t
t
N
t
t
x x x x
r
x x

+
=
=

=

(3)
where
1
N
t
t
x x
=
=

is the overall mean.

+
=
=

=

(4)




The quantity
k
r is called the autocorrelation coefficient at lag k. The plot of the
autocorrelation function as a function of lag is also called the correlogram.

Link between acf and lagged scatterplot. The correlation coefficients for the lagged
scatterplots at lags 1, 2, ...8 k = years are equivalent to the acf values at lags 1,,8.

Link between acf and autocovariance function (acvf). Recall that the variance is the average
squared departure from the mean. By analogy the autocovariance of a time series is defined as
the average product of departures at times t and t+k
( ) ( )
1
1
N k
k t t k
t
c x x x x
N

+
=
=

(5)
where
k
c is the autocovariance coefficient at lag k. The autocovariance at lag zero,
0
c , is the
variance. By combining equations (4) and(5), the autocorrelation at lag k can be written in terms
of the autocovariance:

0 k k
r c c = (6)

Alternative equation for autocovariance function. Equation (5) is a biased (though
asymptotically unbiased) estimator of the population covariance. The acvf is sometimes
computed with the alternative equation

( )
( )( )
1
1
N k
k t t k
t
c x x x x
N k

+
=
=

(7)
Notes_3, GEOS 585A, Spring 2007

4
The acvf by (7)has a lower bias than the acvf by (5), but is conjectured to have a higher mean
square error (Jenkins and Watts 1968, chapter 5).
3.4 Testing for randomness with the correlogram
The first question that can be answered with the correlogram is whether the series is
random or not. For a random series, lagged values of the series are uncorrelated and we expect
0
k
r . It can be shown that if
1
....,
N
x x are independent and identically distributed random
variables with arbitrary mean, the expected value of
k
r is
( ) 1
k
E r N = (8)
the variance of
k
r is
Var( ) 1/
k
r N (9)
and
k
r is asymptotically normally distributed under the assumption of weak stationarity. The 95%
confidence limits for the correlogram can therefore be plotted at 1/ 2 N N , and are often
further approximated to 0 2 N . Thus, for example, if a series has length 100 years, the 95%
confidence band is at 2 100 0.20 = . Any given
k
r has a 5% chance of being outside the
95% confidence limits, so that one value outside the limits might be expected in a correlogram
plotted out to lag 20 even if the time series is drawn from a random (not autocorrelated)
population.
Whether a given
k
r outside the 95% confidence band really indicates the population the
sample comes from is autocorrelated is therefore open to question. Factors that must be
considered are (1) how many lags are being examined, (2) the magnitude of
k
r , and (3) at what
lag k the large coefficient occurs. A very large
k
r is less likely to occur by chance than an
k
r barely outside the confidence bands. And a large
k
r at a low lag (e.g., 1 k = ) is far more likely
to represent persistence in most physical systems than an isolated large
k
r at some higher lag.

3.5 Large-lag standard error
While the confidence bands described above are horizontal lines above and below zero on the
correlogram, the confidence bands you will see on the correlograms in the assignment script may
appear to be narrowest at lag 1 and to widen slightly at higher lags. That is because the
confidence bands produced by the script are the so-called large-lag standard errors of
k
r (Anderson 1976, p. 8). Successive values of
k
r can be highly correlated, so that an individual
k
r might be large simply because the value at the next lower lag,
1 k
r

, is large. This
interdependence makes it difficult to assess just at how many lags the correlogram is significant.
The large-lag standard error adjusts for the interdependence. The variance of
k
r , with the
adjustment, is given by

2
1
1
Var( ) 1 2
K
k i
i
r r
N
=
| |
+
|
\ .

(10)
where K k < . The square root of the variance quantity given by (10) is called the large-lag
standard error of
k
r (Anderson 1976, p. 8). Comparison of (10) with (9) shows that the
adjustment is due to the summation term, and that the variance of the autocorrelation coefficient
at any given lag depends on the sample size as well as on the estimated autocorrelation
coefficients at shorter lags. For example, the variance of the lag-3 autocorrelation coefficient,
3
r ,
Notes_3, GEOS 585A, Spring 2007

5
is greater than 1/ N by an amount that depends on the coefficients at lags 1 and 2. (The
summation is over lags 1 to K , where 2. K = )

3.6 Hypothesis test on r
1
The first-order autocorrelation coefficient is especially important because for physical systems
dependence on past values is likely to be strongest for the most recent past. The first-order
autocorrelation coefficient,
1
r , can be tested against the null hypothesis that the corresponding
population value
1
0 = . The critical value of
1
r for a given significance level (e.g., 95%)
depends on whether the test is one-tailed or two-tailed. For the one-tailed hypothesis, the
alternative hypothesis is usually that the true first-order autocorrelation is greater than zero:

1
H : 0 > (11)
For the two-tailed test, the alternative hypothesis is that the true first-order autocorrelation is
different from zero, with no specification of whether it is positive or negative:

1
H : 0 (12)
Which alternative hypothesis to use depends on the problem. If there is some reason to expect
positive autocorrelation (e.g., with tree rings, from carryover food storage in trees), the one-sided
test is best. Otherwise, the two-sided test is best.
For the one-sided test, the World Meteorological Organization recommends that the 95%
significance level for
1
r be computed by

1,.95
1 1.645 2
1
N
r
N
+
=

(13)
where N is the sample size. User-written Matlab function acf.m , which is called by assign3.m,
(13) for the critical (95%) level for
1
r ; this critical level is returned as an output argument by
acf.m.
More generally, following Salas et al. (1980), who refer to Andersen (1941), the probability
limits on the correlogram of an independent series are


k
1 1.645 1
(95%) one sided
1 1.96 1
r (95%) two sided
k
N k
r
N k
N k
N k
+
=

(14)


where N is the sample size and k is the lag. Equation (13) is just the above equation for the one-
sided test with lag 1. k =

3.7 Effective Sample Size
If a time series of length N is autocorrelated, the number of independent observations is fewer
than N . Essentially, the series is not random in time, and the information in each observation is
not totally separate from the information in other observations. The reduction in number of
independent observations has implications for hypothesis testing.
Some standard statistical tests that depend on the assumption of random samples can still be
applied to a time series despite the autocorrelation in the series. The way of circumventing the
Notes_3, GEOS 585A, Spring 2007

6
problem of autocorrelation is to adjust the sample size for autocorrelation. The number of
independent samples after adjustment is fewer than the number of observations of the series.
Below is an equation for computing so-called effective sample size, or sample size adjusted for
autocorrelation. More on the adjustment can be found elsewhere (WMO 1966; Dawdy and
Matalas 1964). The equation was derived based on the assumption that the autocorrelation in the
series represents first-order autocorrelation (dependence on lag-1 only). In other words, the
governing process is first-order autoregressive, or Markov. Computation of the effective sample
size requires only the sample size and first-order sample autocorrelation coefficient. The
effective sample size is given by:


( )
( )
1
1
1
'
1
r
N N
r

=
+
(15)
where N is the sample size, N is the effective samples size, and
1
r is the first-order
autocorrelation coefficient. For example, a series with a sample size of 100 years and a first-
order autocorrelation of 0.50 has an adjusted sample size of


( )
(1 0.5) 0.5
' 100 100 33 years
1 0.5 1.5
N

= =
+


References

Anderson, R.L., 1941, Distribution of the serial correlation coefficients: Annals of Math.
Statistics, v. 8, no. 1, p. 1-13.
Anderson, O., 1976, Time series analysis and forecasting: the Box-Jenkins approach: London,
Butterworths, p. 182 pp.
Chatfield, C., 1975, The analysis of time series: Theory and practice, Chapman and Hall,
London, 263 pp.
Dawdy, D.R., and Matalas, N.C., 1964, Statistical and probability analysis of hydrologic data,
part III: Analysis of variance, covariance and time series, in Ven Te Chow, ed.,
Handbook of applied hydrology, a compendium of water-resources technology: New
York, McGraw-Hill Book Company, p. 8.68-8.90.
Jenkins, G.M., and Watts, D.G., 1968, Spectral analysis and its applications: Holden-Day, 525 p.
Salas, J.D., Delleur, J.W., Yevjevich, V.M., and Lane, W.L., 1980, Applied modeling of
hydrologic time series: Littleton, Colorado, Water Resources Publications, 484 pp.
World Meterorological Organization, 1966, Technical Note No. 79: Climatic Change, WMO-No,
195.TP.100, Geneva, 80 pp.