Journal of Quality Technology

A Quarterly Journal of Methods, Applications and Related Topics

ISSN: 0022-4065 (Print) 2575-6230 (Online) Journal homepage: [Link]

The Exponentially Weighted Moving Average

J. Stuart Hunter

To cite this article: J. Stuart Hunter (1986) The Exponentially Weighted Moving Average, Journal
of Quality Technology, 18:4, 203-210, DOI: 10.1080/00224065.1986.11979014

To link to this article: [Link]

Published online: 21 Feb 2018.

Submit your article to this journal

Article views: 1

View related articles

Citing articles: 86 View citing articles

Full Terms & Conditions of access and use can be found at

[Link]
 The Exponentially Weighted Moving Average

1. STUART HUNTER
179 Longview Drive, Princeton, Nj 08540

The Shewhart and CUSUM control chart techniques have found wide application in the manufacturing
industries. However, workpiece quality has also been greatly enhanced by rapid and precise individual
item measurements and by improvements in automatic dynamic machine control. One consequence is
a growing similarity in the control problems faced by the workpiece quality control engineer and his
compatriot in the continuous process industries. The purpose of this paper is to exposit a control chart
technique that may be of value to both manufacturing and continuous process quality control engineers:
the exponentially weighted moving average (EWMA) control chart. The EWMA has its origins in the
early work of econometricians, and although its use in quality control has been recognized, it remains
a largely neglected tool. The EWMA chart is easy to plot, easy to interpret, and its control limits are
easy to obtain. Further, the EWMA leads naturally to an empirical dynamic control equation.

Introduction comparisons of charting techniques will assume the

observations to be normally distributed. Thus Yt = 77
THE exponentially weighted moving average
II (EWMA)hasbeenaroundalongtime. The EWMA
+ E, where 77 is the mean of the process and the errors
Et are normally and independently distributed with
in quality control charting was first exposited by Rob-
zero mean and constant variance.
erts (1959) in a landmark paper in which the author
compared the average run lengths of the EWMAchart The Shewhart Control Chart
(there called "the geometric moving average chart")
to the Shewhart chart and other simple moving av- In a Shew hart control chart upper and lower control
erage schemes. The EWMA also found early applica- limits for the plotted points are established around
tion in economics (Muth [1960)), and in inventory the process mean 77 at positions 77 ± 30', the "three-
control and forecasting (Brown [1962)).The industrial sigma" limits where 0' is the standard error of the
uses of the EWMA were discussed by Freund (1962), points plotted. In a process under ideal control the
and modifications by Stewart (1970), and Wortham mean 77 will equal a fixed target value T required by
and Henrich (1972). The EWMA has, however, seldom the process specifications. A signal that the observed
been employed by quality control engineers. This is process requires attention is given whenever the most
doubly unfortunate since the EWMA can be viewed recent plotted point falls outside the control limits.
as a compromise between the Shewhart and CUSUM This method of process control charting is called a
charting procedures, and perhaps more importantly, Shewhart chart to honor its originator, Dr. W. A.
as a method for establishing real-time dynamic control Shewhart (1931, 1938).
of industrial processes. We begin with a brief descrip- In practice the Shewhart control chart is established
tion of the associated Shewhart and CUSUM charting empirically. Usually, repeated random homogeneous
procedures. samples of four or five observations each are gathered
In all that follows, observations Yt (t = 1, 2, ... , from the process, and, based upon the sample averages
T) are assumed to be sequentially recorded and the fj and ranges R (or estimates s of the standard devia-
observations, or some function of the observations, tion), empirical control charts are established with
plotted for the purpose of controlling a manufacturing the center line of the chart the grand average y. Many
process. In each case the plotted point is compared to approaches for establishing the three-sigma upper and
control limits, and if found to be beyond these critical lower control limits for the Shewhart control chart
boundaries, declared a signal that the observed process are in use, each dependent upon knowledge about the
requires attention. In this paper all descriptions and mean, variance and target values and each with its
own set of constants. See ANSI/ASQC Al (1978, Table
Dr. Hunter is a Professor Emeritus at Princeton University. 1), Grant (1952), Hansen (1963), Duncan (1974), and
He is a Fellow of ASQC. Grant and Leavenworth (1980).

Vol. 18, No.4, October 1986 203 Journal of Quality Technology

204 J. STUART HUNTER

The standard Shewhart chart with its ±3O' limits point of the Sr plot falls outside the branches of
only rarely provides a signal when the current process the V.
is at its mean level, thus keeping the Type I error, or Unlike the Shewhart chart, the control limit pa-
"a risk," small. Freund's (1957) acceptance control
rameters (the lead distance and interior angle of the
charts provide control limits that simultaneously con- V mask) take into account both the a and {:J risks along
sider both the Type I and Type II errors (i.e., both a with the size of the shift 6 felt to be economically
and (:J). important. Detailed instructions on the construction
of the V mask are given in Johnson and Leone (1977).
The CUSUM Chart
Many variations of the CUSUM are possible. If in
An alternative method for plotting sequentially re- equation (2) one uses the quantity C = T ± 6/2, the V
corded observations y, is to plot their cumulative sum, mask converts into two separate pairs of parallel con-
r
L y" against time t. Or, rather than the cumulative trollimits similar in appearance to the Shewhart chart,
t=1
one pair of control limits for C = T + 6/2, and a second
sum of the observations, one can use the cumulative
for C = T - 6/2. A single pair of these control limits
sum of the deviations
can be used as a one-sided CUSUM chart. In these
d,=y,-C (1) instances only those values of s; with the same sign
as 6 are plotted. For example, should 6 be positive and
where C is some convenient constant, usually the tar-
the CUSUM Sr go negative because of an observed
get value T. Thus a CUSUM chart is the quantity
series of negative recordings of d, = y, - (T + 6/2),
r r
then Sr is simply recorded as equal to zero. A one-
Sr= L:,(y,-C)= L:,dt (2)
/=1 '=1
sided CUSUM chart thus occasionally "ignores" ob-
served values of d, having a sign opposite to 6. The
plotted against T. plot of Sr restarts with the first d, with the same sign
The CUSUM control chart was first introduced in as 6. Recently Lucas and Crosier (1982) have proposed
England by Page (1954). Other important early con- important modifications to the CUSUM chart that
tributors are Barnard (1959), Ewan and Kemp (1960) provide for early detection of process shifts.
and Johnson and Leone (1962). Ewan (1963) provides Weighting the Historical Data
an excellent expository article on the CUSUM, and a
small text elucidating CUSUM procedures was au- One key to understanding the differences between
thored by Woodward and Goldsmith (1964). the Shewhart, CUSUM, and EWMAcontrol charts rests
in knowing how each charting technique uses the data
If the mean 71 of the observed values Yt equals the generated by the production process. For the classical
r
target value, T, then Sr = L a, will plot as a random Shewhart chart (where by "classical" we mean for
/=1
the moment to ignore the use of runs) the signal that
walk, that is, Sr will wander randomly about zero.
the process may be out of control is the appearance
However, should the mean of the process differ by a
of a single point falling beyond the 30' limits. In terms
slight amount 6 from the target value, then the ex-
of a weighting function, the signal thus depends entirely
pected value of Sr will add 6 with each observation.
on the last plotted point; that is, the weight given the
The plot of Sr thus increases or decreases depending
last plotted point is to, = 1, and the weights given to
on the sign of 6. Therefore, when plotting a CUSUM
all previous values are Wt-k = 0 for k ~ 1. For the
chart the analyst watches for a change in the slope
ordinary CUSUM chart the out of control signal de-
of the plot of Sr as an indication of a shift in mean r
away from target. pends upon the sum Sr = L d,. Of course, in
'=1
Surprisingly, even when the mean is on target, the computing a sum each entry has equal weight, and
CUSUM Sr can wander remarkably far and give the thus the most recent d, is equal in influence to the
appearance of a change in mean. For statistical process most ancient, and io, = l/n for all t.
control purposes the CUSUM chart is often accom- The weighting functions for the classical Shewhart
panied by a "V" mask. The mask is in the shape of a chart and for the ordinary CUSUM chart are displayed
V placed sidewise (>) with its vertex placed a fixed in Figures la and 1b. The reader will note that for the
distance from the last plotted CUSUM point. As long purpose of providing signals the classical Shewhart
as the entire historical track of Sr lies within the chart has no memory. That is, it ignores the immediate
branches of the V, no signal is provided. A signal that history. The ordinary CUSUM chart, however, has an
the process needs attention occurs when any single elephant-like memory, giving equal attention to the

Journal of Quality Technology Vol. 18, No.4, October 1986

 THE EXPONENTIALLY WEIGHTED MOVING AVERAGE 205

(a) (b)
100
100 SHEWHART CUSUM
o:::-!1
:::-!1
o ~
.
~
::r:
::r: o 50
o 50 w
w ~
~

12 t
TIME TIME

(c) (d)
100 MOVING AVERAGE
100 EWMA
:::-!1 k=5
o. o:::-!1
~
::r:
o 50 50
w
~

a L...-.............- --'-..L......l--L......L......IL.....L....L...I...................._ a L...-........................L..-L.........&.....&--&-....L..-L.....L ..10.....:1_

t
TIME TIME
FIGURE 1. Data Weighting for the Shewhart, CUSUM, Moving Average and EWMA Charts.

most ancient datum as well as the most recent. Of The EWMA

course, the Shewhart chart with runs does use the
The exponentially weighted moving average
historical data, and various forms of the CUSUM chart
(EWMA) is a statistic with the characteristic that it
will ignore portions of the history. In both instances
gives less and less weight to data as they get older and
the data weighting functions will therefore differ from
older. A plotted point on an EWMA chart can be given
those displayed. Our purpose here is to illustrate in
a long memory, thus providing a chart similar to the
its simplest form how the charts weight the data.
ordinary CUSUM chart, or it can be given a short
The Moving Average memory and provide a chart analogous to a Shewhart
The desire to employ historical data more resource- chart. An example of a weighting function for a
fully has occasionally led to the use of the moving EWMA is displayed in Figure ld.
average. A plot of a moving average of k = 5 obser- The EWMA is very easily plotted and may be
vations simply displays the average of the five most graphed simultaneously with the data appearing on
recent observations. Each new entering observation a Shewhart chart. The EWMA is best plotted one time
forces the oldest in the group out of the computation. position ahead of the most recent observation. Later
The weighting function for a moving average of k discussion will show the EWMA may be viewed as the
= 5 is displayed in Figure lc. One may question the forecast for the next observation, but that need not
reasonableness of giving equal importance to the most bother us here. Our immediate purpose is only to plot
recent k observations and zero importance to all pre- the statistic. The EWMA equals the present predicted
vious. The moving average smooths a time series. value plus A times the present observed error of pre-
Further, as illustrated in Nelson (1983), the moving diction. Thus,
average can produce cyclic and trend-like plots even
when the original data are themselves independent EWMA = Y'+1 = y, + Ae, (3)
random events with a fixed mean. This characteristic
lessens its usefulness as a control mechanism. = y, + A(y,- y,) (4)

Vol. 18, No.4, October 1986 Journal of Quality Technology

206 J. STUART HUNTER

where 56
Yl+1= predicted value at time t + 1 (the new EWMA) 54
Yt = observed value at time t
Yt = predicted value at time t (the old EWMA)
52 • • • •* • •
e, = Yt - Yl = observed error at time t 50 * * it
time time *time
and A is a constant (0 < A < 1) that determines the 48
depth of memory of the EWMA.
46 • •
Equation (4) can be written as
44
(5) (0 ) (b) (c) (d)
Plotting the EWMA FIGURE 2. The EWMA, Indicated by the *, is Computed
To plot the EWMA, let the symbol * stand for pre- at Time t but Plotted at Time Position t + 1. Observations
dicted values and. for observed values. We now pro- are indicated by •.
pose to plot simultaneously the observations Yt and
the EWMA Yl+1 on the same graph. Data for the ex-
ample are given in Table 1. Let A = 0.5 and suppose value Y3 = Y2 + Ae2 = 51 + 0.5(47 - 51) = 49. The
at t = 0 a process thought to be under control has a fourth predicted value, displayed in Figure 2d, is ob-
target value T = 50. To initiate the EWMA, set the tained once the third observed value Y3 = 53 is in hand.
initial predicted value YI equal to the target value and Thus Y4 = Y3 + 0.5(Y3 - Y3) = 51, and so on. The pre-
plot it at position t = 1. The first observed value YI dicted values, the Yt+I, are the EWMA.
= 52 is now also plotted at position t = 1 as shown in The current EWMA contains all the historical in-
Figure 2a. The observed error el equals 2.0. The next formation and there is no need to keep records of the
predicted value Y2 is obtained using equation (3). Thus previous observations or their weights. The new
Y2 = 50 + 0.5(52 - 50) = 51 as illustrated in Figure EWMA, Yt+I, is simply updated by adding Aet to the
2b. When the second observed value Y2 = 47 is ob- old EWMA, Yl' After a little practice, plotting the
tained, Figure 2c can be constructed. The observed EWMA is almost as easy as plotting the successive
value Y2 is first plotted and then the new predicted observations.

TABLE 1. EWMA Forecast Computations Using Various Values of A

~ ~O.5 ~~O.8 ~=O.2

obs EWMA e, EWMA e, EWMA e,

1 52.0 (50.00) 2.00 (50.00) 2.00 (50.00) 2.00

2 47.0 51.00 -4.00 51.60 -4.60 50.40 -3.40
3 53.0 49.00 4.00 47.92 5.08 49.72 3.28
4 49.3 51.00 -1.70 51.98 -2.68 50.38 -1.08
5 50.1 50.15 -0.05 49.84 0.26 50.16 -0.06
6 47.0 50.13 -3.13 50.05 -3.05 50.15 -3.15
7 51.0 48.56 2.44 47.61 3.39 49.52 1.48
8 50.1 49.78 0.32 50.32 -0.22 49.82 0.28
9 51.2 49.94 1.26 50.14 1.06 49.87 1.33
10 50.5 50.57 -0.07 50.99 -0.49 50.14 0.36
11 49.6 50.54 -0.94 50.60 -1.00 50.21 -0.61
12 47.6 50.07 -2.47 49.80 -2.20 50.09 -2.49
13 49.9 48.83 1.07 48.04 1.86 49.59 0.31
14 51.3 49.37 1.93 49.53 1.77 49.65 1.65
15 47.8 50.33 -2.53 50.95 -3.15 49.98 -2.18
16 51.2 49.07 2.13 48.43 2.77 49.55 1.65
17 52.6 50.13 2.47 50.65 1.95 49.88 2.72
18 52.4 51.37 1.03 52.21 0.19 50.42 1.98
19 53.6 51.88 1.72 52.36 1.24 50.82 2.78
20 52.1 52.74 -0.64 53.35 -1.25 51.37 0.73
21 52.42 52.35 51.52

Journal of Quality Technology Vol. 18, No.4, October 1986

 THE EXPONENTIALLY WEIGHTED MOVING AVERAGE 207

The Weighting Constant A Based upon this very limited evidence the value A
= 0.2 is preferred since it provides the smallest error
As shown in the Appendix, the EWMA can be writ-
sum of squares. The 'best' estimate, the least squares
ten as
estimate, is A = 0.112. In practice, at least fifty obser-
YHI = LWiYi (6) vations should be employed in such estimation pro-
1=0
cedures.
where the Wi are weights and EWMA Control Limits
(7) As shown in the Appendix, the variance of the
EWMA is
The sum of the weights L ui, = 1. The constant A Var(EWMA) = [.\j(2 - A))~
i=O
determines the "memory" of the EWMAstatistic. That and thus
is, A determines the rate of decay of the weights and O'EWMA = [A/(2 - A)f5O'.
hence the amount of information secured from the An estimate of ~ can be obtained from the minimum
historical data. Note now that as A-I, WI - 1 and error sum of squares L e7 obtained while estimating
Yt+1 practically equals the most recent observation Yt. A. That is,
When the process is under control and A = 1, points T

plotted on the classical Shewhart chart and those on 02 = Le7/(T-l).

an EWMA chart are therefore almost equal in their
ability to detect signals of departures from assump- When the Shewhart and EWMAcharts are constructed
tions. As A- 0, the most recent observation has small and plotted simultaneously O'EWMA can be estimated
weight, and previous observations near equal (though from the information used in establishing the three-
lower) weights. Thus, as A - 0, the EWMA takes on sigma Shewhart control limits. That is,
the appearance of the CUSUM. The EWMA control 5uShewhart.
UEWMA = [A/(2 - A)f
chart for values of 0 < A < 1 stands between the
Shewhart and CUSUM control charts in its use of the The three-sigma control limits for the EWMA are
historical data.
'T ± 3[UEWMA).
The choice of A can be left to the judgment of the
For this example, with the target value 'T = 50 and A
quality control analyst. Experience with econometric
= 0.5 the three-sigma limits for the EWMA are
data suggests values of A = 0.2 ± 0.1. The smaller the
value of A the greater the influence of the historical 50 ± 1.73uShewhart.
data. For example, using equation (7) with A = 0.3, The Shewhart and the EWMA control limits may
the most recent Yt has weight 0.3 while Yt-IO has weight be conveniently placed on the same chart, and both
A(1 - A)IO = 0.0085; that is, only W l - IO/Wl = 0.028 or Yl and the EWMA co-plotted as indicated in Figure 3.
approximately 3% the weight of the most recent ob- Using UShewhart = 1.5 the three-sigma Shewhart control
servation. However, with A = 0.1 observation Yt-IO has limits for the plotted Shewhart points (usually aver-
weight W t - IO = 0.0349 or approximately 35% the weight ages) are given by 50 ± 4.5, and the three-sigma control
of the most recent observation. limits for the plotted EWMA forecasts (constructed
The parameter Acan be estimated using an iterative from the Shewhart points) by 50 ± 2.6. In either case
least squares procedure. To illustrate, suppose the ob- a "signal" is produced whenever the last plotted point
servations YI, Y2, ••• , Y20 listed in Table 1 had been falls beyond its appropriate control limits. In this ex-
gathered from a process with target value 'T = 50 before ample, at time position t = 19 the EWMA(the forecast)
A had been chosen. The analyst, by considering the falls beyond the upper control limit for the EWMA, a
data as new data arriving sequentially, could for dif- signal that the process needs immediate attention.
ferent values of A compute the corresponding se- in practice the classical Shewhart chart (only the
quential set of predicted values Ybased on the EWMA last plotted point falling outside the 30' limits providing
as illustrated in Table 1. For A = 0.5 the sum of squares a signal) is aided by the use of runs. For example, two
of the errors associated with the predicted values in out of three successive points falling within the region
20
Table 1 is L e7 = 89.66, for A = 0.8 the error sum of 1/ ± 20' and 1/ ± 30', or six points in sequence above the
t~1 chart's center line, are frequently taken to be signals
20 20
that the process is no longer under control, see Nelson
squares is L e7 = 117.39 and for A = 0.2, L e7 = 78.02.
t~1 t~1 (1984). Employing runs provides an informal use of

Vol. 18, No.4, October 1986 Journal of Qualify Technology

208 J. STUART HUNTER

56
UCL for Yt
54
52 •
. - - - - - - - - - - - - - - - - - - - - - - · .. -;~lt--UCL for EWMA
•
• •• *
• • * *
48 *
---------- -
* *
- - - _....L __ - . _ - - - - - - - - - -LCL for EWMA

46
• •
I - - - - - - - - - - - - - - - - - - - - - - - L C L for Yt
44
FIGURE 3. Observations y Ie, and the EWMA 1*' Co-plotted With Their Upper and lower Control limits.

the recent history and, in the hands of an experienced Of course, if an operator is part of a feedback control
analyst, can make the Shewhart chart take on the loop he must know what corrective action to perform,
aspects of the EWMA chart. However, the EWMA and care must be taken to avoid inflating the vari-
chart provides a regular and formal use of the histor- ability of the process by making changes too often.
ical data. Runs and all other data configurations are But control engineers long ago learned how to close
encompassed in the EWMA forecast. Further, once the feedback loop linking forecast and adjustment to
the EWMA has been computed it contains all the in- target. The same information feedback loop exists in
formation provided by the historical record. There is many situations which only the operator can control.
no need to save or, as in the case of the modified The EWMA chart not only provides the operator with
CUSUM, to give occasional zero weights to recorded a forecast, but also with control limits to inform when
observations. the forecast is statistically significantly distant from
the target. Thus, when an EWMA signal is obtained,
Process Control
appropriate corrective action based on the size of the
The EWMA can be used as a dynamic process con- forecast can often be devised.
trol tool. When the process mean T/ is on the target
(i.e., T/ = r ) all three charting procedures, the Shew- The Empirical Control Equation
hart, CUSUM and EWMA, are roughly equivalent in
their ability to monitor departures from target. How- The EWMA can be modified to enhance its ability
ever, the EWMA provides a forecast of where the pro- to forecast. In situations where the process mean
cess will be in the next instance of time. It thus pro- steadily trends away from target the EWMA can be
vides a mechanism for dynamic process control. improved by adding a second term to the EWMA pre-
diction equation. That is,
To control a process it is convenient to forecast
where the process will be in the next instance of time. Modified EWMA = Yt+1 = Yt + Alet + A2 Let
Then, if the forecast shows a future deviation from
where Al and A2 are constants that weight the error
target that is too large, some electro-mechanical con-
at time t and the sum of the errors accumulated to
trol system or process operator can take remedial ac-
time t. The coefficients Al and A2 can be estimated
tion to compel the forecast to equal the target. In
from historical data by an iterative least-squares pro-
modern manufacturing, and particularly where an
cedure similar to that illustrated earlier for the esti-
observation is recorded on every piece manufactured
mation of A in equation (3).
or assembled, a forecast based on the unfolding his-
torical record can be used to initiate a feedback control A third term can be added to the EWMA prediction
loop to adjust the process. equation to give the empiricalcontrol equation

Journal of Quality Technology Vol. 18, No.4, October 1986

 THE EXPONENTIALLY WEIGHTED MOVING AVERAGE 209

to the referees and the Editor for valuable commen-

tary.
where the symbol \7 means the first difference of the
errors et ; that is \7et = e, - et- I . Note now that the Appendix
forecast Yt+1 equals the present predicted value (zero
Substituting Y, = XYt-1 + (1 - X)Yt-1 into equation
if the process has been adjusted to the target) plus
(5) gives for the EWMA
three quantities: one proportional to e, the second a
function of the sum of the e, and the third a function Y'+l = XYt + (1- X)[XYt_1 + (1- X)Yt-d
of the first difference of the e.. These terms are some-
times called the "proportional", "integral", and "dif- = Xy, + X(1- X)Y'_I + (1- X)2Y'_1
ferential" terms in the process control engineer's basic but
proportional, integral, differential ("PID") control
equation. The parameters x., X2 and X3 weight the Yt-I = XY'-2 + (1- X)Y'-2'
historical data to give the best forecast. Thus
The EWMA may thus be viewed as more than an
alternative to either the Shewhart or the CUSUM con- Y'-l = XYt + X(1- X)Y'_I + X(1- XYYt-2 + (1- XfYt-2'
trol charts. The EWMA may also be viewed as a dy- Employing equation (5) recursively we find that the
namiccontrol mechanism to help keep a process mean EWMA is a linear combination of the observations of
on target whenever discrete data on manufactured the form
items are sequentially available. An extensive litera- t
ture exists describing discrete data process control, Yt+ 1 = EWMA= L WiYi
two important statistical references being Box and ;=0

Jenkins (1976) and more recently, Pandit and Wu where the weights
(1983). In this literature one finds that the EWMA is
an important member of a large class of time series ui, = X(1- Xy-i.
models identified by Box and Jenkins as ARIMA mod- Given the errors e, are independent with constant
els (Autoregressive, Integrated, Moving Average variance, the variance of the EWMA is
models).
V(EWMA)=(LWD~
Conclusion i=O

The technology of industrial quality control is a = [X 2+ X2(1- XY+ X2(1- X)4+ ... ]0- 2
continuum, beginning with simple time plots of data
and proceeding to control chart methodologies of in- = X2~/[I -(1- XY]= [A!(2- X)]~.

creasing complexity. The Shewhart, CUSUM and

EWMA control charts all represent stages in this con- References
tinuum. The EWMA chart may be viewed as an al- ANSI/ ASQC A: Table 1 (1978). American Society For Quality
ternative, or used in addition to, the more familiar Control, Milwaukee, WI 53203.
Shewhart and CUSUM charts. But the EWMA chart BARNARD, G. A. (1959). "Control Charts and Stochastic Pro-
also provides the quality control engineer with the cesses." Journal of the Royal Statistical Society, Series B 21, pp.
opportunity to begin to consider real-time dynamic 239-271.
control of processes using discrete data. Most appli- Box, G. E. P. and JENKINS, J. M. (1976). Time Series Analysis: Fore-
casting and Control. Holden-Day, Inc., San Francisco, CA.
cations of control charting techniques require that the
Cox, D. R. (1961). "Prediction by Exponentially Weighted Moving
operator leave the process alone. If desired, the EWMA
Averages and Related Methods." Journal of the Royal Statistical
can be employed to make the operator part of the Society, B 23, pp. 414-422.
feedback control loop. DUNCAN, A. J. (1974). Quality Control and Industrial Statistics. Rich-
ard D. Irwin, Inc., Homewood, IL.
Acknowledgments EWAN, W. D. and KEMP, K. W. (1960). "Sampling Inspection of
Continuous Processes with no Autocorrelation between Suc-
The author has discussed the EWMA with many cessive Results." Biometrika 47, pp. 363-380.
associates, most particularly G. E. P. Box. He is also EWAN, W. D. (1963). "When and How to Use CUSUM Charts."
indebted to Jack Wood and Donald Gardiner for their Technometrics 5, pp. 1-22.
reading of an earlier manuscript, to the students in FREUND, R. A. (1957). "Acceptance Control Charts." Industrial
many industrial short courses, and most particularly Quality Control 14, pp. 13-23.

Vol. 18, No.4, October 1986 Journal of Quality Technology

210 J. STUART HUNTER

FREUND, R. A. (1962). "Graphical Process Control." Industrial PANDIT, S. M. and Wu, S. (1983). Time Series and System Analysis
Quality Control 18, pp. 15-22. with Applications. John Wiley & Sons, Inc., New York, NY.
GRANT, E. L. (1952). Statistical Quality Control. McGraw Hill, ROBERTS, S. W. (1959). "Control Chart Tests Based on Geometric
Hightstown, NJ. Moving Averages." Tecnnometrics 1, pp. 239-250.
HANSEN, B. L. (1963). Quality Control: Theory and Applications. SHEWHART, W. A. (1931). Economic Controlof Quality. D. Van Nos-
Prentice Hall, Englewood Cliffs, NJ. trand Co., New York, NY.
JOHNSON, N. L. and LEONE, F. C. (1962). "Cumulative Sum Control SHEWHART, W. A. (edited by W. E. Deming) (1939). Statistical
Charts, Parts I, II & III." Industrial Quality Control 18, pp. 15- Methodfrom the Viewpoint of QualityControl. The Graduate School,
21; 19, pp. 29-36. Department of Agriculture, Washington, DC.
JOHNSON, N. L. and LEONE, F. C. (1977). Statistics and Experimental STEWART, K. B. (1970). "A New Weighted Average." Technometrics
Design in Engineering and thePhysical Sciences, 2nd ed. John Wiley 12, pp. 247-258.
& Sons, Inc., New York, NY. WOODWARD, R. H. and GOLDSMITH, P. L. (1964). Cumulative Sum
LUCAS, J. M. and CROSIER, R. B. (1982). "Fast Initial Response Techniques. Published for Imperial Chemical Industries, Oliver
for Cusum Quality-control Schemes: Give Your Cusum a Head & Boyd, London, England.
Start." Technometrics 24, pp. 199-206. WORTHAM, A. M. and HENRICK, G. F. (1972). "Control Charts
MUTH, J. F. (1960). "Optimal Properties of Exponentially Using Exponential Smoothing Techniques." ASQC Annual Tech-
Weighted Forecasts." Journal of the American Statistical Associ· nical Conference Proceedings, pp. 451-458.
ation 55, pp. 299-306.
NELSON, L. S. (1983). "The Deceptiveness of Moving Averages."
Journal of Quality Technology 15, p. 99.
PAGE, E. S. (1954). "Continuous Inspection Schemes." Biometrika Key Words: Cumulative Sum Control, Forecasting, Moving
41, pp. 100-115. Averages, Process Control, Shewhart Control Charts.

Journal of Qualify Technology Vol. 18, No.4, October 1986