International Journal of Forecasting 20 (2004) 5 – 10

www.elsevier.com/locate/ijforecast

Forecasting seasonals and trends by exponentially weighted

moving averages
Charles C. Holt
Graduate School of Business, University of Texas at Austin, Austin, TX, USA

Abstract

The paper provides a systematic development of the forecasting expressions for exponential weighted moving averages.
Methods for series with no trend, or additive or multiplicative trend are examined. Similarly, the methods cover non-seasonal,
and seasonal series with additive or multiplicative error structures. The paper is a reprinted version of the 1957 report to the
Office of Naval Research (ONR 52) and is being published here to provide greater accessibility.
D 2004 Published by Elsevier B.V. on behalf of International Institute of Forecasters.

Keywords: Exponential smoothing; Forecasting; Local seasonals; Local trends

1. Introduction number of forecasts are needed for sales of indivi-

dual products.
An exponentially weighted moving average is a The simplest application of an exponentially mov-
means of smoothing random fluctuations that has ing average would be to the following stochastic
the following desirable properties: (1) declining process. Consider the problem of making an expected
weight is put on older data, (2) it is extremely easy value forecast of a random variable whose mean
to compute, and (3) minimum data is required. A changes between successive drawings. The following
new value of the average is obtained merely by rule might be proposed: take a weighted average of all
computing a weighted average of two variables, the past observations and use this as your forecast of the
value of the average from the last period and the present mean of the distribution, i.e.,
current value of the variable. This paper utilizes
these desirable properties both to smooth current S̄t ¼ B½St þ ASt1 þ A2 St2 þ A3 St3 þ A4 St4 þ . . . bb
random fluctuations and to revise continuously sea-
sonal and trend adjustments. These may then be where B is a constant between 0 and 1, A is (1  B),
extrapolated into the future for forecasts. The flex- the S’s are observations of the variable and the t
ibility of the method combined with its economy of subscript indicates the time ordering of the observa-
computation and data requirements make it especial- tions. S̄t is the estimate of the expected value of the
ly suitable for industrial situations in which a large distribution, ESt. If the distribution mean is subject to
large changes, A should be small so as to quickly
attenuate the effect of old observations. However, if A
E-mail address: [email protected] (C.C. Holt). is too small, S̄t is subject to so much random variation

0169-2070/$ - see front matter D 2004 Published by Elsevier B.V. on behalf of International Institute of Forecasters.
doi:10.1016/j.ijforecast.2003.09.015
6 C.C. Holt / International Journal of Forecasting 20 (2004) 5–10

that it would be a poor estimator of the mean. The where the constant, B, determines how fast the expo-
following relation is convenient in minimizing com- nential weights decline over the past years—one
putations: period drawn from each year. N is the number of
periods in a year (or whatever is the length of the
S̄t ¼ BSt þ ð1  BÞS̄t1 periodic pattern).
Clearly, Eqs. (1) and (2) are interdependent, and
hence must be solved simultaneously. This is a result
A weighted moving average with exponential weights
of the fact that a new observation leads both to a
is clearly a sensible mode of behavior in dealing with
revision of the sales rate. However, the determination
this simple forecasting problem. An exploration of the
of each of these depends on the other.Substitute Eq.
exact conditions under which this behavior is optimal
(2) in Eq. (1).
will not be considered, but rather the question of
whether this approach to forecasting seems to hold
promise in coping with trends and seasonals in S̄t ¼ A½BS̄t þ ð1  BÞPtN St  þ ð1  AÞS̄t1 : ð3Þ
forecasting.
Solve for the current sales rate:

2. Forecasting ratio seasonals    

Að1  BÞ 1A
S̄t ¼ PtN St þ S̄t1 : ð4Þ
1  AB 1  AB

Thus the current sales rate turns out to be a weighted

average of the sales rate of the last period and the
current sales seasonally adjusted by the index of a
year ago. Numerical values of S̄t from Eq. (4) can be
substituted in Eq. (2) to determine Pt. However,
substituting Eq. (4) in Eq. (2) gives us an explicit
Let St = sales in period t and let S̄t = smoothed and analytic expression for the new seasonal ratio:
seasonally adjusted sales rate in period t. This is an
estimate of ES̄t. Pt = periodic (seasonal) adjustment    
1B Bð1  AÞ S̄t1
ratio for the tth period. This is an estimate of ES̄t/ESt. Pt ¼ PtN þ : ð5Þ
1  AB 1  AB St
The sales rate is obtained by combining the current
seasonally adjusted sales with the sales rate from the
previous period The new seasonal ratio is a weighted average of the
old seasonal ratio and the ratio of the past sales rate
S̄t ¼ APt St þ ð1  AÞS̄t1 ð1Þ to current sales.
The values of A and B can be chosen indepen-
where the constant, A, determines how fast the expo- dently depending upon how fast the level of sales
nential weights decline over the past consecutive changes and how fast the seasonal patterns change.
periods. A high value of A will minimize lag in following
changes in the level of sales, but there will be less
0VAV1:
smoothing of random fluctuations. A high value of
The current seasonal adjustment ratio is obtained A gives weight only to very recent observations
by combining the current ratio of sales rate to sales thereby minimizing lag in following changes in the
with the seasonal adjustment rate from a year ago: expected value of sales. A low value of A gives
weight to many and therefore older observations;
S̄t the larger sample size will reduce the random
Pt ¼ B þ ð1  BÞPtN : ð2Þ
St variability thereby giving greater accuracy in esti-
 C.C. Holt / International Journal of Forecasting 20 (2004) 5–10 7

mating the mean—unless the mean has changed in 3. Forecasting a ratio trend
the meantime. Thus A is chosen for the best
combination of smoothness and lag in following In order to explore the application of the exponen-
changes in seasonal patterns. The trial of several tially weighted moving average to forecasting a trend,
values of A and B with past sales data should we will first consider the simplest case in which there
suggest suitable values. Furthermore, if a large is no seasonal fluctuation. The sales rate is obtained
number of products were involved, it is likely that by combining the current sales with the sales rate from
some relationships can be found between the sto- the previous period corrected for trend.
chastic parameters of the time series that would
indicate desirable values for A and B. S̄t ¼ ASt þ ð1  AÞRt S̄t1 ; ð7Þ
Forecasts may be made of the expected value of
sales T periods in the future by using the following where Rt is the trend adjustment ratio for the tth
extrapolation formula: period. This is an estimate of ES̄t/ESt  1. Note the
implicit assumption that the trend has a constant
S̄t percentage change.
EStþT ¼ T ¼ 1; 2; . . . N 1 : ð6Þ The current trend ratio is obtained by combining
PtþTN
the current trend ratio with the trend ratio from the
previous period.
This amounts to assuming that the present sales rate
will persist in the future and the sales forecast is made
S̄t
by applying the seasonal ratio that is applicable to the Rt ¼ C þ ð1  CÞRt1 ; ð8Þ
period. S̄t1
Examination of Eq. (5) shows that the seasonal
ratio is gradually modified by the factor, S̄t  1/St. where the constant, C, determines how fast the expo-
While this is intended as a current estimate of the nential weights applied to trend ratios decline over the
seasonal ratio for the tth period, actually there is a past consecutive periods. Substitute Eq. (8) in Eq. (7)
time difference of one period between the two varia- to obtain:
bles. Thus, if there is any steady trend in sales, it will
   
become incorporated in the seasonal ratio. This would S ð1  AÞð1  CÞ
not appear undesirable until it is remembered that S̄t ¼ St þ Rt1 S̄t1 :
1  ð1  AÞC 1  ð1  AÞC
there is a lag of N periods in the adjustment of the
ð9Þ
seasonal ratios.
This is clear in Eq. (5) where Pt is obtained by Substitute Eq. (9) in Eq. (8) to obtain:
modifying Pt  N. With a periodic pattern repeating
every N periods, there is little alternative to intro-    
AC St 1C
ducing a lag of N periods in adjusting the seasonal Rt ¼ þ Rt1 :
1  ð1  AÞC S̄t1 1  ð1  AÞC
ratios—for example, January comes but once a year
and to learn anything new about Januarys requires a ð10Þ
year’s wait in between. In contrast, some new
Forecasts may be made of the expected value of sales
information is available on the trend every period,
T periods in the future by using the following extrap-
and hence there is a decided disadvantage in lumping
olation formula:
trend and seasonal factors together and introducing
the lag of N periods in modifying the trend factor.
EStþT ¼ St RTt T ¼ 1; 2; . . . ; N : ð11Þ
This problem can be avoided by introducing a trend
variable into the analysis. This assumes that the present estimate of sales rate will
continue in the future modified by the percentage trend
1
The forecasts can be readily extended beyond N period by factor that is currently estimated. The two analyses can
reusing the N seasonal ratios, Pt + 1  N,. . .,Pt. now begin.
8 C.C. Holt / International Journal of Forecasting 20 (2004) 5–10

4. Forecasting a ratio trend and seasonals The exponential weighting coefficients, A, B, and
C, may be adjusted independently depending on
The sales rate is obtained by combining the current random variability on one hand and on the other,
sales adjusted for seasonal with the sales rate of the the speed with which the sales rate changes aside
previous period adjusted for trend: from trend, the speed with which the seasonal pattern
changes, and the speed with which the trend changes
S̄t ¼ ASt Pt þ ð1  AÞRt S̄t1 : ð12Þ respectively. The use of exponentially weighted
moving averages for forecasting is not limited to
Each seasonal ratio is revised every N periods as ratio trend and seasonal adjustments. Trends may
before in Eq. (2). The trend ratio is revised each change by constant increments rather than constant
period as before in Eq. (8). Now substituting Eqs. percentages and seasonal fluctuations may be addi-
(2) and (8) in Eq. (12) we obtain: tive rather than multiplicative.

 
Að1  BÞ
S̄t ¼ St PtN 5. Forecasting a linear trend and additive
1  AB  ð1  AÞC seasonals
 
ð1  AÞð1  CÞ
þ S̄t1 Rt1 : ð13Þ We consider the case of both trend and seasonal, but
1  AB  ð1  AÞC
as before either of these can be worked out separately.
The sales rate is obtained from the following relation:
Substituting Eq. (13) in Eq. (2) yields:
S̄t ¼ aðSt þ Pt Þ þ ð1  aÞðS̄t1 þ rt Þ: ð17Þ
 
½1  B½1  ð1  AÞC where Pt = periodic (seasonal) adjustment increment
Pt ¼ PtN
1  AB  ð1  AÞC for the tth period. This is an estimate of ES̄t  ESt.
  rt = trend adjustment increment for the tth period. This
Bð1  AÞð1  CÞ S̄t1 is an estimate of E S̄t  ES̄t1 :
þ Rt1 : ð14Þ
1  AB  ð1  AÞC St Note that the lower case letters are used to denote
the additive case while upper case letters denote the
Substituting Eq. (13) in Eq. (8) yields: multiplicative case above.
The periodic adjustment increment is estimated as
  follows:
ACð1  BÞ St PtN
Rt ¼
1  AB  ð1  AÞC S̄t1 Pt ¼ bðS̄t  St Þ þ ð1  bÞPtN ð18Þ
 
½1  C½1  AB The trend adjustment increment is estimated as
þ Rt1 : ð15Þ follows:
1  AB  ð1  AÞC

rt ¼ cðS̄t  S̄t1 Þ þ ð1  cÞrt1 : ð19Þ

Forecasts may be made of the expected value of
sales T periods in the future by using the following Substituting Eqs. (18) and (19) in Eq. (17), we obtain:
extrapolation formula:
 
að1  bÞ
S̄t ¼ St
1  ab  ð1  aÞc
S̄t RTt  
EStþT ¼ T ¼ 1; 2; . . . N : ð16Þ ð1  cÞð1  aÞ
PtþT N þ ½S̄t1 þ rt1 
1  ab  ð1  aÞc
This assures that the present sales rate may be  
adjusted for a percentage trend and a seasonal both ð1  bÞa
þ PtN : ð20Þ
of which continue into the future. 1  ab  ð1  aÞc
 C.C. Holt / International Journal of Forecasting 20 (2004) 5–10 9

Substituting Eq. (20) in Eq. (18), we obtain: The seasonal rate is estimated by Eq. (2), and the trend
increment is estimated by Eq. (19). Substituting these
  in Eq. (24) yields the formula for calculating S̄t:
abð1  bÞ
Pt ¼  b St
1  ab  ð1  aÞc  
  Að1  BÞ
ð1  cÞð1  aÞb S̄t ¼ St PtN
þ ½S̄t1 þ rt1  1  AB  ð1  AÞc
1  ab  ð1  aÞc  
  ð1  AÞð1  cÞ
abð1  bÞ þ ½S̄t1 þ rt1 : ð25Þ
þ þ ð1  bÞ PtN : ð21Þ 1  AB  ð1  AÞc
1  ab  ð1  aÞc
Values of Pt and rt may then be calculated numerically
Substituting Eq. (20) in Eq. (19), we obtain: by substitutions in Eqs. (2) and (19), respectively.
Forecasts for sales T periods in the future would be
obtained by the formula:
 
acð1  bÞ
rt ¼ St S̄t þ rt T
1  ab  ð1  aÞc EStþT ¼ T ¼ 1; 2; . . . ; N : ð26Þ
  PtþTN
ð1  cÞð1  aÞc
þ  c S̄t1
1  ab  ð1  aÞc
 
ð1  cÞð1  aÞc 7. Comments on the theory
þ þ ð1  cÞ rt1
1  ab  ð1  aÞc
  The foregoing derivations illustrate the flexibility
ð1  bÞac
þ PtN : ð22Þ of exponentially weighted moving averages in fore-
1  ab  ð1  aÞc casting. They make possible a simple integrated ap-
proach to the estimation of both trends and seasonals.
Forecasts of sales T periods in the future may be made The underlying stochastic theory has not been
with the following relation: explored here. A great deal of work has been done
on optimal filter design for stationary time series
EStþT ¼ S̄t þ rt T  PtþTN T ¼ 1; 2; . . . ; N : ð23Þ using the criteria of minimizing the sum of the square
of the errors.2 The solution of this problem leads to
The trend adjustment solution without the seasonal the optimality of linear filters of which exponential
may be obtained from Eqs. (20) – (22) by letting b and weights are but a special case. However, the implicit
Pt  N equal zero. The seasonal adjustment solution models assumed by the analysis of this paper are a
without the trend may be obtained by letting c and good deal more complicated than those assumed in
rt  1 equal zero. The combination of a multiplicative Wiener’s book.
seasonal and a linear trend is very frequently found to Many linear forecasting rules can be constructed
be suitable for forecasting. The following derivation using the general approach that is presented here. The
extends the exponentially weighted moving average to rules that have been developed are all based on first
this case. order difference equations. Second- and higher-order
equations would allow even greater flexibility. Many
other extensions are possible. For example, a macro-
6. Forecasting a linear trend and ratio seasonals economic variable like GNP could be forecasted

The sales rate is estimated by the following 2

relation: For example, see Norbert Wiener, The Extrapolation,
Interpolation, and Smoothing of Stationary Time Series, 1949.
Note particularly Levinson’s appendix on the solution for difference
S̄t ¼ ASt Pt þ ð1  AÞðS̄t1 þ rt Þ ð24Þ equations.
10 C.C. Holt / International Journal of Forecasting 20 (2004) 5–10

independently, and used to modify the forecast of an Biography: Charles C. HOLT is Professor of Management
individual product through a self-adjusting parameter. Emeritus at the Graduate School of Business, University of
Texas at Austin. His current research is on quantitative decision
methods, decision support systems, and financial forecasting.
Previously he has done research and teaching at M.I.T., Carnegie
8. Conclusion Mellon University, the London School of Economics, the Uni-
versity of Wisconsin, and the Urban Institute. He has been active
in computer applications since 1947, and has done research on
This exploratory analysis indicates the great flex-
automatic control, the simulation of economic systems, schedul-
ibility of exponentially weighted moving averages in ing production, employment and inventories, and the dynamics
dealing with forecasts of seasonals and trends. Further of inflation and unemployment.
study seems fully justified both on empirical and
theoretical levels.