Journal of Quantitative Economics, Vol. 10 No.

2, July 2012

DAMPED TREND EXPONENTIAL SMOOTHING:

PREDICTION AND CONTROL
GIACOMO SBRANA1
Abstract
Damped trend exponential smoothing models have gained importance in
empirical studies due to their remarkable forecasting performance. This paper
derives their theoretical forecast error variance, based on the implied ARIMA
model, as algebraic function of the structural parameters. As a consequence, the
minimum mean squared error (MMSE) forecasts as well as the h-step ahead
theoretical forecast error variances can also be expressed as algebraic (and
unique) functions of the structural parameters. Analytical results are provided for
the random coefficient state space model, as introduced by McKenzie and
Gardner (2010) Damped trend exponential smoothing: A modeling viewpoint,
International Journal of Forecasting, 26 (4), 661-665, in the single source of error
context. Moreover, algebraic results are also given for standard Holt-Winters
(damped) trend models in the multiple sources of errors context.
Keywords: Exponential smoothing, State-space models, Time Series, ARIMA
models, Forecasting
JEL Classification: C22, C53

1. Introduction
The use of trend exponential smoothing models has been widely employed in forecasting
time series. Those models, also known as Holt-Winters methods of exponential smoothing, have
been further developed in the last twenty years. In particular, during the eighties, the pioneering
contributions of Everette S. Gardner & Eddie Mckenzie suggested to damp the trend as the
forecast horizon increases (see for example Gardner &Mckenzie 1985, 1988,1989). More
specifically, given that forecasting a series into the future using a straight line is not necessarily
appropriate, the authors suggested to add an autoregressive (damping) parameter.
Subsequently, damped trend exponential smoothing models gained importance in empirical
studies due to their remarkable forecasting properties. This is confirmed by Armstrong(2006),
who recommended these models as they improve forecasting accuracy. In addition, Fildes,
Nikolopous, Crone & Syntetos (2008) identify the damped trend as a benchmark model to beat.
The importance of using damped trend is also recognized by Hyndman, Koehler, Ord & Snyder
(2008). The choice of adding a damping parameter is particularly important from a modeling point
of view. This is because the damping term makes the model extremely flexible in fitting the

1
NEOMA Business School, 1 rue du Marchal Juin, BP 215, 76130 Mont Saint Aignan, France. E-mail
address: gsb@rouenbs.fr
 DAMPED TREND EXPONENTIAL SMOOTHING: PREDICTION AND CONTROL 153

dynamics of the series. More specifically, as described in Gardner and McKenzie(2011), the
damped parameter can be interpreted as a measure of the persistence of different linear trends.
Recently, McKenzie & Gardner (MG)(2010) suggest a random coefficient state space model. This
model, as described below, has the same ARIMA(1,1,2) reduced form of the standard damped
trend but it introduces further flexibility.
Yet, despite the increasing development of the parametrization of those models, we still
know little on the theoretical forecast error variance of (damped) trend exponential smoothing
models. In other words, the lack of knowledge of the linkage between the ARIMA reduced form
parameters and those belonging the data generation process (structural process) represents a
clear gap. This paper fills this gap. In what follow an algebraic procedure that allows deriving the
theoretical variance of different damped trend exponential smoothing models is shown. These
algebraic results are relevant since they provide full control of the forecasting properties of the
underlying models. The paper is organized as follows: Section 2 shows the algebraic results
when the single source of error (SSOE) framework is adopted. More specifically, the linkages
between the reduced form ARIMA parameters and those of the random coefficient state space
models are fully described and discussed. Section 3 provides algebraic linkages when the
multiple sources of errors (MSOE) framework is employed. In particular, the algebraic relations
among the theoretical forecast error variances and the structural parameters of both trend and
damped trend exponential smoothing are shown.

2. SSOE
This section reconsiders and extends the results as in MG. More specifically, the
theoretical forecast error variance, based on the implied ARIMA model, is derived for the random
coefficient state space model. Accordingly, the same notation as in MG is fully adopted. Here,
results are shown in the context of SSOE. Yet, in the next section this assumption will be relaxed
by allowing for the more general MSOE context.
Consider the following error-correction form of a linear trend with additive errors:
yt =lt-1 +bt-1 + t
lt =lt-1 +bt-1 + 1 t ... (1)
bt =bt-1 + 1 t
Note that the coefficients of the innovations in the level and gradient equations are written
slightly different compared with the standard notation adopted by the exponential smoothing
literature. Yet, considering that this paper is an extension of MG, the same notation is fully
adopted here. The damped version of the previous model can be expressed suchthat:
yt =lt-1 + bt-1 + t
lt =lt-1 + bt-1 + 1 t ... (2)
bt = bt-1 + 1 t

MG show both the parameters of the implied ARIMA(0,2,2) belonging to (1) and those of
the ARIMA(1,1,2) belonging to (2) are trivial to derive. In addition, in both cases the error term of
the ARIMA model corresponds to as in the error-correctionform. The random coefficient state-
space model can be expressed as follows:
154 JOURNAL OF QUANTITATIVE ECONOMICS

yt =lt-1 +At bt-1 + t

lt =lt-1 +At bt-1 + 1 * t ... (3)

*
bt =At bt-1 + 1 t

where At is a sequence of independent, identically distributed binary random variates with

P At =1 = and P At =0 =1- . Note that (3) is a variant of the standard damped trend model. Yet,
this model is particularly flexible since it is a stochastic mixture of standard linear trend and
simple exponential smoothing (i.e. random walk plus noise) models. This feature is not only
appealing from a modeling viewpoint, but it is also in line with Brown (1963) who argued that the
parameters of the model may change from one segment to another one as processes are thought
to be locally constant.
The reduced form of (3) is a random coefficient ARIMA as follows:
*
(1-At B)zt = t -(* +At ) *
t-1 +At t-2 ... (4)
where B is the backshift operator and zt =(1-B)yt . In addition, after some algebra, the following
autocovariances can be obtained:
* 2
* 2 * 2 (1- )
E zt zt =2 1+*2 + (1- ) -2* 1- +
1-

* 2
* * 2 (1- )
E zt zt-1 =2 -* + 1- -* 2
1- +
1- ... (5)
* 2
* 3 (1- )
E zt zt-2 =2 2
1- 1- * +
1-

E zt zt-3 = E zt zt-2
Therefore, as shown by MG, can be generated by the following stochastic difference
equation (i.e. ARIMA(1,1,2)):
1- B zt =at +1 at-1 +2 at-2 ... (6)
Note that the variance of at is crucial since it is the theoretical forecast error variance of
the random coefficient state space model. MG claim that the parameters of (6) are complicated
functions of the four parameters of (3). Yet, using the algebraic procedure provided in the
appendix, it can be shown that the moving average parameters of (6) are exact functions of *;
*; ; 2 .To see this, note that the autocovariance function of 1 are:
2
0 = 1+ E zt zt -2 E zt zt-1
2
1 = 1+ E zt zt-1 - E zt zt-2 - E zt-1 zt-1
... (7)
2 =E zt zt-2 - E zt-1 zt-2 =* 2
k =0 for k3
At this stage,2 given the results in the Appendix, the autocovariances are sufficient to
derive the ARIMA parameters. Thus:

2 2
For the sake of clarity, note that 2 = 1+ E zt zt-2 - E zt zt-3 - E zt-1 zt-2 , but given that
2 *
E zt zt-3 = E zt zt-2 , it collapses to .
 DAMPED TREND EXPONENTIAL SMOOTHING: PREDICTION AND CONTROL 155

* 2
2 =
2a
1+ 2 E zt z - E zt zt-2 - E(zt-1 zt-1 )
... (8)
t-1
1 =
* 2 +2a

with:
1
2a = (0 -20 +G+2 20 +0 G-2 21 +2 22 +G ... (9)
4

And

G= (0 -21 +22 )(0 +21 +22 ) ... (10)

where E(zt zt ), E(zt zt-1 ) and E(zt zt-2 ) are those provided in (5). These expressions give full control
of the theoretical variance of the random coefficient state space model.3 As a consequence, the
minimum mean squared error (MMSE) forecasts as well as the h-step ahead theoretical forecast
error variance are also algebraic functions of the error-correction parameters.
Interestingly, the algebraic expression of can be used to derive the set of parameters
that allows to obtain the following equality 2a = . This set is shown in Figure 1. More specifically,
each point shown in Figure 1 is a combination of *, *and that yields 2a =2 . This finding
implies that it is not necessarily true that, as claimed by MG, the moving average parameters in
(6) must be different from those of the error-correction specification as in (3).
In empirical applications, one might be interested in comparing the implied parameters as
in (8) and (9) with the ARMA estimated parameters. A similar exercise was carried out for
example in Morley, Nelson & Zivot (2003) to compare the Beveridge-Nelson decomposition with
unobserved components models.

3. MSOE
The previous section has focused on exponential smoothing processes assuming the
single source of error framework. This section shows that, even in the context of multiple sources
of errors, it is possible to derive the theoretical variance of standard (damped) trend exponential
smoothing processes. First, algebraic results are provided for the damped trend exponential
smoothing. Subsequently, results are given for the standard Holt-Winters methods.
Consider the following MSOE version of the damped trend exponential smoothing:
yt =lt-1 + bt-1 + t
lt =lt-1 + bt-1 +(1+)t ... (11)
bt = bt-1 +(1+)t
This specification differs from (2) since the model is characterized by three different
shocks. The reduced form of (11) can be written as:
1- B 1-B yt = t - 1+ t-1 + t-2 + 1+ t-1 - t-2 + (1+)t-1 ... (12)

3
An Eviews 7 program containing a simulation running the whole procedure can be provided by the
author upon request. In addition, an Excel file computing the autocovariances as in (5) as well as the
moving average parameters, given the error-correction parameters, is also available upon request.
156 JOURNAL OF QUANTITATIVE ECONOMICS

*
Figure 1. Region of parameters ( ; * ; ) allowing 2 =2a

having the following autocovariance functions:

2 2 2
0 = 1+ 1+ 2
+ 2
2 +((1+) + 1+ ) 2 +( (1+)) 2
2
1 =- 1+ 2 - 1+ 2 - (1+) 2
... (13)
2 = 2
k =0 for k3
Therefore:
2
2 =
2a
... (14)
- 1+ 2 - 1+ 2 - (1+)2 2
1 =
2 +2a

with:
1 2
2a = [22 +22 2 +2 2 +22 2 +2 2 + 1+ 2
1+ 2
2 +H+
4
1 4 1
[- -64 2 -42 2 2 -82 2 2 -42 2 2 2 -4 4
+ H
2 2
2
-4 2
1+4+62 +43 +4 +22 2 3
1+2+2 +M
 DAMPED TREND EXPONENTIAL SMOOTHING: PREDICTION AND CONTROL 157

1
1
+ M2 ]2 ] ... (15)
2
and
1
2 2 2 2
H=[(1+) 2 (-1+ ) +(1+) 2 ] [(1+) 2
2 2 2
+
1
2 2 2
42 (1+ ) +(1+) 2 (1+ ) ] 2 ... (16)
2 2
M=(1+) 2 2
+22 1+ 2
+(1+) 2 (1+ 2
) ... (17)
Note that when =1 the damped trend (11) collapses to the standard Holt-Winters trend
model whose reduced form is:
2 2
(1-B) yt = t (1-B) + 1+ t-1 -t-2 +(1+)t-1 ... (18)
With autocovariance functions:
2 2
0 =62 +2(1+) 2 +(1+) 2
2
1 =-42 -(1+) 2
... (19)
2 =2
k =0 for k3
Such that:
2
2 =
2a
... (20)
-4 -(1+)2 2
2
1 =
2 +2a

with:

1 2 2 1 1 2 2 4 4 2 2
2a =2 + (1+) + 1+ J + 2 (1+) + [ (1+) + 4(1+) (1+) 2 2 +
2 4 4 4
4 3 2
2(1+) 4 +(1+) 3 J+2(1+) 1+ 2 J
1
2
+42 2(1+) 2 + 1+ 3 1+ +J ]2 ... (21)
and
2 2
J= 162 +(1+) 2 +4(1+) 2 ... (22)

Therefore, once the parameters of the error-correction form are obtained, the forecasting
performance of the model can be computed instantaneously. To conclude, these results shed
light on the forecasting properties of damped trend exponential smoothing models. More
specifically, even in the MSOE context we have full control of the reduced form parameters and
thus on the MMSE forecasts relative to the data generation process.

4. Conclusions
As recognized in Gardner (2006), the damped trend exponential smoothing has gained
importance in empirical studies due to its remarkable forecasting performance. This paper derives
158 JOURNAL OF QUANTITATIVE ECONOMICS

its theoretical forecast error variance based on the implied ARIMA model. Algebraic results are
provided for the standard Holt-Winters trend models, the damped trend exponential smoothing
and for the random coefficient state space model. These algebraic results are relevant since they
provide full control of the ARIMA reduced form parameters as exact functions of the structural
parameters of the underlying models.

Acknowledgements
I wish to thank Jeyananthan Sivakumaran for his editorial assistance.

Bibliography
Armstrong, J. S. (2006) Findings from evidence-based forecasting: Methods for reducing
forecast error, International Journal of Forecasting,22,583-598.
Brown, R.G. (1963) Smoothing, forecasting and prediction of discrete time series, Englewood
Cliffs, NJ: Prentice-Hall, Inc
Fildes, R., Nikolopous, K., Crone, S., and Syntetos, A. (2008) Forecasting and operational
research: a review, Journal of Operational Research Society,59,1-23
Gardner, E. S., Jr. (2006) Exponential smoothing: The state of art-Part II, International Journal
of Forecasting,22,637-666.
Gardner, E. S., Jr. and McKenzie, E. (1985) Forecasting trends in time series, Management
Science, 31,1237-1246
Gardner, E. S., Jr. and McKenzie, E. (1988) Forecasting trends in time series, Journal of the
Operational Research Society, 39,863-867
Gardner, E. S., Jr. and McKenzie, E. (1989) Seasonal exponential smoothing with damped
trends, Management Science, 35,372-376
Gardner, E. S., Jr. and McKenzie, E. (2011) Why the damped trend works, Journal of
Operational Research Society,62, 1177-1180
Hyndman, R., Koehler, A., Ord, J. K., and Snyder, R. D. (2008)Forecasting with exponential
smoothing: the state space approach, Berlin: Springer-Verlag.
McKenzie, E. and Gardner, E. S., Jr. (2010) Damped trend exponential smoothing: A modeling
viewpoint, International Journal of Forecasting, 26 (4), 661-665
Morley, J. C., Nelson, C. and E. Zivot (2003) Why are the Beveridge-Nelson and unobserved-
components decomposition of GDP so different?Review of Economics & Statistics,
LXXXV, 2, 235-243
Sbrana, G. (2011), Forecasting Aggregated Moving Average Processes with an Application to the
Euro-Area real interest rate, Journal of Forecasting, 31(1), 85-98
Sbrana, G. (2012), Structural time series models and aggregation: some analytical results,
Journal of Time Series Analysis, 32(3), 315-316
 DAMPED TREND EXPONENTIAL SMOOTHING: PREDICTION AND CONTROL 159

APPENDIX
Let be an invertible moving average processes of order two such that:
zt =et +1 et-1 +2 et-2
where et is white noise process. Considering the following autocovariance functions: Where
0 =E(zt zt ), 1 =E(zt zt-1 ) and2 =E(zt zt-2 ). The moving average parameters can be recovered solving
the following system of three equations (autocorrelations):
2 2
= 2 2
1 (1+1 +2 )
1 1 +1 2
= 2 2
0 (1+1 +2 )
0
2e = 2 2
(1+1 +2 )

These equations represent respectively the second and first order autocorrelations of the
process and the variance of et .
First, the following analytical solutions can be easily obtained:
2
2 =
2e
1
1 =
(2e +2 )

Secondly, substituting these solutions in the last equation of the system, the following
quartic equation in x (with x=2e ) can be obtained:
x4 + 22 -0 x3 + 222 -22 0 +21 x2 + 232 -22 0 x+42
=0
x(x+2 )2

This equation has four different solutions. Yet, the only solution leading to the invertible
process4 is:
1
2e = (0 -22 +G+2 20 +0 G-2 21 +2 22 +G ) ... (23
4

with:

G= (0 -21 +22 )(0 +21 +22 ) ... (24

and:
42
2 =
(0 -22 +G+2 20 +0 G-2 21 +2 22 +G )
41
... (25
1 =
(0 +22 +G+2 20 +0 G-2 21 +2 22 +G )

Similar results can be found in Sbrana (2011, 2012).

4
A process with roots of the characteristic function that lie outside the unit circle (i.e. 1 +2 <1; 2 -2 <1; -1
< 2 <1