Nepal Journal of Science and Technology 3 (2001) 1 4

Application of Triple Exponential Smoothing Technique

in the Analysis of Time Series with a Quadratic Trend
S. L. Shrestha
Central Department of Statistics, Tribhuvan Universi~Kirtipul; Kathmandu

Received March 2000; acceptedAugust 2001

Abstract
Exponential Smoothing Technique is a useful forecasting tool in time series analysis. In contrast to the Regression
Approach of forecasting, the technique is based upon unequally weighting the observed data with recent observations
getting higher weightage than the remote ones. One such technique is theTriple Exponential Smoothing which is
applied to the time series with a quadratic trend. The paper aims to highlight this method of forecasting with
reference to time series of the Annual Number of Tourist Anivals in Nepal during the period 1971 - 1997.The
predicted series of data arising from theTriple Exponential Smoothing statistically fits to the observed time series
considered.
Keywords: Autocorrelation, Diagnostic Checking, Forecasting, Regression, Smoothing Constant

Introduction of one among them depends upon the nature of the

Among various forecasting techniques, the multiple time series. If a time series under investigation does
regression, exponential smoothing and Box-Jenkins not exhibit any particular trend i.e. the average level
methodology are considered as the basic and useful of the time series is not changing over time or changing
forecasting techniques in the process of analysis very slowly then simple exponential smoothing is
involving time series data. The regression approach applied. If the time series follow a linear trend over
uses the least square estimation which minimizes the time i.e. the average level of time series changes in
sum of squared residuals and are based upon equally linear fashion over time then double exponential
weighting each of the observed values of the time smoothing is used. Lastly, if the time series exhibits a
series. Where as the exponential smoothing is a quadrate trend i.e. the average level of the time series
forecasting method that weights each of the values changes over time in a curvilinear fashion then triple
unequally, with recent observations being weighted exponential smoothing is adopted.
more heavily than the relatively remote ones. This Among the three exponential smoothing
procedure allows the forecaster to update the estimates techniques, the triple exponential smoothing technique
of the parameters as time changes. The Box-Jenkins is illustrated here with reference to the available time
methodology, generally regarded as one of the most series data of the annual number of tourist arrivals in
accurate forecasting methodology relies in the Nepal during the period 1971-1997.
functions known as Autoregressive Function (ACF),
Partial Autocorrelation Function (PACF) and the Methodology
resulting Correlograms. The common models under
this methodology are Autoregressive Models (AR), Consider a time series, which can be represented by a
Moving Average Models (MA), Autoregressive quadratic function of time as follows
MovingAverage Models (ARMA) andAutoregressive Y t = Po+ Prt + 112 P2t2+pt ...(1)
Integrated Moving Average Models (ARIMA). Suppose that at the end of the periodT, we have a
However, the use of such models essentially require a set of observationsy 1,Y2, ....YT, If bo(T), bl(T). and
large amount of data (generally at least 50 observations) b2(T) are the estimate of Po, PI, and &, respectively,
to build an accurate model (Chatfield 1996). then the estimated forecast for a future periodT+t is
While considering the Exponential Smoothing given by
technique, we come across its various types. Choice E~t.y~+~Q=b&I')+b i(T) (T+t)+lI2 b2(T) ( ~ + t..(2)
)~
 S.L. ShresthaINepal Journal of Science and Technology 3 (2001) 1-4

Changing the origin to the time periodT , we get forecasts is generated. These are then compared to the
the estimated forecast as actual set of observation in the time series.The value
Est. YT+t (T)=AO(T)+Al(T) jt)+ 112A2(T) (t12 (3) . of a which generates a minimum sum of squared
forecast residuals is finally chosen.
The triple exponential smoothing involves the use
three smoothing statistics. These are as follows.
ST = aYT + (I-a) ST-l ... (4) Results
s ~ ( 2=) asT+ ( 1 4 ) ST-1(2) ...(5) This section covers the results of the application of
s ~ =( ~ ) + (1-a) s ~ ~ ( .~
.. (6)
) the triple exponential technique with reference to the
time series data of the annual number of tourist arrivals
Where ST is single smoothed statistic, S$2) is
in Nepal during the period 1971-1997.Data of the time
double smoothed statistic and %(3) is triple smoothed
series is extracted from different statistical year books
statistic and a is the exponential smoothing constant.
and statistical pocket books (statistical year book of
The estimates AO(T),A1(T) and A2(T) are given by
Nepal, 1987, 1989, 1991 and 1995, statistical pocket
Ao(T)=3ST- 3 ~ ~4- ( ~ ) ...(7) book 1982,1996 and 2000, publications o m e Central
~ ~ ( T ) = a ( 2 ( 1 -((6-5a)~~2(5-4a)s~(~)+(4-3a)S+~))
a)~]-~ ...(8) Bureau of Statistics, Thapathali, Nepal).
A2(T)= a2(1-a)-2( ~ ~ - 2 ts ~~ '~ ~ () ~ 1 ) ...(9) Fig No.1 shows the time series exhibiting an
The t period ahead forecasting equation is given approximate quadratic trend (can be verified by least
by replacing the values ofAo(T), Al(T) and A2(T) in square estimation).
equation (3).
Est. YT+t (T) 4 2 (1-a)2)-1{6(1-a12+(6-5a)at+a2t2]sT
- (2(1-a)2)-1{6(1-a12+ 2(5-4a) at + 2a2t2)
+ (2(1-a)2)-1{2(1-a)2+ (4-3a) at + a421 s ~ (...(~10) )
In order to begin the procedure, the initial values
of the Smoothed Statistics must be supplied.They are
obtained by solving the equations given below 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27

so=~ ~ ( [a~-~(l-a)
0)- ~ ~ ( 0 ) + ( 2 a ~ ) -(2-a) - a ) ...(11)
~ ( l A2(0) Year

Figure 1 Line Graph of the Time Series

O ) - ~ A1(0)+2(2a2)-l(l-a)(3-2a)
S ~ ( ~ ) = A ~ ((a]-l(1-a) A2(0)...( 12)
s ~ ( ~ ) = A ~-3{a]-'(1-a)
(o) ~~(0)+3(2a~)-l(l-a)(4-3a)~~(0) ...( 13)
Where Ao(O), A1(0) and A2(0) are the initial In X axis, 1=1971,27=1997
estimates of the model parameters. One way of The initial values of the least square estimates of
obtaining them are by the principle of least squares the model parameters are found to be
estimates of the time series. Once the initial values of b0=41816.78718,
the smoothed statistic are obtained, the forecasting b1=7570.1732 and
procedure is quite straightforward. In each time period,
the single, double and triple smoothed statistics are
computed. These smoothed statistics are then entered Changing the origin to time periodT, the estimates
into the forecasting equation. The equation yields a t are
period ahead forecast (Bowerman & Connell 1979). Ao(0) = 41816.78718,
A1(0) = 7570.1732 and
Choice of Smoothing Constant (a) A2(0) =422.92998
When an exponential smoothing procedure is used as For the chosen values of the smoothing constants
a forecasting tool, the forecaster must specify a value a = 0.05,0.08 and 0.15, the initial Smoothing Statistics
of the smoothing constanta. The constant determines are given in table 1.
the extent to which the past observations influence the
forecast. In practice, it has been found that values of Table 1. Initial Smoothing Statistics for different
between 0.01 and 0.3 work quite well (Abraham & Values of Smoothing Constant.
Ledolter 1983). Now the question may be if we choose
between .O1 and 0.3 which value should one choose?
One way of choosing involves the use of simulation.
For each value of a (at an interval of say 0.03), a set of
 S.L. ShresthaJNepal Journal of Science and Technology 3 (2001) 1-4

The one period ahead forecasting equations for n b l e 3. Sum of Squared Residuals (in I@) for
different values of the smoothing constant are. different Values of Smoothing Constant
a=0.05 Est.~+~(T)=3. 16067 ST-3.2687 S-+2) e0.05 ~-0.08 ac-0.15

DiagnosticChecking for ModelAdequacy

A model that fails in diagnostic checking for model
adequacy will always remain a suspect.Therefore, it
is essential that model selected for forecasting should
The one period ahead forecasts and the residuals satisfy the important diagnostic checking. The
for a=0.05 are shown in table 2. Table 3 shows the following diagnostic checking are considered here.
sum of squared residuals (in 109) for different values
Test for Goodness of Fit
of the smoothing constant. The value of a for which
the sum of squared residuals is minimum is found to The computed Coefficient of Determination (2)for
be 0.05. smoothing constant 0.05 is equal to 0.9776, which is
very high. Also if the data set of predicted values and
Table 2. Estimated Values for different Values of
the observed values is classified into six classes of
Smoothing Constant
equal intervals using Sturge's rule of classifying data,
Observed Est.1 Est.2 Est.3 Res. 1
the computed value of Chi Square is only 2.7 ( I b Chi
Square value for level 0.05 is 11.07) which is
insignificantsuggesting that the predicted values fit to
the observed time series (for a=0.05).
Autocorrelation and Partial Autocorrelation Test
for Residuals
The assumption of no serial correlation between error
variables (residuals) should not be violated while using
the technique of exponential smoothing as in case of
linear regression approach. The evidence of lack of
significantautocorrelationsand partial autocorrelations
for different lag periods is essential for verification of
model adequacy test. The residual analysis shows that
non of the autocorrelations and partial autocorrelations
are individually statistically significant applying
Bartlett approximation for significance of correlation
at 95% confidence interval.This implies that the model
can be accepted for forecasting ('hble 4 & 5).
Table 4. Autocorrelationsfor Residuals for Smoothing
Constant 0.05
Lag Correlation Error
1 0.118 0.192
2 -0.256 0.195
3 -0.252 0.207
4 0.206 0.218
5 0.070 0.225
6 -0.094 0.226
7 -0.334 0.228
8 0.020 0.245
9 0.087 0.245
10 0.05 1 0.246
11 -0.095 0.247
12 0.176 0.248
13 -0.179 0.253
* 1P.AF 419956.52 421213.29 427992.76 14 0.075 0.257
* lP.A.F= one period ahead forecast 15 0.075 0.258
16 0.004 0.259
 S.L. ShresthaINepal Journal of Science and Technology 3 (2001) 1-4

Table 5. Partial Autocorrelations for Residuals for Discussion and Conclusion

Smoothing Constant 0.05 The results of using the triple exponential soothing
Lag Correlation Error technique to the time series of the annual number of
1 0.118 0.192 tourist arrivals in Nepal demonstrates that the method
is indeed useful and suitable to such time series data
2 -0.274 0.192 exhibiting an approximate quadratic trend. This is
3 -0.198 0.192 statistically supported by performing some important
4 0.218 0.192
diagnostic checks. Howeve~the use of this technique
consumes a lot of time and computations without the
5 -0.106 0.192 support of appropriate computer programming. This
6 -0.059 0.192 is due to the simulation arising from different values
of the smoothing constant. For three different values
7 -0.262 0.192
of a considered for simulation the value equal to 0.05
8 0.022 0.192 generates a minimum sum of squared residuals. Other
9 -0.084 0.192 values of a has been left out because of lengthy
computations arising due to simulation. For the reason
10 -0.060 0.192
there can be another value of a which may generate a
11 0.015 0.192 smaller sum of squared residuals.
12 -0.267 0.192
References
13 -0.254 0.192
Abraham, B. & J. Ledolter 1983. Statistical Methods
14 -0.100 0.192 for Forecasting.John Wiley and Sons, Inc., USA.
15 -0.144 0.192 Bowerman, B. L. & R.T. 0' Connell1979. Eme Series
16 -0.115 0.192 and Forecasting : An Applied Approach.
Wadsworth, Inc., Belmont, California, USA.
Chatfield, C. 1996. The Analysis of lime Series, fifth
edition. Chapman and Hall, U.K.

View publication stats