10(g).

Time Series Analysis & Forecasting

Learning objectives
1. To recognize and define different component of a time series.
2. To obtain trend, seasonal index, cyclical and irregular movements by using appropriate
methods.
3. To understand the importance as well as the different methods of forecasting.
4. To measure methods of forecasting error.
. To understand and to apply criteria for choosing an appropriate forecasting method.
Introduction
!rom economic situation if it relates to an individual form, industry or a nation as a whole, we can
observe a continuous movement of economic activity. "n order to describe this flow of economic
activity, the statistician uses a time series. The term #Time $eries% means a set of observations
concurring any activity against different periods of time. The duration of time period may be
hourly, daily, wee&ly, monthly or annually.
!ollowing are few examples of time series data'
a( )rofits earned by a company for each of the past five years.
b( *or&ers employed by a company for each of the past 1 years.
c( +umber of students registered for ,- examination in the institute for the past five years.
d( The wee&ly wholesale price index for each of the past 3. wee&.
e( +umber of fatal road accidents in /elhi for each day for the past two months.
Im!ortance o" Time series Analysis
There are several reasons for underta&ing a time series analysis. !irstly, the analysis of a time
series helps to understand the past performance. $econdly, a time series analysis helps directly
in business planning. - firm can &now the long0term trend in the sale of its product.
Thirdly, a time series analysis helps one to study such movement as cycles that fluctuates in two
or more series regarding the rate or type of growth.
!rom the above discussion, we can conclude that time series analysis has great advantages in
business and industry.
#$%&$'('TS $F A TI%( S()I(S
- time series may contain one or more of the following four components'
1. $ecular trend 1T(
2. $easonal variation 1$(
3. ,yclical variation 1,(
4. "rregular variation 1"(
There are two approaches for the relationship amongst these components.
1a( 2 3 T$," 1multiplicative model(
1b( 2 3 T4$4,4" 1additive model(
*here 2 is the result of the four components.
The effects of these components might be multiplicative or additive or might be a combination of
several other form. 5n account of this, different assumptions will give different results. 6owever,
multiplicative components is most fre7uently used.
T)('*
The trend is the long0term movement of a time series. -ny increase or decrease in the values of
a variable occurring over a period of several years gives a trend. "f the values of a variables
remain statutory over several years, then no trend can be observed in the time series. To study
the growth in industrial production from the year 188 to 2.., we need to find the trend values
in industrial production for this time period which may be increasing or decreasing.
These trends may be either linear or non0linear. There are other types of trends li&e parabolic 1or
7uadratic( and logarithmic 1or exponential(. 6owever we are more concerned here with straight
line 1i.e. linear( trends.
The various methods of fitting a straight line to a time series such as free hand method, the
method of semi0averages, the method of moving averages and the method of least s7uares.
1a( The !reehand 9ethod' "t is the simplest method of finding a trend line. The procedure
involves first the plotting of the time series on a graph and fitting a straight line through the
plotted points in such a way that the straight line shows the trends of the series.
e.g. $ales of products #-% 188802..
2ear 1x( $ales 1y( 1:s. in 9illion(
1888 1.
2... 2.
2..1 1
2..2 2
2..3 3;
2..4 3
2.. 4.
Ta&ing < axis as time 1in years( and y axis to the sale of product in million rupees , sales figures
are plotted and =oining them by straight line we get fluctuating straight lines through which we get
an average straight line. "t may be noted that the free hand method is not an accurate method of
filling a trend as different persons may fit different trend times to the same set of data.
y
.
4.
3.
2.
1.
x
1888 ,.. ,.1 ,.2 ,.3 ,.4 ..
12ears(
$
a
l
e
s

1
9
i
l
l
i
o
n

i
n

:
s
.
(
1b( T+e %et+od o" Semi,averages *hen the method of semi0averages is used, the given
time series is divided into two parts preferably with the same number of years. The average
of each part is calculated and then a trend line through these average is filled.
)roduction from 188802..>
2ear < 2 -verage
1888 . 1.
2... 1 12 >.?4 3 1
2..1 2 1@
2..2 3 2.
2..3 4 2.
2..4 2
2.. > 23 1..?4 3 2
2..> ; 32
The average of first part of the data is 1 and that of the second part is 2. -s 1 is the average
of 1888, 2..., 2..1 and 2..2. 1 is plotted in between 2... and 2..1 which is the middle of the
4 years period. $imilarly 2 is plotted in between 2.4 and 2... Then these points are =oined by
a straight line which is a semi0average trend line.
3
3.
2
2.
1
1.

. 1 2 3 4 > ; @
2ear 1188802..>(
1c( T+e met+od o" moving average The method of moving averages is used not only to fit
trend lines but also to seasonal and cyclical variation.
e.g. !ind out the three year moving averages starting from 18@8.
2ear' 18@8 188. 1881 1882 1883 1884 188 188> 188; 188@ 1888 2...
$ales' 1. 1 2. 2 1 12 1 24 1 21 1 24
1:s. in million(
$olution' Ta&ing the first three years 18@8, 188. and 1881, we add up their sales 1. 4142. 3
4. This figure is written is column 13( against the mid year that is 188.. +ow dropping the year
18@8 and including the year 1882 in the calculation and thus moving in this manner the process
is continued until we reach the last three years 188@, 1888 and 2.... "t may be noted that there
are no moving total for the 1
st
and last year.
The next step involves the calculation of moving average by dividing the moving total by 3.
)
r
o
d
u
c
t
i
o
n
Table
2ear $ales 3 year moving 3 year moving
:s. in 9illion total average
18@8 1. 00 00
188. 1 4 1
1881 2. >. 2.
1882 2 >. 2.
1883 1 2 1;
1884 12 42 14
188 1 1 1;
188> 24 4 1@
188; 1 >. 2.
188@ 21 4 1@
1888 1@ >3 21
2... 24 00 00
(""ective A!!lication o" %oving Average %et+od
To ensure result of moving average method to be appropriate and effective, it is re7uired to
ascertain first whether a regular periodic cycle in the time series exists. "n several cases one
would find that there is a certain regularity in the series to allow the use of the moving average
method. "f may be also noted that if the basis nature of the time series is linear, it will give a
linear trend. "n case of curvilinear nature, the trend will be curve. 9oreover moving average
method helps to element seasonal fluctuation, for a time series.
%et+od o" Least s-uares -mong the method of fitting straight line to a series of data, this
method is the most fre7uently used method. The e7uation of a straight line is 2 3 a4 b where <
is the time period, say, year and 2 is the value of the item measured against time, a is the 20
intercept and b is the coefficient of < indicating slope of the trend line.
"n order to find a and b the following two e7uations are solved'
A2 3 ax 4 b Ax
A<2 3 a Ax 4 bAx
2
*here n is the total number of observations in a series. These e7uations are called normal
e7uations.
e.g. 5ld number of years'
2ear' 188> 188; 188@ 1888 2...
2 2;. 2@ 28 31 33.
!it a linear trend to these data using method of least s7uares'
Solution
2ear x y xy x
2

188> 02 2;. 04. 4
188; 01 2@ 02@ 1
188@ . 28 . .
1888 1 31 31 1
2... 2 33. 0>>. 4
Total . 3Ax 1,48 3 Ay 1. 3Ax y 1. 3 Ax
2

$ubstituting these values in two normal e7uations.
*e get
148 3 a 4b 1.( BBBBBB..11(
1. 3 a 1.( 4b 11.( BBBBB12(
C7uations 11( gives a 3 1,48? 3 288
C7uations 12( gives b 3 1.?1. 3 1
The straight line trend we get is
2 3 288 41x 1origin' year 188@(
1x 3 10 year unit(
e.g. (ven 'umber o" years
2ear' 18@ 18@> 18@; 18@@ 18@8 188.
2 1 14 1@ 2. 1; 24
!it a trend to these data using method of least s7uares.
Solution
2ear x y xy x
2

18@ 0 1 0; 2
18@> 03 14 042 8
18@; 01 1@ 01@ 1
18@@ 1 2. 2. 1
18@8 3 1; 1 8
188. 24 12. 2

Total . 3Ax 1.@ 3 Ay > 3Ax y ;. 3 Ax
2

Two normal e7uations are'
A2 3 an 4b Ax
A<23 aAx 4bAx
2
$ubstituting the values in the two e7uations, we get
1.@ 3 >a 4b 1.( BBBB11(
> 3 a1.( 4b 1;.(BBB12(
$olving we get a 3 1.@?> 3 1@
-nd b 3 >?;. 3 ..@
The trend e7uation is 2 3 1@ 4..@
5rigin year' 9id of 18@;0@@
Dnit of < 3 E year
+ote' 1i( )utting values of x for different years we get values of y which are trend values 1y(
1ii( !or forecasting, to calculate the value of x, for a given future year and substituting the value of
x in the trend e7uation to get the forecasted value.
.ig+er *egree &olynomial Trends
$o far we have discussed linear trend which in fact is a polynomial trend of the first degree.
6owever instead of straight line trend, some parabolic trend may be best fit when we have to use
second degree polynomial trend. This trend e7uation is
2 3 a4bx4cx
2
"n this also method of least se7uence is to be used where there are normal e7uations
A 2 3 na 4b Ax 4 c Ax
2
Axy 3 a Ax 4b Ax
2
4c Ax
3
Ax
2
y 3 aAx
2
4bAx
3
4c Ax
4
e.g. fit 7uadratic trend'
2ear ' 188 188> 188; 188@ 1888
)roduction ' 2 4 @ 14 22
5f cars
1#....(
Solution
2ear x x
2
y xy x
2
y x
3
x
4

188 02 4 2 04 @ 0@ 1>
188> 01 1 4 04 4 01 1
188; . . @ . . . .
188@ 1 1 14 14 14 1 1
1888 2 4 22 44 @@ @ 1>
Total .3Ax 1.3Ax
2
.3Ay .3Axy 1143Ax 2y .3Ax
3
343Ax
4

-pplying these values in the above e7uations, we get
. 3 a 41.c BBBBBBB..11(
. 3 1.b BBBBBBBBB.12(
114 3 1.a 434c BBBBBB.13(
!rom 12( we find b 3 .?1. 3
-gain, . 3 a 41.c BBBBBB..11(
114 3 1.a 434 cBBBBB...12(
13( 114 3 1.a 434c
11(2 1.. 3 1.a 42.cFFFF
14 3 14,
c 3 1
!rom 11(, . 3 a 41. a 3 @
:e7uired trend e7uation is
2 3 a4 bx 4cx
2
2 3 @4x 4x
2
origin year 188;
Dnit of x 3 1 year
(/ercise
11( -ssuming an additive model, obtain 3 year0moving average to obtain the trend free series for
year 1882 to 188> from the following data'
2ear' 1881 1882 1883 1884 188 188> 188;
Cxport' 12> 13. 13; 141 14 1 18
1:s. in la&hs(
-nswer' 131, 13>, 141, 14;, 13
12( Dse a 4 yearly moving average method to calculate trend for the following data'
2ear' 18;@ 18;8 18@. 18@1 18@2 18@3 18@4 18@ 18@> 18@; 18@@
)roduction' >14 >1 >2 >;@ >@1 > ;1; ;18 ;.@ ;;8 ;;
-nswer' FFF, FFF, >4@, >>1.4, >;4.>3, >@;.8, >8..1, ;1.3, ;3.@, FFF, FFFF.
13( Give below are the figures of production 1in million tonns( of wheat
2ear' 18@8 188. 1881 1882 1883 1884 188
)roduction' @. 8. 82 @3 84 88 82
!it a straight line trend to these figures.
H-ns' 2 3 8. 42x, origin year 1882, x unit 3 1 yearI
14( !it a straight line by the least s7uares method to the following figures of production of $ugar
factory'
2ear 18;8 18@. 18@1 18@2 18@3 18@4 18@
)roduction ;> @; 8 @1 81 8> 8.
Cstimate the production for 18@>.
-ns' 23@@42xJ 8> 1#... tons(
1( The following table given the annual profit 1in thousand :s.( in a factory
2ear' 1881 1882 1883 1884 188 188> 188;
)rofit >. ;2 ; > @. @ 8.
!it a straight line trend by the method of least s7uares. -lso find annual trend of the profit. -lso
calculate the pro=ected profit for 188@.
H-ns.' 2 3 ;.2@> 44.321x, 4.321, 82.; 1#... :s.(
1>( Kelow are given the annual consumption in thousand tons of a town
2ear 18@ 18@> 18@; 18@@ 18@8 188. 1881
,onsumption ;. ; 8. 81 8 8@ 1..
1#... tons(
!it a straight line trend by the method at least s7uare and indicate the trend values by ta&ing
18@@ as origin. $hift the origin to 188. and find out the revised trend e7uation.
,onstruct the annual trend e7uation into monthly trend e7uation.
H-ns. 2 3 @@.43 4 ..4x, 2 3 [email protected] 4..x, 2 3 ;.3; 4...3 xI
1;( !it a 7uadratic trend to the following data'
2ear 18@@ 18@8 188. 1881 1882
Lalue 1@. 182 18 2.4 2.2
H-ns.' 2 3 18;.;4 4.>x 01.;x
2
I
*etrending The process of eliminating the trend is referred to as detrending and the trend itself can be
represented either as a straight line or some type of smooth curve. *hat is left after elimination of trend
is short term variation, expressed as '
$ x , x " Dnder multiplicative model,
-nd $ 4 , 4 " Dnder additive model.
T+e #alculation o" Seasonal 0ariations
These are short0term movements which recur after a period. The period is normally less than a year0it
may be a month or a 7uarter of the year. !or example, the prices of grain vary between the harvest
season and the non0harvest season.
!or purpose of time series analysis we isolate the seasonal element. There are two ma=or reasons for
this ' 1i( to study seasonal variationJ and 1ii( to eliminate them.
". $easonal Lariations in themselves give a clear idea about the relative position of each season on
the basis of which it is possible to plan for the season.
"". "t is necessary to eliminate seasonal factor to get the cyclical or irregular factors. The is called
deseasonalization of the data.
!irst we ta&e an illustration of isolating short0term and seasonal fluctuations by the additive model.
Illustration
2ear Muarter 5utput 9oving 9oving 9oving $hort0term
M 2 total total -verage !luctuation
14( 1@( 1T( 20T3$4,4"
18;1 1 3
2 @>
312
3 >; >2; ;@ 011
31
4 124 >3 @2 442
33@
18;2 1 3@ ;.. @@ 0.
3>2
2 1.8 ;;> 8; 412
414
3 81 @3; 1. 014
423
4 1;> @8 112 4>4
4;2
,alculations of seasonal fluctuations '
2ear M.1 M.2 M.3 M.4
18;1 011 42
18;2 0. 412 014 4>4
18;3 0;3 43. 032 4@>
18;4 0@ 42> 02. 4@1
18; 082 433
Total 03.. 41.1 0;; 42;3
-verage 0; 42 018 4>@
$t+er %et+ods o" "inding s+ort !eriod variations
". $imple -verage ' The method is very simple' average the data by months or 7uarters or years
and them calculate the average for the period. Then find out, what percentage it is to the grand
average.
2ear Muarter 5utput 9oving 9oving 9oving $hort0term
M 2 total total -verage !luctuation
14( 1@( 1T( 20T3$4,4"
18;3 1 4; 8; 12. 0;3
4@
2 1@ 1.2. 12@ 43.
3
3 1.4 1.@4 13> 032
48
4 22> 111; 14. 4@>
>@
18;4 1 >1 11>> 14> 0@
8@
2 1;; 121. 11 42>
>12
3 134 123 14 02.
>23
4 24. 12; 18 4@1
>2
18; 1 ;2 1311 1>4 082
>8
2 2.> 13@ 1;3 433
;2>
3 141
4 3.;
$easonal "ndex 3
100
Monthlyorquarterlyaverage
GrandAverageofthemonthsorthequarters

$ame results are obtained if the totals of each month or each 7uarter are obtained instead of the
average of each month or each 7uarter.
Illustration
2ear 1st Muarter 2nd Muarter 3rd Muarter 4th Muarter
18;. 3.; 4.1 3.3 3.
18;1 3.; 3.8 3.> 3.
18;2 4.. 4.1 3.3 3.1
18;3 3.3 4.4 4.. 4.1
Total 14.; 1>. 14.2 14.2
Muarterly 3.>@ 4.12 3. 3.
-verage
$easonal "ndex 8@.; 11..@ 8.3 8.3
The grand average 3
3.68 4.12 3.55 3.55 14.90
3.72
4 4
+ + +

$easonal "ndex 3
100
QuarterlyAverage
GrandAverage

$easonal "ndex for the 7uarters '
1
st
'
3.68
100 98.7
3.72

2
nd
'
4.12
100 110.8
3.72

3
rd
and 4
th

3.55
100 95.3
3.72

(/am!le
Dsing the method of monthly averages determine the monthly indices for the following data of production
of a commodity for the year 18;8,18@.,18@1.
Solution
6ere, instead of dividing the monthly average by the overall average, we divide the monthly totals by the
average of totals.
&roduction in Lac1s o" tons
2ear 18;8 18@. 18@1 $easonal "ndex
9onth Totals 1percentages(
Nanuary 12 1 1> 43 1.4.8
!ebruary 11 14 1 4. 8;.
9arch 1. 13 14 3; 8..2
-pril 14 1> 1> 4> 112.2
9ay 1 1> 1 4> 112.2
Nune 1 1 1; 4; 114.>
Nuly 1> 1; 1> 48 118.
-ugust 13 12 13 3@ 82.>
$eptember 11 13 1. 34 @2.8
5ctober 1. 12 1. 32 ;@
+ovember 12 13 11 3> @;.@
/ecember 1 14 1 44 1.;.3
Total 482
-verage 41
Oimitation' -lthough simple, this method is not very scientific for it assumes as if there is no trend
component in the series, i.e., the original series comprise, , x $ x ". $ince most economic series have
trends, the index computed by this method is actually an index of seasonal variation plus trend. !urther,
the effect of cycles on the original values may not be eliminated by the averaging process. "t is only in
case that the duration of the cycle coincides with the number of months or 7uarters included in the
average that the cyclical fluctuations will be avoided. "n the absence of this the seasonal index will also
include the effect of trend.
1ii( :atio0to0Trend 9ethod ' This method is an improvement over the previous method because
this assumes that seasonal variation for a given month is a constant fraction of trend. This method
presumably isolates the seasonal factor in the following manner '
T S C I
S C I
T


:andom elements 1"( are supposed to disappear when the ratios are averaged. !urther, a carefully
selected period of years used in computation is expected to eliminate the influence of cyclical
fluctuations 1,(.
9onth 18;8 18@. 18@1
)roduction in lacs of tons
Nanuary 12 1 1>
!ebruary 11 14 1
9arch 1. 13 14
-pril 14 1> 1>
9ay 1 1> 1
Nune 1 1 1;
Nuly 1> 1; 1>
-ugust 13 12 13
$eptember 11 13 1.
5ctober 1. 12 1.
+ovember 12 13 11
/ecember 1 14 1
(/am!le !ind seasonal variations by the ratio0to0trend method from the data given below.
Solution
!irstly , we have to determine the trend value for each 7uarter as follows '
2ear 2ear Muarterly /eviation Trend Lalues
Total -verage from mid0
year
2 < <2 <2
18>8 14. 3 02 0;. 4 32
18;. 1@. 4 01 04 1 44
18;1 2.. . . . . >
18;2 2>. > 41 4> 1 >@
18;3 34. @ 42 41;. 4 @.
2@. 12. 1.
The value of the constants a and b in the e7uation of a straight line are as following '
280
56
5
Y
a
N

,
2
120
12
10
XY
a
X

The 7uarterly increment therefore is 12?433.

$ince there are 4 7uarters we have to ta&e the point in between the 2
nd
and the 3
rd
7uarter each year.
Thus following are trend values for various 7uarters of the respective years.
Muarters
2ear 1st 2nd 3rd 4th
18>8 3. 4. 3> 34
18;. 34 2 . 44
18;1 4. @ 4 4@
18;2 4 ;> >@ >2
18;3 @. 82 @> @2
$ince the total of the seasonal index is 4.3..@ 3 182.@@[email protected]>[email protected](.
Cach index has to be ad=usted by multiplying it by
400
403.08
and the final indices are thus available.
1iii( :atio0to09oving -verage 9ethod ' "n this method instead of calculating the annual trend by the
method of least s7uares the moving average is used. "t is more suitable when seasonal variations for
the months are to be calculated'
Lin1 )elative %et+od
1i( To ,alculate $easonal Oin& :elative for each season where Oin&
:elative 3
100
Pr
CurrentSeasonFigure
eviousSeasonFigure

1ii( ,alculate average of the lin& relatives for each season. -rithmetic mean is generally used but
even median could be used.
1iii( ,onvert there averages into chain relatives on the basis of first season.
1iv( ,alculate the ,hain relative of the first season on the basis of the last season.
1v( !or correction, the chain relative of the first season calculated by the first method is deducted
from the chain relative of the first season calculated by the second method. The difference is
divided by the number of seasons. The resulting figure multiplied by 1,2,3, etc. are deducted
respectively from the chain relatives of 2
nd
,3
rd
,4
th
, etc. seasons.
2ear 1st 2nd 2early Trend 3rd 4th
Lalues
18>8 2;. 3.. 32 33. 3>.
18;. 38. 42. 44 4. 4@.
18;1 1. 4. > 4. >..
18;2 >3. >>. >@ >8. ;2.
18;3 ;. ;@. @. @1. @4.
-ctual Muarterly Lalues -s P of 7uarterly Trend Lalues '
18>8 1.8.1 131.1 1.;. 83.1
18;. @>.1 122.4 1.8.8 8..;
18;1 ;;.; 1.>.4 83.8 ;8.3
18;2 @.. 114.3 8;.@ @.
18;3 1.>.. 11;.1 1.. 8;..
Total 4>3.8 81.3 14.> 44.>
-verage 82.;@ [email protected]> 1.2.82 @8.12
$easonal "ndex
-d=usted 82.. 11;.4 1.2.1 @@.4
1vi( The seasonal indices are available when we express the corrected chain relatives as
percentages of their respective averages.
(/am!le -pply the method of lin& relatives to the following data and calculate seasonal indices.
Muarter 18>8 18;. 18;1 18;2 18;3
" 4. 4.@ 4.8 .2 >..
"" .4 .> >.3 >. ;..
""" ;.2 >.3 ;.. ;. @.4
"L >.. .> >. ;.2 >.;
Solution Oet us first calculate lin& relatives of the seasonal figures.
These are given below in the table '
Muarters
2ear " "" """ "L
18>8 12. 133 @3
18;. @. 11; 113 @8
18;1 @@ 128 111 83
18;2 @. 12 11 8>
18;3 @3 11; 12. @.
9ean @2.@ 121.> [email protected] @@
,hain :elatives 1..
100 121.6
100

121.6 118.4
100

143.9 88
100

3 121.> 3 143.8 3 12>.>

,orrected chain 1.. 121.>01.2 143.802.4 12>.>03.>
3 12..4 3 141. 3 123
$easonal "ndices
100 100
121.2
120.4 100
121.2
141.5 100
121.2
123.0 100
121.2

3 @2. 3 88.4 3 11>.; 3 1.1.

'ote The lin& relatives are calculated by dividing the value of one 7uarter by the value of the previous
7uarter and expressing them as percentages.
!or example, Oin& :elative of 18>80Muarter "" 3
5.4
100
4.5

3 12.
Oin& :elative of 18>80Muarter """ 3
7.2
100
5.4

3 133
The figure correction has been calculated as follows '
,hain relative of the first 7uarter on the basis of first 7uarter 3 1..
,hain relative of first 7uarter on the basis on the last 7uarter 3
82.8 126.6
104.8
100

The difference between these chain relatives 3 1.4.@ Q 1.. 3 4.@

/ifference per 7uarter
4.8
12
4

$easonal variation indices have been calculated as follows
-verage corrected chain relatives 3
100 120.4 141.5 123.0
121.2
4
+ + +

$easonal variation indices therefore will be calculated as '

100
121.2
Correctedchainrelative

*eseasonali2ation The process of eliminating seasonal fluctuations or deseasonalization of data

consists of dividing each value in the original series by the corresponding value of the seasonal index.
The following table will illustrate this.
Muarter 5riginal Lalue $easonal "ndex /eseasonal Lalue
" 2;2 ;8.... 344
"" 3@1 12..13 31;
""" 34> 1.8.@3 31
"L 2;8 81.2>4 3.
Smoot+ing met+ods in Time Series
(/!onential Smoot+ing
The methods of analyzing the time series discussed earlier are usually employed for long term
forecasting. Kut there are numerous situations e.g. , warehouse, consumer stores and manufacturing
organizations which need future forecasts =ust over the next month, wee& or even a day, for each of the
item individually.
"t is not uncommon to find thousands and even la&hs of items in such organizations and, therefore, as
many forecasts may have to be raised to maintain sufficient stoc&s of items. Cxcess stoc&ing would
adversely affect the service to the customer or hold up assembly of the final product.
9oving -verages method could indeed be applied in such situations, but it would re7uire &eeping in the
records of the past sales figures as many in number as the moving average period. This would re7uire
too much clerical recording or in a computerized system, storage space in the computer files.
Cxponential smoothing is a 7uic&er and theoretically more sound method of short0term forecasting. "ts
ma=or, though not exclusive, use, however, is in )roduction and "nventory control.
(/!onential Smoot+ing as an Average
,onsider the 9oving -verage 1)eriod 3 4 months( for the following series '
2. 24 2@ 3.
9oving -verage 3
1
(20 24 28 30)
4
+ + +
3
1 1 1 1
20 24 28 30
4 4 4 4
+ + +
3
1 2 3 1
20 24 28 30 + + +
"t may be noted that the weight ages *1, *2, *3 and *4 assigned to each observation is e7ual
and
1
4
of that is reciprocal of the 9oving -verage period. $ince the sum of the weight is unity the
9oving -verage is a true average.
Cxponential $moothing is also an average in this sense. Kut the weight ages assigned to the past
figures are not e7ual. The most recent observation is assigned the highest weight age and it decreases
in geometric as we move towards the older observations. This is obviously more reasonable because
the most recent observation is more relevant for forecasting the future sales than an older observation.
The ratio of the geometric progression of the series of the weights is (1 ) , where

is a constant
called smoothing coefficient. Theoretically all the past observations are needed for computing the
exponentially smoothed average, that is the se7uence of the past sales figures may run indefinitely to
infinity. /espite this, as will be noted below, the method is remar&ably easy to apply in practice.
The discussion should be more clear from the following table '
1The sum of the column , gives the exponentially smoothed average, ut(
5bservations at time t *eightage -ssigned ,ol.1-( x ,ol 1K(
1-( 1K( 1,(
t
y
t
y
1 t
y

(1 )
1
(1 )
t
y

2 t
y

2
(1 )
2 2
(1 )
t
y

3 t
y

3
(1 )
3
3
(1 )
t
y

t n
y

(1 )
n
(1 )
n
t n
y

Cxp. $m. -verage.

2
1 1 2
(1 ) (1 ) .... (1 ) ...
n
t t t t n
u y y y y

+ + + +
This average may be used as the forecast for the next period , a in theory may be between . to 1' but in
practice between ..1 to ..3.
"t is to be contrasted with the moving average method in that infinite number of the past figures are being
utilized, here. This loo&s awesome but a great deal of simplification is possible.
!rom ,ol. K it is obvious that the weights are decreasing exponentially
2
[ , (1 ), (1 ) ,...] . 6ence
the name exponential smoothing.
-n essential feature of a true moving average is that the sum of the weights should be unity. Oet us
ascertain this for Cxponential smoothing by summing up the column K.
2
(1 ) .... (1 ) ...
{1 (1 ) 1 ) ...}
1 (1 )
1 (1 )
1 (1 )
n
n
n

+ + +
+ +
1

1

]
-s n approaches , (1 ) being less than 1, approaches .. Therefore, the sum of the weights is unity.
&rocedure "or #om!utations
1 1 1
(1 ) ..... (1 ) .....
n
t t n
y y y

+ + +
1 1 1 2
(1 ){ (1 ) .....}
t t
y y y

+ + +
1 1 1
(1 )
t
y

+
1 1 1 1
( )
t t t
y

+
The term in the brac&et on :.6.$R is the differences between latest observation and its forecast raised a
period earlier. "t is therefore, the error, and is denoted by e1
1 1 t t
e

+
The procedure for computations is summarized below'
1i( !ind the error by subtracting the recent average from the incoming observation.
1ii( 9ultiply error e
1
by a. This is the correction to be applied to the past average.
1iii( -dd the correction to the past coverage
1 1 t


. This gives new average u1 as the forecast for the next period. The rationale behind
this is further clarified in the following figure'

e
1
The figure gives the forecasts prior to the period t.
1 t

is the forecast for the period t raised at period 1t0
1(. The actual sales that materializes between the period t01 and t are average yt. The sales have
exceeded the forecast. "n view of this we should tend to modify the average
1 t

'e is the error. The
modification is done by increasing the average
1 t

by e,. Thus the difference between each new
incoming observation and its forecast is being utilized in &eeping the average up0to0date.
0alue o" t+e Smoot+ing #o,e""icient
,onsider the two extremes 3 . and 3 1
1i( 3.
1ii( The correction to be applied to the forecast is

3.. This implies that the earlier average stands

and no account of any difference between actual observation and its forecast is ta&en

3., therefore ,
implies that the forecasting system does not react to the incoming observations.
1iii(
1 1 1 1 1
( )
t t t t t
e y y

+ +
6ere the observation itself is ta&en as the forecast for the next period. The past series is
completely ignored.
This should suggest that a low value of gives more weightage to the past figures and less
consideration to the incoming observation. Oow values are therefore, to be used where the series is
stable and high values where the series is rather fluctuating.
Cg'1. The actual figures for the sales of bleaching, powder for the year 18;3 are shown below
in ,olumn - in the table below. The forecasts or the estimates using Cxponential smoothing
are re7uired , 3..3. "nitial forecast may be ta&en as 28.>.
+ote' 1i( Cach fig in col 1d( is obtained by adding to its earlier figure the correction in column 1c(.
1ii( "nitial estimate of 28.> was given. *hen exponential smoothing is introduced this estimate would
invariably be needed to start the system. "t may be sub=ectively guessed or simple arithmetic mean of
the past figures may be ta&en.
To summarise the discussion on Cxponential smoothing thus far'
1-( Cxponential smoothing, as against its contender the 9oving -verages 9ethod, is
appealing because of the better weight age scheme.
1K( The greatest disadvantage inherent in the 9oving -verages is relatively large amount of
historical data that must be retained to compute them. The longer, the averaging period,
the more the data that must be retained. "n Cxponential smoothing, records need to
be &ept =ust for and the latest average.
1,( Kasically, Cxponential smoothing provides a convenient, systematic method for revising the
forecast for the next period whenever discrepancy exists between the previous forecast sales for the
current period and the actual demand for the current period. "f actual demand in the current period is
higher than the forecast demand for the current period, the forecast for the next period should be
ad=usted upwards. The amount of ad=ustment is determined by the selected smoothing coefficient, the
9onth $ales Crror ,orrection !orecast
1-( 1K( 3 1-( 0 1/( 1,( 3 1K( x ..3 1/(
Nan. 33 33028.>33.4 3.4 x ..3 3 1.. 28.> 1"nitial(
!eb 31 31.3..>3.4 ..4 x ..3 3 ..1 28.>41 3 3..>
9ar 34 3403..;33.3 1 3..>4..133..;
-pr 32 32031.;3..3 ..1 3..;41331.;
9ay 3; 3;[email protected] 1.> 31.@
Nun 3> 3>033.432.> ..@ 33.4
Nul 34 34034.230..2 0..1 34.2
-ug 32 32034.1302.1 0..> 34.1
$ep 41 41033.3;. 2.2 33.
5ct 44 44.3.;[email protected] 2. 3.;
+ov 44 [email protected].@ 1.; [email protected]
/ec . .038.831..1 3.. 38.8
greater the constant the greater the ad=ustment and vice versa. "n fact, if the behavior of the sales series
changes say, from a stable one to a fluctuating one, may be changed to be responsive to the new
behavior.
(/istence o" Trend
The students might note in the solved example above by comparing column - with / that the forecast
tends to be lower than the actual sales. This is because there exists a trend in the sales figures. The
sales which is around 3. at commencement of the year 18;3 have risen to . at the end.
The diagram below brings out the behaviour of the Cxponentially smoothed average with regard to
series depicting trend.
e
1
Thus the Cxponential smoothed average lags the sales series with trend. The following
e7uations Hproof having been omittedI may be used, where trend is expected'
t
u gives the necessary forecast for the next period at time period t.
(/am!le 3 (/!onential Smoot+ing "or series 4it+ Trend.
-ssumed ' "nitial Trend 3 03..., "nitial -verage 3 ;...., 0.1
1i(
1 t t t
u u e

+
1ii(
1 1
( ) (1 )
t t t t
u u

+
1iii(
1
1
t t t
u u

_
+ +

,
#onsider &eriod 1.
1 1 1
1 1 0
1
1
( ) 750 0.1(750 750) 747
( ) (1 ) 0.1( 3.00) 0.9( 3.00) 3.00
1
1 747 10( 3.00) 717
o o
o
t
u u e y u
u u
u u

+ +
+ +

_
+ + +

,
!orecast Crror 3 ;1;0>;. 3 4;. 1Cntered in the last column(
#onsider &eriod 3
2 1 1 1 2 1
2 2 1 1
2 2 2
( ) 747 0.1(670 747) 739.30
( ) (1 ) 0.1(739.30 747) 0.9( 3.00) 3.47
10 739.30 34.70 704.60
u u e u y u
u u
u u

+ + +
+ +
+
!orecast Crror. 3 ;.4.> Q >@. 3 24.>
"t is to be noted that the trend factor ,
1
is also recursively exponentially smoothed for every
period.
The student may carry on with the rest of calculation for practice by verifying with the figures in
the table.
)eriod $ales $moothed $moothed !orecast !orecast
-verage Trend Crror
. ;. 03...
1 ;2. ;4; 03...
2 >;. ;38.3 03.4; ;1;.. 4;...
3 >@. ;33.3; 03.;2 ;.4.> 24.>.
4 ;4. ;34..3 03.2@ >8>.2 043.;8
;2. ;32.>3 03..8 ;.1.3 01@.;
> 84. ;3.3; 0..;1 ;.1.; [email protected]@
; 1.2. ;@...3 2..3 ;4>.3 02;3.;1
@ 122. @24..3 >.23 @...3 0418.>3
8 12>. @>;.>2 8.8> @@>.3 03;3.;1
1. 13.. 81..@> 13.28 8>;.3 0332.;4
11 118. 83@.;@ 14.; 1.43.@ 014>.23
12 1.@. 82.8. 14.>8 1.@>.3 >.3.
$!timum 0alue o" t+e Smoot+ing #oe""icient
"n the above example, the forecast errors are given in the last column. These, when s7uared
and accumulated, give a measure of the discrepancy between the forecasts and the actual sales.
$imilarly, other values of
0.2, 0.3,
etc. may be tried. The value that gives the least s7uared
diviation of error would be optimum

. "n practice, it is not infre7uent to assign the value of a

from experience. *here the computer activities are available, the above method of least
s7uared deviations may be carried out for groups of items depicting similar sales behaviour.
Limitation o" (/!onential Smoot+ing
1i( The method is useful for short0term forecasting only.
1ii( "t relies solely on the past history of sales. There are cases where sub=ective estimates
may provide better forecasts. There have been attempt to complement the twoJ forecasts
from Cxponential smoothing and sub=ective estimates, successful to a great extent, but
not fully satisfactory.
Furt+er (/tension
The method of Cxponential smoothing is 7uite versatile and' trend can be extended to account
for seasonal variations. Nust as the initial, forecast and trend factors are estimated and then
recursively ad=usted in the light of the discrepancy between the forecast and the incoming
observationJ seasonal factors can be similarly initially estimated, incorporated in the model and
then recursively ad=usted. !or short0term forecasting, it has almost universally superseded the
9oving -verage method. The method is particularly favoured with computerized applications
because of the minimal data items that need to be stored.
!orecasting using Time $eries
$ince economic and business conditions very over time, business leaders must find ways to
&eep in touch with the effects that such change will have on their particular operations. 5ne
techni7ue which business managers may use as an aid in planning for future needs is
forecasting. There are basically two types of forecasting called 7ualitative and 7uantitative.
Mualitative forecasting methods are useful when historical data are unavailable. Muantitative
methods ma&e use of historical data for forecasting. Muantitative forecasting methods can be
subdivided into two types, namely, time series and causal. ,ausal forecasting methods involved
the determination of factors which relate to the variable to be predicted. 5n the other hand, time
series forecasting methods involve the pro=ection of future values of a variable based entirely on
the past and present observation of that variable. "n this section we will see the various
forecasting methods using time series.
1i( 9ean !orecast ' The simplest forecasting method in which for the time period t. we
forecast the value of the series to be e7ual to the mean of the seriesJ
t
y y
This method is not ade7uate as trend effects and the cyclical effects are not ta&en into
account in this.
1ii( +aSve forecast ' "n this method, by ta&ing advantage of the fact that there may be high
correlation between successive pairs of values in a time series, we forecast the value,
for the time period t, to0be e7ual to the actual value observed in the previous period t
that is, time period 1t Q 1('
1 t t
y y

1iii( Oinear Trend !orecast ' "n this method , a linear relationship between the time and the
response value has been found from the linear relationship.
t
y a !X +
where < will be found from the value of t and a and b are constants.
1iv( +on0linear Trend !orecast ' "n this method, a non0linear relationship between the time
and the response value has been found again by least0s7uares method. Then the
value, for the time period t , will be calculated from the non0linear e7uation . i.e.,
2
t
y a !X cX + +
where <0value will be calculated from the value of t.
1v( !orecasting will Cxponential $moothing ' "n this method, the forecast value for the
time period t is found using exponential smoothing of time series. $pecifically, at the time
period t.
1 1
( )
t t t t
y y y y

+
where
t
y
the forecasted value for time period t 4 1 J
1 t
y

3 the forecasted value for time period t. '
t
y
the observed value for time period t.
*e will see the usefulness of these formulae with an example.
(/am!le 5 The following time series data are given. !ind the forecast for various years using
mean forecast, naSve forecast, linear trend forecast, non0linear trend forecast.
Solution
*e list the various forecasting values in a table'
!rom the table, one can easily see that the observed and the forecast values using 7uadratic
trend setting method do not differ much. 6ence, here forecasting for future years can be done
using 7uadratic trend setting. !or example the forecast for 1881 will be :s.4@.>8 crores. This is
found from the 7uadratic e7uation
t
y
24.@; Q ..1; < 4 ..>8 <
2
with base year as 18@ i.e. <3 t
Q 18@.
Table 3 #om!arison o" "rom "orecasting met+ods.
2ear 18@. 18@1 18@2 18@3 18@4 18@
$ales 1:s. "n ,rores( 24. 2.8 2;.> 3..1 34.@ 41.
TI%( S()I(S (6)
1. ,alculate the seasonal indices in the case of the following 7uarterly data in certain units
assuming no trend in the data'
2ear Muarter
" "" """ "L
18@. 38 21 2 @1
18@1 4 23 >3 ;>
18@2 44 2> >8 ;
18@3 3 2> >4 @4
1-ns' @>.4, 44.4, [email protected] , 1..@(
2. The number of production in Kombay in from Muarters of a year during the period 188;0 188;
are given below'
2ear Muarter
" "" """ "L
18@; 1> 13 14. 1@.
18@@ 12 121 12; 1>3
18@8 44 2> >8 ;
!ind the seasonal indices by Trend ratio method assuming a linear trend for the data.
1 -ns' 1., @3, @8, 123(
2ear 5bserved !orecast value
value 9ean +aSve Oinear Muadratic
y1 yt3y33..;3 yt3yt01 yt322.@ 4 yt 3 24.@;0
3.2> < ..1;< 4 ..>8<2
18@ 24. 3..;3 22.@ 24.@;
18@> 2.8 3..;3 24. 2.@4 2.3@
18@; 2;.> 3..;3 2.8 28.1 24.2;
18@@ 3..1 3..;3 2;.> 32.3> 3..3
18@8 34.@ 3..;3 3..1 3.>3 3.1;
188. 41. 3..;3 34.@ 3@.@8 41.18
3. ,ompute $easonal indices from the following data using method of Oin& :elatives.
2ear Muarter
" "" """ "L
18;@ > @ > >1
18;8 >@ >3 >3 >;
18@. ;. 8 > 2
18@1 >. 1 @
1-ns' 1.8,8@,84,88(
-n old forecast of 38 is given initially and the smoothing co0efficient is ..1. $mooth
exponentially the following observations' 41,41,34,38,3>,3,4.,3>,41,33.
-ns' 38.2,38.3@,3@.@4,3@.@>,3@.;,[email protected],[email protected],3@,1,3@,43,3;,@8(