Unit-I Forecasting SCM Date: 27-Jan-2025

Lecture 4: EC3211 – Forecasting Trend by – Holt’s

method
• In the previous lecture we Forecasting Techniques

discussed about the

New Products Existing Products
Qualitative Causal
Exponential Smoothing methods methods

Method. Bass
Diffusion Time series
Regression

• In this lecture we shall

Method for methods Grass
New Analysis
roots
Products

consider another method Naïve

Method
Market
research

for analysis of Time Moving Panel

Series namely, Trend Average Consensus

Projections. Exponential
Smoothing
Historical
Analogy

• The method discusses Trend Delphi

decomposing of a Time Projections Method

Different Forecasting Methods

1
 Series into its components.
• Decomposing helps in forecasting the trends effectively.
A time series is a set of values observed sequentially
through time which we have studied in Exponential
Smoothing. 𝐹𝑡+1 = 𝐹𝑡 + 𝛼(𝑥𝑡 − 𝐹𝑡 )
• Trend component: At time t, that reflects the long-term
progression of the series (secular variation).
Linear

Growing Declining
Market share Market share

Sales

Sales
➢ A trend exists when there is an increasing
Time Time

or decreasing direction in the data. Growing

Non-Linear
Declining
Market share
Market share

Sales

Sales
Time Time

2
 ➢ The trend component does not have to be linear.
• Cyclical component: A Cyclic Cyclic variation of sales

component is a repeated but non-periodic

Sales
fluctuation.
• Seasonal component: At time t, Time

reflecting seasonality (seasonal variation).

A seasonal pattern exists when a time series is influenced
by seasonal factors. Seasonality is
Seasonal variation of sales
always of a fixed and known period
(e.g., the quarter of the year, the

Sales
month, or day of the week).
• We shall next discuss about the trend Time

component is particular

3
Trend Projection
• The time-series forecasting method fits a trend line to a
series of historical data points and then projects the line
into the future for medium-to long range forecasts.
• The series may contain (a) Trend Component, (b)
Seasonal component; (c) Cyclic component and; (d)
Random component.
• We have also discussed about the Exponential Smoothing
where weights are chosen to predict a Forecast.
• We shall discuss an extension of the Exponential
Smoothing method which is known as Holt’s method.

4
Holt’s Method:
• In the Holt’s model, we fit a
Forecast
straight line between the points of Ft
sales data at the previous time
bt-1
period (𝑡 − 1), and the forecast at 𝑡:
𝐹𝑡 = 𝑎𝑡−1 + 𝑏𝑡−1 at-1 time
• 𝑎𝑡−1 , is called the level, which t-1 t
represents the smoothed value up
to 𝑡 and including the last data;
• And 𝑏𝑡−1 , is the trend on the slope of the straight line that,
we are fitting at the point 𝑡.
• So, 𝑎𝑡−1 is the representative value of the level or the
constant and we add a trend to it.

5
• So that the forecast of the next period will be a level at the
end of the previous period and then we add a trend to it.
• Now, both 𝑎𝑡 and 𝑏𝑡 are updated for every data point
using exponential smoothing, so 𝑎𝑡 which is the level
value up to and including the last
point is given by: Dt
at
𝑎𝑡 = 𝛼𝐷𝑡 + (1 − 𝛼)𝐹𝑡 bt
Ft
𝑎𝑡 = 𝛼𝐷𝑡 + (1 − 𝛼 )(𝑎𝑡−1 + 𝑏𝑡−1 )
𝑏𝑡 = 𝛽(𝑎𝑡 − 𝑎𝑡−1 ) + (1 − 𝛽)𝑏𝑡−1 bt-1

Where
at-1 time
𝐹𝑡 = 𝑎𝑡−1 + 𝑏𝑡−1 t-1 t

6
• In fact, this is our very familiar exponential smoothing
equation where 𝛼 is the smoothing constant.
• The normal exponential smoothing equation has two
parts, one is the demand part and
the other is the forecast part 𝐹𝑡+1 =
𝛼𝑥𝑡 + (1 − 𝛼)𝐹𝑡 .
• Here, 𝐷𝑡 is the demand during the
previous period.
• And, 𝛼𝐷𝑡 represents the demand with a
weight added to it.
• In the exponential smoothing (1 − 𝛼 )𝐹𝑡 is
used to account for the previous forecast.

7
• Here, the last term (𝑎𝑡−1 + 𝑏𝑡−1 ) represents, the forecast
for period 𝑡, just as forecast for period 𝑡 + 1 is represented
by (𝑎𝑡 + 𝑏𝑡 ).
• So, here the equation is like 𝛼𝐷𝑡 + (1 − 𝛼)𝐹𝑡
• Now, 𝑏𝑡 , which is the trend, which also has two
components, one is in some sense the component that
comes out of the level and the other is the component that
comes out of the trend.
• So, 𝛽 is another exponential
smoothing constant like 𝛼, and
𝛽(𝑎𝑡 − 𝑎𝑡−1 ), is another value of level.

8
• If we take the present level value 𝑎𝑡 and take the previous
level value 𝑎𝑡−1 , the difference between them in some
sense represents a trend as the slope.
• So, that is one component of 𝑏𝑡 , and the other
component comes from the actual slope that
was computed.
• So, the 𝑏𝑡 also has two components, that are
related through exponential smoothing just like 𝑎𝑡 which
has two components that are related, through exponential
smoothing.
• In the Holt’s model, we are redefining the slope at every
point, up-to and including the last point.

9
• Similarly, in the Holts’ model for every point 𝑡, we are
actually trying to find out a level and a slope and we add
them, so that we get the forecast for the next period then
we know that forecast for the next period and then we
know that demand point and keep iterating this
till, we get to the final answer.
• Let us now consider an example to understand
the method.
• The forecast of the near future is then
calculated from:
𝐹𝑖+ℎ = 𝑎𝑖 + ℎ × 𝑏𝑖
Where ℎ stands for the time period.

10
Example
• Let us consider a case where the data is given as:

Time, months Sale data

t D
1 35
2 52
3 95
4 210
5 124
6 176
7 229
8 240
9 525
10 415
11
 11 575
12 760
13 620
14 760
15 870

If we plot the data it shows a trend

as shown:

12
• The analysis of the data is done as shown
below:
• To start the Holt’s method of prediction we
assume that the 𝑎1 = 𝐷1 = 35 and 𝑏1 = 0.
• Assume the values of coefficients as: 𝛼 =
0.4; 𝑎𝑛𝑑 𝛽 = 0.6
• This gives us the values of 𝑎2 𝑎𝑛𝑑 𝑏2 by the Holt’s method
as, for 𝐷2 = 52;
𝐹𝑡 = 𝑎𝑡−1 + 𝑏𝑡−1
𝑎𝑡 = 𝛼𝐷𝑡 + (1 − 𝛼 )(𝑎𝑡−1 + 𝑏𝑡−1 )
𝑎2 = 𝛼𝐷2 + (1 − 𝛼)(𝑎1 + 𝑏1 )
= 0.4 × 52 + (1 − 0.4) × (35 + 0) = 41.8
𝑏𝑡 = 𝛽(𝑎𝑡 − 𝑎𝑡−1 ) + (1 − 𝛽)𝑏𝑡−1
13
𝑏2 = 𝛽 (𝑎2 − 𝑎1 ) + (1 − 𝛽) × 𝑏1
= 0.6 × (41.8 − 35) + (1 − 0.6) × 0 = 4.08
𝐹3 = 𝑎2 + 𝑏2 = 41.8 + 4.08 = 45.88
𝐷3 = 95
𝑎3 = 𝛼𝐷3 + (1 − 𝛼 )(𝑎2 + 𝑏2 )
= 0.4 × 95 + (1 − 0.4) × (41.8 + 4.08)
= 65.528
𝑏3 = 𝛽 (𝑎3 − 𝑎2 ) + (1 − 𝛽) × 𝑏2
= 0.6 × (65.528 − 41.8) + (1 − 0.6) × 4.08 = 15.8688
𝐹𝑡 = 𝑎𝑡−1 + 𝑏𝑡−1
Hence
𝐹4 = 𝑎3 + 𝑏3 = 65.528 + 15.8688 = 81.3968

14
In the same manner we calculate the values of
𝑎𝑖 , 𝑏𝑖 𝑎𝑛𝑑 𝐹𝑖 . The table of these values are shown below:

α 0.4
β 0.6
t D at bt Forecast Ft |e|=|D-Ft| e^2
1 35 35 0
2 52 41.8 4.08 35 17 289
3 95 65.528 15.8688 45.88 49.12 2412.774
4 210 132.8381 46.73357 81.3968 128.6032 16538.78
5 124 157.343 33.39637 179.5716 55.57165 3088.208
6 176 184.8436 29.85893 190.7394 14.73936 217.2488
7 229 220.4215 33.29032 214.7025 14.29746 204.4173
8 240 248.2271 29.99947 253.7118 13.71184 188.0146
9 525 376.9359 89.22509 278.2266 246.7734 60897.12
10 415 445.6966 76.94644 466.161 51.16104 2617.452
11 575 543.5858 89.51211 522.6431 52.35693 2741.248
12 760 683.8588 119.9686 633.098 126.902 16104.13
15
13 620 730.2964 75.85003 803.8274 183.8274 33792.5
14 760 787.6879 64.77488 806.1465 46.14645 2129.495
15 870 859.4777 68.98382 852.4628 17.53725 307.555
16 928.4615
17 997.4453
18 1066.429
19 1135.413
20 1204.397

The values are plotted in the graph shown below.

From the data, level and trend values the forecast of the
future is given as,
𝐹𝑖+ℎ = 𝑎𝑖 + ℎ × 𝑏𝑖
𝐹15+2 = 𝑎15 + 2 × 𝑏15 = 859.4777 + 2 × 68.98382 = 997.4453
𝐹15+3 = 𝑎15 + 3 × 𝑏15 = 859.4777 + 3 × 68.98382 = 1066.429

16
 𝐹15+4 = 𝑎15 + 4 × 𝑏15 = 859.4777 + 4 × 68.98382 = 1135.413
𝐹15+5 = 𝑎15 + 5 × 𝑏15 = 859.4777 + 5 × 68.98382 = 1204.3968
1400
Sale Data and Forecast

• The

Sale data and Forecast, numbers

1200

forecasted 1000

values are 800

also shown in 600

Data

the table. 400 Predicted

• The forecast is
Forecast
200

calculated as, 0
0 2 4 6 8 10 12 14 16 18 20

𝐹𝑖+ℎ = 𝑎𝑖 + ℎ × 𝑏𝑖 Time, months

• Hence the forecasts are:

17
𝐹15+2 = 𝑎15 + (𝑡17 − 𝑡15 )𝑏15
= 859.4777 + (17 − 15) × 68.9838 = 997.4453
𝐹15+3 = 𝑎15 + (𝑡18 − 𝑡15 )𝑏15
= 859.4777 + (18 − 15) × 68.9838 = 1066.4291
𝐹15+4 = 𝑎15 + (𝑡19 − 𝑡15 )𝑏15
= 859.4777 + (19 − 15) × 68.9838 = 1135.4129
𝐹15+5 = 𝑎15 + (𝑡20 − 𝑡15 )𝑏15
= 859.4777 + (20 − 15) × 68.9838 = 1204.3967

18
 Forecasting – Seasonality by Winter Holt’s Method:
• In the previous lecture we discussed about the Trend
Projections in Time Forecasting Techniques
Series using Holt’s
method. New Products Existing Products
Qualitative
methods
Causal
methods

• In this lecture we Bass

Diffusion Time series
Regression
shall discuss Method for
New
methods Grass
roots
Analysis
Products
seasonal models. Naïve Market

• Let us consider a
Method research

case where the data Moving

Average
Panel
Consensus

is divided into 4 Exponential Historical

quarters and we are Smoothing Analogy

considering the Trend

Projections
Delphi
Method

Different Forecasting Methods

19
 pattern of changes during 3 years.
SEASONAL
70
CHANGES IN DATA
60 Seasonal Changes
Year 50
Quarter 1 2 3 40

Data
1 53 58 62
30
2 22 25 27 Year-1
20
3 37 40 44 Year-2

4 45 50 56 10 Year-3

Total-> 157 173 189 0

0 1 2 3 4 5
Quarter

• From the data given in the table we note that, there is a

trend, in the sense that, for all the years the total sale
during the year also seems to go up.

20
• However, for each quarter of the year the data shows a
kind of variation which goes through peaks and troughs.
• This variation during the quarters is called seasonality as
the nature of variation is during a Charles C. Holt (21 May 1921 – 13
December 2010) was Professor at
season. the Department of Management at
the McCombs School of Business at

• Products such as air conditioners, the University of Texas at Austin. He

is well known for his contributions

coolers, room-heaters, ice creams, cool- (and for the contributions of his
student, Peter Winters) to
exponential smoothing.[1]
drinks etc. show such variations as the Holt holds BS and MS degrees from
market requirement varies with the the Massachusetts Institute of
Technology (1944). He later earned a
MA (1950) and a PhD (1955) from
season. the University of Chicago.

• In our Forecasting for such data we

have to capture the seasonality.
• Hence we will build a simple model for predicting
seasonality.
21
 • We will develop model to predict seasonality that uses
ideas from exponential smoothing.
Holt-Winter’s Model for Forecasting Seasonal changes
• In this method we shall use Exponential Smoothing in
some form to get a model which can, in principle,
forecast better, in the sense that, a recent data will have
more weightage.
• For example, when we use the method of moving
average to fit a line, we give equal weightage to all data
points.
• We have explored if the Exponential Smoothing can be
used to build another model where the most recent data
has more weightage.

22
 • We shall now introduce another model which attempts to
do similar treatment to the sale data and that is named as
Holt-Winter’s model.
• The Holt-Winter’s model gives adequate D
a
t
t

representation to the seasonality in data F t

b t

such that it may be used to forecast along b t-1

with the seasonality.

at-1 time
• The model is stated mathematically as t-1 t

follows.
𝑊𝑖𝑛𝑡𝑒𝑟 ′ 𝑠 𝑀𝑜𝑑𝑒𝑙 𝑓𝑜𝑟 𝑎 𝑝𝑒𝑟𝑖𝑜𝑑 𝑝:
𝐷𝑡
𝐿𝑒𝑣𝑒𝑙: 𝑎𝑡 = 𝛼 ( ) + (1 − 𝛼 )(𝑎𝑡−1 + 𝑏𝑡−1 )
𝐶𝑡−𝑝
𝑇𝑟𝑒𝑛𝑑 (𝑠𝑙𝑜𝑝𝑒): 𝑏𝑡 = 𝛽(𝑎𝑡 − 𝑎𝑡−1 ) + (1 − 𝛽)𝑏𝑡−1

23
 𝐷𝑡
𝑆𝑒𝑎𝑠𝑜𝑛𝑎𝑙: 𝐶𝑡 = 𝛾 ( ) + (1 − 𝛾)𝐶𝑡−𝑝
𝑎𝑡
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛: 𝐹𝑡 = (𝑎𝑡−1 + 𝑏𝑡−1 )𝐶𝑡−𝑝
𝑆𝑒𝑎𝑠𝑜𝑛𝑎𝑙 𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡: 𝐹𝑡+ℎ = (𝑎𝑡 + ℎ𝑏𝑡 )𝐶𝑡+ℎ−𝑝
𝑃𝑒𝑟𝑖𝑜𝑑: 𝑝
• The Winter’s model is specified by the equations stated
above.

24
• In the equation 𝑎𝑡 represents the level or the base of the
data.
• Here 𝑏𝑡 represents the trend or the
projection of the slope and 𝑐𝑡 represents
the seasonality.
• There are now three parameters in the
Winter’s model, a level a trend and
seasonality.
• We shall deliberate on how to extract
the three parameters 𝑎, 𝑏 & 𝑐 used in the above example
where the yearly sum of data is given by 157, 173, 189.
• The level is governed by the first value 157.
• It is also evident that the trend is an increasing type.

25
• And there is a slope which increases the level from 157
to 173 in the second year and then to 189 in the third
year.
• Following the same lines as we have done in the case of
Exponential Smoothing and Holt Model, the parameters
are calculated here.
• Thus the level for the next period is calculated as 𝛼 times
the data for the recent period and (1 − 𝛼) times the
forecasted data as it has been done in the Holt’s method.
• Thus the predicted value of the level becomes:
𝐷𝑡
𝐿𝑒𝑣𝑒𝑙: 𝑎𝑡 = 𝛼 ( ) + (1 − 𝛼 )(𝑎𝑡−1 + 𝑏𝑡−1 )
𝐶𝑡−𝑝

26
 • In the above equation 𝐷𝑡 is the data for the period and
𝐶𝑡−𝑝 is the seasonality coefficient which reflects the value
one period prior the present data 𝐷𝑡 .
• The trend of the slope of the next period is also
calculated in similar manner.
• Let us consider an example and compute with Holt
Winter’s model:
Example:
• Consider the data given in the table below which contains
a seasonal variation and following the Holt Winter’s
method the Forecast is calculated as shown:
Consider the data in the table given below:

27
 Time, quarter Sale Data
1 115
2 130
3 70
4 260
5 170
6 225
7 120
8 380
9 140
10 280
11 180
12 460
13 250
14 410
15 220
16 690
The scatter plot of the data is shown in the graph below:
28
 800
Sale Data
700

600

Data, numbers
500

400

300

200

100

0
0 2 4 6 8 10 12 14 16 18
Time, quarter

To begin the Holt-Winter’s forecasting start calculation of the

seasonal coefficient.
The period in the above data is 𝑝 = 4 as the data is
presented for every quarter. Assume the values of 𝛼, 𝛽 𝑎𝑛𝑑 𝛾
as

29
 𝛼 = 0.2; 𝛽 = 0.4; 𝛾 = 0.6;
Take the starting value of seasonal coefficient for quarter-1
as 𝐶𝑡 :
𝐷1 115 115
𝐶1 = = =
𝐴𝑣𝑒𝑟𝑎𝑔𝑒(𝐷1 𝑡𝑜 𝐷4 ) 𝐴𝑣(115,130,70,260) 143.75
= 0.8
Then calculate the seasonal coefficient for quarter-2,3&4 as:
𝐷2 130 130
𝐶2 = = =
𝐴𝑣𝑒𝑟𝑎𝑔𝑒(𝐷1 𝑡𝑜 𝐷4 ) 𝐴𝑣(115,130,70,260) 143.75
= 0.9043
𝐷3 70 70
𝐶3 = = =
𝐴𝑣𝑒𝑟𝑎𝑔𝑒(𝐷1 𝑡𝑜 𝐷4 ) 𝐴𝑣(115,130,70,260) 143.75
= 0.487

30
 𝐷4 230 260
𝐶4 = = =
𝐴𝑣𝑒𝑟𝑎𝑔𝑒(𝐷1 𝑡𝑜 𝐷4 ) 𝐴𝑣(115,130,70,260) 143.75
= 1.8087
For the 4th quarter also calculate the coefficients 𝑎4 𝑎𝑛𝑑 𝑏4 .
The expressions for 𝑎𝑡 𝑎𝑛𝑑 𝑏𝑡 are:
𝐷𝑡
𝐿𝑒𝑣𝑒𝑙: 𝑎𝑡 = 𝛼 ( ) + (1 − 𝛼 )(𝑎𝑡−1 + 𝑏𝑡−1 )
𝐶𝑡−𝑝
𝑇𝑟𝑒𝑛𝑑 (𝑠𝑙𝑜𝑝𝑒): 𝑏𝑡 = 𝛽(𝑎𝑡 − 𝑎𝑡−1 ) + (1 − 𝛽)𝑏𝑡−1
Since the trend value 𝑏𝑡 may be calculated only when two
values of levels are known so assume 𝑏4 = 0, however,
estimate 𝑎4 as:
𝐷4 260
𝑎4 = ( ) = = 143.75
𝐶4 1.8087

31
Now the table is ready with the starting estimates so
calculate the subsequent values by the expressions given
above:
𝐷𝑡
𝐿𝑒𝑣𝑒𝑙: 𝑎𝑡 = 𝛼 ( ) + (1 − 𝛼 )(𝑎𝑡−1 + 𝑏𝑡−1 )
𝐶𝑡−𝑝
𝑇𝑟𝑒𝑛𝑑 (𝑠𝑙𝑜𝑝𝑒): 𝑏𝑡 = 𝛽(𝑎𝑡 − 𝑎𝑡−1 ) + (1 − 𝛽)𝑏𝑡−1
𝐷𝑡
𝑆𝑒𝑎𝑠𝑜𝑛𝑎𝑙: 𝐶𝑡 = 𝛾 ( ) + (1 − 𝛾)𝐶𝑡−𝑝
𝑎𝑡
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛: 𝐹𝑡 = (𝑎𝑡−1 + 𝑏𝑡−1 )𝐶𝑡−𝑝
Thus,

32
 𝐷5
𝑎5 = 𝛼 × ( ) + (1 − 𝛼 )(𝑎4 + 𝑏4 )
𝐶5−4
170
= 0.2 × + (1 − 0.2) × (143.75 + 0) = 157.5
0.8
𝑏5 = 𝛽(𝑎5 − 𝑎4 ) + (1 − 𝛽)𝑏4
= 0.4 × (157.5 − 143.75) + (1 − 0.4) × 0 = 5.5
𝐷5 170
𝐶5 = 𝛾 ( ) + (1 − 𝛾)𝐶5−4 = 0.6 × + (1 − 0.6) × 0.8
𝑎5 157.5
= 0.9676
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛: 𝐹𝑡 = (𝑎𝑡−1 + 𝑏𝑡−1 )𝐶𝑡−𝑝
𝐹5 = (𝑎4 + 𝑏4 )𝐶5−4 = (143.75 + 0) × 0.8 = 115
In the same manner the subsequent values are calculated
and the complete table is generated as given below:

33
Holt-Winters analysis
Parameters
α β γ
0.2 0.4 0.6

Time Sale Data Level Trend Seasonal Forecast Abs. error Sqr. error
quarter numbers at bt Ct Ft |e| e^2
1 115 0.8
2 130 0.904348
3 70 0.486957
4 260 143.75 0 1.808696
5 170 157.5 5.5 0.967619 115 55 3025
6 225 180.1596 12.36385 1.111075 147.4087 77.5913 6020.411
7 120 203.3045 16.67625 0.548931 93.75056 26.24944 689.0334
8 380 218.0038 15.88549 1.769331 397.8782 17.87821 319.6302
9 140 216.0485 8.749147 0.775849 226.3158 86.31575 7450.409
10 280 230.2397 10.926 1.174104 249.7669 30.23308 914.0392
11 180 258.5146 17.86554 0.637344 132.3834 47.6166 2267.341
12 460 273.1011 16.55394 1.718347 489.008 29.00803 841.466
13 250 296.1696 19.15974 0.816806 224.7287 25.27132 638.6396
14 410 322.1039 21.86959 1.23337 370.2294 39.77058 1581.699
15 220 344.2153 21.96631 0.638418 219.2294 0.770575 0.593787

34
16 690 373.255 24.79567 1.7965 629.2272 60.77279 3693.332
17 325.1303
18 521.5262 MAD MSE
19 285.783 41.37314 2286.799
20 848.7344

The absolute values of error is calculated as the absolute

difference between data and prediction, as,
|𝑒𝑡 | = 𝑎𝑏𝑠(𝐷𝑡 − 𝐹𝑡 ) = 𝑎𝑏𝑠(𝐷5 − 𝐹5 ) = 𝑎𝑏𝑠(170 − 115) = 55
Thus the mean value of absolute errors is generated as,
𝑀𝐴𝐷 = 𝑎𝑣𝑒𝑟𝑎𝑔𝑒(|𝑒5 |, |𝑒6 |, −−, |𝑒16 |)
The best way to estimate the parameters 𝛼, 𝛽 𝑎𝑛𝑑 𝛾 is use a
solver to calculate the values for minimum value of MAD.
However in this calculation the values have been manually
generated for the minimum value of MAD.

35
The value of forecast from the 17th quarter to the 20th quarter
is generated as below:
The expression used is:
𝑆𝑒𝑎𝑠𝑜𝑛𝑎𝑙 𝐹𝑜𝑟𝑒𝑐𝑎𝑠𝑡: 𝐹𝑡+ℎ = (𝑎𝑡 + ℎ𝑏𝑡 )𝐶𝑡+ℎ−𝑝
Thus
𝐹16+1 = [𝑎16 + (𝑡16+1 − 𝑡16 )𝑏16 ]𝐶16+1−4
= [373.255 + (17 − 16) × 24.7957] × 0.8168 = 325
𝐹16+2 = [𝑎16 + (𝑡16+2 − 𝑡16 )𝑏16 ]𝐶16+2−4
= [373.255 + (18 − 16) × 24.7957] × 1.2334 = 521.53
𝐹19 = [373.255 + (19 − 16) × 24.7957] × 0.6384 = 285.77
𝐹20 = [373.255 + (20 − 16) × 24.7957] × 1.7965 = 848.73

36
Thus the values of sale data, predicted values and the
forecast for the next year are calculated and are shown in
the graph given below.

37
 900
Forecast of Seasonal Changes using Holt-Winter method
800

700

600
Data, number

500

400

300

200
Data

Predicted
100
Forecast

0
0 2 4 6 8 10 12 14 16 18 20
Quarter of an year, number

38
Conclusion
• The Holt-Winter’s method of forecasting is useful for
planning purpose using historical data;
• The smoothing constants allow the results to forecast
keeping in view the features for the type of data to be
analysed.
• The Holt-Winter’s model can be used as a part of the
Enterprise Resource Planning (ERP) as it takes care of
the changing buying behaviours, market completion and
other factors.
• Helps planning advertisement in marketing, repeat
telecasts, game show etc.
• The power generating companies use the method
extensively for their ERP.
39
 Forecast of Electricity Consumption by Holt-Winter's Method
20000

18000
Electricity Consumption in GWhr

16000

14000

12000

10000

8000

6000

4000

2000

0
1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998
Time, year

40
•

41