n, s\".Ae..,!

Dt. M. Ahsan Akhtar Hasin, IPE, BUET.

Chapter 13
Forecasting
Dr. Md. Altsan Akhtar Hasin
professor,
Industrial and Production Engineering, BUET.

Introduction
Forecast means prediction o. estimation about an)'thing in future. In operations management.
forecast is necessary for anticipating demand of products or services. It is in fact the first
level planning, since all subsequent plans depend on this. Some level of inaccuracy may be
present in forccast. However, it is better to minimize this inaccuracy as much as possible.

Bear in mind that a perfect forecast is usually impossible, because too many factors in the
business environment can not be known in advance and predicted with certainty.

Forecast is vital, because it is the basis for long-tem planning. It is required for every
department in the organization. This demand estimation acts as the input for other functions
or departments, as explajned below:

Marketing/sales need to forccast demand of products in the market, because

appropriate sales arangement must be in place beforehand.
. Distribution needs to know the amount to stock to be kept to fulfill demand in the next
pe ods. Because all logistical arrangements, including transportation, must be in
place on time.
. Production planner also needs this forecast demand for production planning putpose,
which will determine the target production volume in a period. All production
resources must be planned accordingly.
. Finance also needs to know forecast values for budgetary allocation.
. Inventory relies on production plan, based on forecast, to stock amount ofmaterials.
. Procurement department needs to kflow forecast, because amount and timings of
procurement ofmaterials depend on that.

Thus, forecast is vital for many depa.tments.

Components ofDemand / Demand Patterns

In most cases, demand for products and services can be broken down into five components,
or can follow five patternsi

1. Trend Trend lines are the usual starting point in developing a forecast. These tend
lines are then adjusted fo, seasonal effects, cyclical, and any other expected events
that may influence the final forecast. Figure 13.1 shows four of the most common
t)?es of trends. Trends are such a pattem that increase or decrease with time.
Demands ofmost ofthe products or services in the world have some pattem oftrend.
For instance, demands oftelevisions, mobile phones, camera, apparels, chairs/tables,

-8-
Dr. M. Ahsan Akhtar Hasin, IPE, BUET.

computers, etc. have trend. However, increase or decrease may take place either
linearly (in a straight line), or non-linearly (in a curve).

a) Linear trend A linear trend is obviously a straight continuous relationship. lf

demand increases, or decreases unifomly with time, then it has Linear trend,
either Linear Increasing trend, or Linear decreasing trend. A linear trend is
obviously a straight continuous relationship. Linear increasing trend means
that demand ofthe product increases at a constant rate with time. For example.
demands ofchair, television, etc. increase Iinearly with time.

On the other hand, linear decreasing trend means that demand ofthe product
decreases at a constant rate with time. For example, demand ofwrist watch is
decreasing Iinearly with time.

b) Nonlinear trend If demafld changes at a constant percentage, then it looks

Iike a curve, either Asymptotic or Exponential. ln case ofAsymptotic change,
initially after introduction ofthe product, demand increases rapidly, then slows
down fbr any reason. For instance, demand ofTeletalk mobile phone.

On the otherhand, in case ofExponential change, initially after introduction of

the product, demand increases very slowly, and then grows very fast after
some time for any reason. For instance, demand ofmobile phone in general.

2. Seasonal - Demand ofthis trpe ofproduct goes up and down depending upon seasons
ofthe year. For instance, demand ofan umbiella goes up in the rainy season, whereas
its demand goes deep down in the winter. Demand of cold drinks goes up in the
summer, but goes down in the winter. Although demand changes twice a year, its
pattem changes four times a year.

3. Cyclical - Graphically, Cyclical pattem looks like seasonal, although Cyclical

demand pattem goes up or down once after several years. Cyclical factors are more
difficult to determine, because the time span may be unknown. Cyclical influence on
demand may come fiom such occurrences, as political elections, \,,,ar, economic
conditions such as recession, sociological prcssures, or sociological evolution such as
change in fashion. A very few products belong to this category. Fashion products
follow this pattem.

Random Random variations are caused by chance events. Statistically. it is thc

unknown part ofdemand, which is beyond eKplanation. In facr. rhis is the unerplain.^d
portion of demand. ln this case, demand goes up and down frequently in small
margin. All most all t]?es ofproducts have this pattem in their demand; e.g. grocery
items, foods, etc.

5. Combined In reality, every product follows a combination of many pafterns

It is rather uncommon to find a product which follows a single pattem of
together.
demand change. For instance, cold drinks have lncreasing Trend, Seasonal pattem, as
well as Random pattem (random change).

-9-
 ar Hasin,IPE, BUET.

2) As! mplotic trund/Exponenti al tend

Increasing

EI
EI
Decaeasing ol
Time

3) Seasonal lrehd (Qualtelly change) 4) Cyclic compoaenl pattem

EI
EI
ol
-.--------------
Time (Quarters) Time (Years)

5) Randoht pdtte ,

_t
EI
al

Time

6) Cottbi ed (Combination ofmany pattems)

Time ----------------

Figure 13.1. Components ofdemand, or demand patterns.

-10-
Dr. M. Ahsail Akhtar Hasin, IPE, BUET.

Tl pes ofForecasting
Two basic types:
1. Quatitative
2. Quantitative
a Qualitative

These are Subjective, judgmental techniques, based on estimations or opinions, derived from
experience, especially when past data ar€ not available.

a) Grass root - Dtiyes a forecast by compiling inputs from the persons at the end ofthe
hjerarchy. For example, an overall sales forecast may be derived by combining inputs
from each salesperson, who works closest to the market or consumer. Forecasts at the
bottom level are summed and given to the next higher level. This is usually a district
warehouse, which then adds in safety stocks and any effects of ordering quantity
sizes. This information is then fed to the next higher level, such as rcgional
\nrarehouse-

b) Market Research It is done by conducting Survey or Interview, using questionnaire.

The collected data are analyzed statistically, and then a decision is taken. Firms often
hire outside companies that specialize in market rcsearch. Market research is used
mostly for product research in the sense of looking for new product ideas; iikes and
dislikes about existing products, consumer r€quirements, etc.

c) Historical analogy - An analogy (comparison) is prepared from past historical

experience, leamt by selling a similar product in the past, or in a similar market.

d) Panel consensus - [t is based on the idea that two heads are better than one. The basic
idea is that a panel ofpeople from a variety ofpositions can develop a more reiiable
forecast than a narrower group. Thus, a committee of expefis is formed, comp sing
top executives having direct relevance with demand. Free open exchange of views
take place at face-to-face meetings.

When decisions in forecasting are at a broader and higher level (as when jnhoducing a
new product line, or conceming strategic product decisions such as new marketing
areas), the term "Jury of Executive Opinion" or "Executive Judgment', is generally
used.

e) Delphi method - This is partly similar to Panel consensus. However, meetings are
avoided. Group of experts responds to questionnaire. It can avoid influence of top
executives.

The idea behind Delphi method is a statement or opinion at a higher level person
will likely be weighted more than that of a lower-level pe$on. The worst is when
lower level people feel threateried and do not contribute their true beliefs. In this
method, everyone has the same weight. A moderator takes the central role of
collecting, analyzing the datrinformation from the group members.

-11-
Dr. M. Ahsan Akhtar Hasi , IPE, BUET.

a QuaDlitalive

These are based on analysis ofpast records or data or mathematical methods. A person has no
scope to express subjective judgment. While recognjzing that tems such as short, medium,
and long arc relative to the context! or situation, in uhich they are used, in business
forccaslirg short term usually refers to less than 3 months, mediutl-telm 3 months to 2 years,
and. long-term gteatet than 2 years.

These may be ofmany types again i

a) Time series arollsis - These techniques are based on the idea that past records will
continue in future, or past history will be applicable in future. Time series forecasting
models try to predict the future based on past data. There are different techniques:

i. Simple Moving Average (MA)

ii. Weighted Moving Average
iii. Exponential Smoothing
iv. Regression Analysis
v. Etc.

Apart from the above time series quantitative methods, there are some other time series
analysis techniques, such as Holt's method and Winter's method.

Holt's two-parameter model, also known as linear exponential smoothing, is a popular

smoothing model for forecasting data with trend variation, but no seasonality variation.
Holt's model has three separate equations that work together to generate a final forecast.
The method is also called double exponential smoothing ot te17d-enhanced exponential
smoothing.

Winter's method is applicable with data set having both trend and seasonality.

Holt-Winter's combined exponential smoothing method covers three aspects of demand

variation - an average demand, seasonal variation and tend. This means that based on
average demand, two types of additive variations, namely trend variation ard seasonal
variations, can be analyzed together. That's why this is also called ,rple exponential
smoothing nethod. However, this is often considered as highly contusing.

b) Causal relaliohships These techniques recognize that forecast or demand is not ooly a
variable of "time", but also many other factors. Thus, these try to understand how
surrounding elements affect forecastjng, e.g. Sales may be affected by advertising; sales
promotion or discount may suddenly influence demand, etc. There are many such
methods, e.g.
i. Econometric models.
ii. Input-Output model, etc.

c) Simulation - These are based on Computer-based modeling (using computer

programming) and analysis on huge amount of past data, which can simulate the
behavior of data (paftem of demand). Then, based on that behavior, leamt through
simulation, the program can estimate future demand.

-12-
Dr. M. AhsanAkhtar Hasin, IPE, BUET.

How to choose a method ?

Which forecasting method/model should be used, depends on some characteristics or

attributes. Those are:

. Time horizon to forecast - Length of period of estimation, such as long{erm, shott-

term, medium, etc.
Data availability - How many periods! data is available on-hand for analysis.
Accuracy requircd How much accurate we want to be in our futule prediction, or
forecast/output. Generally, the higher the accuracy required, the more complex a
technique should be selected.
Demand chamcteristics The pattern ofdemand, such as increasing trend, asymptotic
trend, seasonal pattem. etc. largely influence selection ofa method.
Size offorecasting budget, and availability ofexpertise for analysis.

O Simple Moving Ayerage (MA)

When demand for a product is neither growing, nor declining rapidly, and if it does not have
seasonal effect, then this method is applicable. This assumes that recent demand data are
more usefirl in predicting future demand. If there is a trend in the data ether increasing or
decreasing the moving average has the adverse characteristic oflagging the trend.

Here, an average of3 periods is selected for calculating average demandl

Weeks Demand/actual 3 week MA

sales forecast
1200
2 1400 ---1
3 1000
4 1500 L+ 1200
5 r 500 1300
6 1300 1333
7 t600 1433
8 t700 1467
9 1500 1533

O ExponentialSmoothing
The major drawback ofthe previous method is the need to continually carry a large amount
of historical data. If limited amount of data is available. and the most recent data should be
given more importance, then Exponential Smoothing is more suitable. In many applications,
(perhaps in most), the most recent occurrences are more indicative ofthe future than those in
the more distant past. Tfthis premise is valid, then Exponential Smoothing is the most logical
solution. That's why this is the most used ofall forccasting techniques. For example, it is
widely used in ordering inventory in retail firms, wholesale companies, and service agencies.

-t 3-
Dr. M. Ahsan Akhtar Hasin, IPE, DUET.

Here. F. 0A- - (l-o)l .

where, Fr = Forecast fo. the next period (week, month, quarter. year, etc.),
Ft I = Forecast for the previous period,
At.r = Actual demand/sales for the previous period
(t: Smoothing constant (0-1)

The Smoothing constant o detemines the level of smoothing and the speed of reaction to
ditTercnces between forecasts and actual occurrences. Smoothing constant is basically a
"weightage" or "importance!' assigned to actual demand alld forecast. The value of o can be
lound through computer simulation using past pattem analysis! or expert opinion.

This method is surprisingly good, especially ifno rapid changes in demand take place.

The value for the constant is determined both by the nature of the product and by the
managerrs sense of what constitute a good response rate. For example, if a fim produced a
standard item with relatively stable demand the reaction rate to differences between actual
and forecast demand would tend to be small, perhaps just 5 or l0 percentage points.
However, if the firm were experiencing groMh, it would be desirable to have a higher
reaction rate, perhaps 15 to 3opercentage points, to give greater impofiance to recent growth
experience (or slump experience).The more rapid the grox,th rate, the higher the reaction rate
should be. A more specific discussion is given below under the topic ,,smoothing
Coeffi cient Selection".

ExamDlg: Last month's forecast and actual sales were 1050 units and 1000 units respectively.
Given, o = 0.05 (past experience). What is the fo.ecast amount for this morth ?

Ft : o Ar I + (l- c.) Fr-r : 0.05 x 1000 + (1-0.05) 1050 : 1048 uoits

Smoothing Coefficient Selection

[Ref: Everett E. Adam and Ronald J. Ebert, 5th Editjon Prentice Hall, India. p-95].

In Exponential Smoothing, there is a problem with parameter selection, that is, we must fit
the forecasting model to the data. To begin forecasting, some reasonable estimate of a
"beginning forecast" is necessary. Likewise, a shtoothing coellicient d, must be selected.
This choice or selection is critical. If you notice, the basic equation of (first level)
Exponential Smoothing, a high o places heavy weight on the most recent actual demand; a
low o weights recent actual demand less heavily. A high smoothing coefficient could be more
approp ate for new products or items for which the underlying demand is fluctuating or
unstable or d1,namic. An o of0.7, 0.8 or 0.9 might be best for these conditions, although one
may question the use of exponential smoothing method at all if unstable conditions are
known to exist. If demand is very stable and believed to be representative ofthe future, the
forccaster wants to select a low o value to smooth out any sudden noise (variations due to
market disturbances) that might have occurred. The forecasting procedure, then, does not
overreact to the most recent (actual) demand. Under these stable conditions, an appropriate
smoothing coelficietrt o might be 0.1,0.2, or 0.33. When demand is slightly unstable,
smoothing coellicient of0.4, 0.5, or 0.6 might provide the most accurate forecasts.

-14-
Dr. M Ahsan Akhtar Hasin, IPE, BUET-

O Adaptive Exponential Smoothing

[Ref: Everett E. Adam and Ronald J. Ebert. 5th Edition Prentice Ha]1, India, p-971.
and [Chase, Aquilano and Jacobs, 8th Edition, Irwin McGraw Hill, USA, p-510].

lfthe forecaster or the manager is unsure about the stability, or form ofthe demand pattern,
Adaptire Exponential Smoothing methods provide good forecasting alternative. In adaptive
exponential smoothing, the smoothing coefficient o, is not fixed. An initial value is set, and
then allowed to fluctuate over time, based on changes in the demand pattern.

The changes in the demand pattem is due to trend and seasonality on the basic avemge
demand. These two components of demand (i.e. presence of taend and seasonality on the
average demand) can be incorporated in Doable E ponenlial Smoothi g method, and Tiple
Exponentitl Stuoothing method. ln these two cases, Exponential Smoothing is done twice
(i.e. 2-level smoothing) and thrice (3level smoothing) respectively. There are slight
variations in representatjons ofsuch two models outlined by Brown (Brown,s method), Holt
(Holt's method) and (Holt's student) Winter (Holt-Winter) merhod.

The Exponential Smoothing that we learnt earlier is known as the basic or lstJevel
smoothing.

O Regression Analysis

This is also known as "Linear Regession Analysis". The Least Squares Method is a
Regression analysis method. This basically conelates two important demand data: demand
(amount) vs. time.

This assumes that what ever happened in the past, will continue to happen in the future. This
may not always be true.

This is specially useful for Iinear trend (uniform inqeasing or decreasing). lt is useful for
long-term forccasting ofmajor occurrences and aggregate production planning. For example,
linear regression would be very useful to forecast demands for product families (a group of
similar prcducts).

Here, Y : a+ bx ....equation of a straight line, Y

(amount)
Where, y is demand, x is time; a and b are constants.

Z*v *v o=y-E
Lx -nx x atime)

-15-
 Dr. M. Ahsan Akhtar Hasin, IPE, BUET.

llq4plq Demands ofthe past 1 year (in 12 months, starting flom January ofthat year, up to
December ofthat year) are given.

a) Find the Re$ession equation

b) Calculate the demand ofJanuary ofnext year

x (month) y (Sales) x) x_
600 600 I
2 1550 3100 4
3 I500 4500 9
4 I500 6000 t6
5 2400 t2000 25
6 3100 18600
7 2600
8 2900
9 3800
10 4500
1l 4000
12 4900 58,800 t44
tx:78 !v = 33.350 t xy:268,200 tx' = 650

. 268.200 112\t6.5\12779.t7 | -15S.6l:a_2?7q.17-(l5q.6t)16.5)-441.'l

b-': --
650 (12)t6.s)'

a) The regaession equation is -

y = a +bx :441.7 + 359.6lx ..... (Ans.)

b) Forecast for January next year, i.e. the next period is

yr::441.7+(359.61)(13)=5116units, ... (Ans.)

Although, linear regression method is widely used in business, this is suitable for demands of
those products having more or less stable demand pattem. ln case of rapid and large-scale
ups-downs. this is not suitable.

O Forecast Errors

-, There may be differences between actual sales and forecast amount. In statistics, these
variations are called "variance" or "Standard deviation,,or simply ,,residuals,,. While talking
about Forccast Eror, we are refel.ring to the difference between the forecast value and what
-_ acfually occurred.

It is necessary not only to forecast, but also to measure error for future adjustment, and
usefulness ofa method.

-16-
Dr. M. Ahsan Al,:htar Hasin, IPE, BUET.

Forecast Error can be measured and understood in many ways:

1) Mean Absolute Deviation (MAD) - This is the most widely used error measurcment
technique, although very simple. This measures the difference between actual demand
and forecast, without regard to sign (absolute value).
2) Mean Squared Error O4SE) - lt measures the average ofthe squares ofthe differences
between actual and forecast that is, the average squared difference between the
-
forecast demand and the actual demand. This is very similar to measuring Standard
Deviation or Variance ofstatistics. The objective ofsquare is to remove the negative
values.
3) Mean Absolute Percentage Enor (MAPE) - It measures the accuracy as a percentage.
It can be calculated as the average difference ofactual and forecast, divided by actual
demand again to find the percentage difference. It is very similar to MAD, where
N4-{D is divided by actual demand again to find percentage MAD. That's why it is
often called Mean Absolute Percentage Deviation (MAPD).
4) Tracking Signal (TS) - Wlile other ero$ measurc the amourt or quantity or
magnifude ofdifference (between actual and forecast demand), the tracking signal can
tlack not only the amount, but also whether it (forecast) is above or below the actual
demand-

MAD= L I
Lr -r'1. rvheren : no. of perioils. ;

Ar : actual sales in period t: Fr forecast demand for period t.

Demand/actual 3 week MA lA.F,

forecast
1 1200
2 1400
3 1000
4 1500 I200 300
5 1500 1300 200
6 1300 1333 33
7 1600 1433 167
8 1700 t46',7 233
9 1500 I533 33
TotalDeviation 966

So. MAD:966 / 6:161 units

-1',7 -