Forecasting

in  Supply  Chains
Dr Niniet  Indah  Arvitrida

Graduate Program
Department  of Industrial  Engineering
Institut  Teknologi  Sepuluh  Nopember
2019

1
In  this  course  students  will be  able  to

1. Understand  the  role  of  forecasting  for  both  an  enterprise  and  a  
supply  chain
2. Identify  the  components  of  a  demand  forecast
3. Forecast  demand  using  time-­‐series  methodologies
4. Analyze demand  forecast  error

Main  source:

Chopra,  S.,  Meindl,  P.,  (2009)  Supply  chain  Management:  Strategy,  

Planning,  and  Operation,  New  Jersey;;  Pearson  Prentice  Hall.

2
 What  is Forecasting  ?
Y

Forecasting

n n+1 n+2 Function

of Time
Forecasting is an approach to estimate something
that contains uncertainty and related to time in the
3
future
 Forecasting :
— The  process  of  estimating  future  demand in  terms  of  the  
quantity,  timing,  quality,  and  location for  desired  products  and  
services
— Rely  on  logical  methods  of  manipulating  data that  have  been  
generated  by  historical  events à for  quantitative  method
— Forecasting  is  an  essential  element  of  capital  budgeting.

Sales  will  be  

$200  Million!
Everyone  needs  
forecasting  prior  to  
planning.

4
 Role of Forecasting
in a Supply Chain
— The basis for all planning decisions in a supply
chain
— Used for both push and pull processes
— Production scheduling, inventory, aggregate planning
— Sales force allocation, promotions, new production
introduction
— Plant/equipment investment, budgetary planning
— Workforce planning, hiring, layoffs
— All of these decisions are interrelated
 Characteristics of forecasting
— Forecasts are usually wrong/inaccurate, seldom
correct, and involves error
à find the best method
— Long-term forecasts are usually less accurate than short-
term forecasts.
— Family / grouped / aggregate forecasts are usually more
accurate than item forecast
— In general, the further up the supply chain a company is (the
further they are from the customers), the greater the
distortion of information they receive.
— Forecast should include a measure of forecast error
6
 Components and Methods
— Companies must identify the factors that influence
future demand and then ascertain the relationship
between these factors and future demand
— Past demand
— Lead time of product replenishment
— Planned advertising or marketing efforts
— Planned price discounts
— State of the economy
— Actions that competitors have taken
 Components and Methods
1. Qualitative
— Primarily subjective
— Rely on judgment
2. Time Series
— Use historical demand only
— Best with stable demand
3. Causal
— Relationship between demand and some other factor
4. Simulation
— Imitate consumer choices that give rise to demand
 Assumptions of Forecast

— Information  (data)  about  the  past  is  available

— The  pattern of  the  past  will  continue into  the  
future  (Time  Series  Models).
— Forecasting  methods  assume there  is  some  
underlying  stability in  the  system

9
 Which  item  should  we  forecast?
Independent  Demand: Dependent  Demand:
Finished  Goods;;  spare  parts Raw  Materials,  
Component  parts,
Sub-­assemblies,  etc.

B(4) C(2)

D(2) E(1) D(3) F(2)

10
 Production  System

Considering the order fulfilment strategy

1. Engineer To Order (ETO / Project)

2. Make To Order (MTO / Job Order)
3. Assemble To Order (ATO)
4. Make To Stock (MTS / Mass Production)

Each  of  these  production  systems  requires  different  forecasting  

approach
Customer Order Decoupling Point

Inventory   Raw Work  in  process  (WIP)   Finished  

Suppliers
location materials Parts  and  Component goods

Customer  
order  
decoupling  
point

Environment ETO MTO ATO MTS

 Manufacturing  Lead  Time

ETO

MTO

ATO

MTS

Engineering Procurement Fabrication Assembly Delivery

• MTS  – Forecast  end  items.

• ATO  – Forecast  product  mix  and  calculate  expected  

assembly  and  part  requirements.  

• MTO  –company  resources.

14
Forecasting  Steps
• Decide  what  needs  to  be  forecast
• Level  of  detail,  units  of  analysis (weekly,  monthly,  etc) &  time  horizon  
required
• Evaluate  and  analyze  appropriate  data
• Identify  needed  data  &  whether  it’s  available
• Select  and  test  the  forecasting  model
• Cost,  ease  of  use  &  accuracy
• Generate  the  forecast
• Monitor  forecast  accuracy  over  time

15
Basic  approach  to  demand  forecasting

1. Understand  the  objective  of  forecasting

2. Integrate  demand  planning  and  forecasting  throughout  the  supply  
chain
3. Understand  and  identify  customer  segments
4. Identify  the  major  factors  that  influence  the  demand  forecast
5. Determine  the  appropriate  forecasting  technique
6. Establish  performance  and  error  measures  for  the  forecast

16
Characteristics  of  a  Good  forecast

• Accuracy  à bias  (when  forecast  are  persistently  too  high  or  
too  low)  &  consistency  (concerned  with  the  size  of  forecast  
error)
• Cost
• Response à stability  &  responsive
• Simplicity à easy  to  make,  understand  &  use

17
 Measuring  Accuracy,  Forecast  Errors
To  compare  different  time  series  techniques  or  to  select  the  “best”  set  of  
initial  values  for  the  parameters,  use  a  combination  of  the  the  following  3
metrics (assumption:    errors  are  normally  distributed):

n
• Mean  Absolute  Deviation (MAD)
— Most  popular å A
i =1
i - Fi
— Standard  deviation  of  errors  is  assumed MAD =
by  1.25  times  MAD
n
— Assumption:  error  is  normally  distributed
n
• Mean  Square  Error (MSE) (
å i i
A - F )2

— Is  similar  to  variance of  a  random  sample  (of  errors) MSE = i =1

— Should  be  used  in  tandem  with  MAD n
— Assumption:  mean  of  error  is  zero

• Mean  Absolut  Percent  Error (MAPE) 100 n A i - Fi

MAPE = å
n i =1 A i
Note  for  this  slide: A  is  actual  demand;;  F  is  forecasted  demand
18
Measuring  Accuracy,  Forecast  Errors

Note  for  this  slide: E  is  error;;  F  is  forecasted  demand;;  D  is  actual  demand
19
• Forecast  error  measurements:  
If  you  are  working  with  a  low-­‐volume  or  sporadic  item  then  MAD  is  a  
good  choice,  while  the  MAPE  should  be  avoided.
MAPE  is  good  for  high  volume  /  fairly  consistent  and  regular  demand  
pattern.

20
 Examples  of  Production  Resource  Forecasts
Forecast   Units  of  
Time  Span Item  Being  Forecast
Horizon Measure
• Product  lines
• Factory  capacities
• Planning  for  new  products Dollars,  tons,  
Long-­Range Years • Capital  expenditures etc.
• Facility  location  or  expansion  
• R&D
• Product  groups
• Department  capacities
Medium-­ Dollars,  tons,  
Months • Sales  planning
Range • Production  planning  and   etc.
budgeting
• Specific  product  quantities
• Machine  capacities  
• Planning
• Purchasing Physical  units  
Short-­Range Weeks • Scheduling of  products
• Workforce  levels
• Production  levels
• Job  assignments
21
 Demand  Behavior
— Trend
◦ a  gradual,  long-­‐term  up  or  down  movement  of  demand
◦ Due  to  population,  technology  etc

— Random  variations
◦ movements  in  demand  that  do  not  follow  a  pattern

— Cycle
◦ an  up-­‐and-­‐down  repetitive  movement  in  demand
◦ Non-­‐annual;    multi-­‐year
◦ Due  to  interactions  of  factors  influencing  economy

— Seasonal  pattern
◦ an  up-­‐and-­‐down  repetitive  movement  in  demand  occurring  periodically
◦ Occurs  within  one  year
◦ Due  to  weather,  customs  etc

22
 Forms  of  Forecast  Movement
often  regarded  
as  part  of  the  
trend

Demand
Demand

Random  
movement

Time Time
(a)  Trend (b)  Cycle

Demand
Demand

Time Time
(c)  Seasonal  pattern (d)  Trend  with  seasonal  pattern
23
 Types  of  Forecasts
— Qualitative  (Extrinsic) :  
◦ requiring  no  manipulation  of  data,  
◦ used  only  judgment of  the  forecaster,  
◦ Evaluating  factors  other  than  historical  data
◦ concern  with  identifying  various  factors  that  can  influence  
demand  
◦ suitable  for  long  term forecasting

— Quantitative  (Intrinsic) :  
◦ no input  of  judgment,  
◦ mechanical  procedure that  produce  quantitative  result,  
◦ Rely  on  historical  information
◦ using  statistical  techniques  
◦ suitable  for  short  term forecasting

24
 1.  Qualitative  Methods  (Extrinsic)

¨ Applications: New product Development, new

technology à An unusual product or a unique project
is being contemplated
— Technique  :  
◦ Management  Decision  – Jury  of  executive  opinion
◦ Delphi  Technique
◦ Sales  Force  Composite
◦ Growth  curve
◦ Scenario  writing
◦ Market  Research
◦ Historical  Analogies
◦ Life  Cycle  Curves
◦ Focus  groups

25
 Qualitative  Methods :  Example
Type Characteristics Strengths Weaknesses
Executive   A  group  of  managers   Good  for  strategic  or   One  person's  opinion  
opinion meet  &  come  up  with   new-­product   can  dominate  the  
a  forecast forecasting forecast

Market   Uses  surveys  &   Good  determinant  of   It  can  be  difficult  to  
research interviews  to  identify   customer  preferences develop  a  good  
customer  preferences questionnaire

Delphi   Seeks  to  develop  a   Excellent  for   Time  consuming  to  

method consensus  among  a   forecasting  long-­term   develop
group  of  experts product  demand,  
technological  
changes,  and  

26
 1.  Qualitative  Methods
-­‐in  general  -­‐

Drawbacks :
Advantages :
– No historical data is
required.
– Involving experts for
modeling, so the results are
quite accurate.
– High cost à for paying
labour time of the
experts
– Time consuming,
particularly in building
the models and selecting
appropriate methods for
validation

27
 2.  Quantitative  methods  (Intrinsic)

Used  when  sufficient  historical  data  are  available  and  when  

these  data  values  are  judged  to  be  representative  of  the  
unknown  future

Assumption  : Future  pattern  following  past  pattern

Can  be  categorized  :

— Time  series  methods :  focus  entirely  on  pattern,  pattern  
change,  disturbance  caused  by  random  influence
◦ Moving  average,  exponential  smoothing,  Holts,  winter
— causal  methods :  identification  and  determination  of  
relationship  between  the  variable  to  be  forecast  and  other  
influencing  variable
◦ Regression
28
 Demand  Components  :
Observed  demand  (O)  =
Systematic  component  (S)  + Random  component  (R)

Level (current deseasonalized demand)

Trend (growth or decline in demand)

Seasonality (predictable seasonal fluctuation)

• Systematic component: Expected value of demand
• Random component: The part of the forecast that deviates
from the systematic component
• Forecast error: difference between forecast and actual demand
29
The  systematic  component  may  take  various  
forms  in  time-­‐series  forecasting  methods
• Multiplicative  :  
Systematic  component  =  level  x  trend  x  seasonal  factor

• Additive  :  
Systematic  component  =  level  +  trend  +  seasonal  factor

• Mixed  :  
Systematic  component  =  (  level  + trend  )  x  seasonal  factor

30
 Time  Series  Models

Forecasting  Method Applicability

Moving  average No  trend  or  seasonality
Simple  exponential   No  trend  or  seasonality
smoothing
Holt’s  model Trend  but  no  seasonality
Winter’s  model Trend  and  seasonality
1.  Moving  Average  (MA)

How  it  works  approach :

If  the  number  of  period  used  is  3  years (MA  =  3  years), forecasted  
demand  for  next  year  will  be  the  average  demand  of  the  total  
demand  over  the  past  3  years.
• Naive  forecast (N=1)
• demand  the  current  period  is  used  as  next  period’s  
forecast
• Simple  moving  average (N>1)
• stable  demand  with  no  pronounced  behavioral  patterns
• Weighted  moving  average (N>1)
• weights  are  assigned  to  most  recent  data
32
 Example  :  MA(1)
• MA  (1)  à N=1,  the  more  popular  term  for  this  is  NAIVE  METHOD  

ORDERS
FORECAST
MONTH PER  MONTH
Jan 120 -­‐
Feb 90 120
Mar 100 90
Apr 75 100
May 110 75
June 50 110
July 75 50
Aug 130 75
Sept 110 130
Oct 90 110
Nov                                          -­‐ 90

33
Simple  Moving  Average  

n
S
i =  1
Di
MAn =  
n
where

n = number  of  periods  in  

the  moving  average
Di = demand  in  period  i

34
Example  :  MA(3)

ORDERS MOVING   3
MONTH PER  MONTH AVERAGE S
i =  1
Di
Jan 120 – MA3 =  
Feb 90 – 3
Mar 100 –
Apr 75 103.3 90  +  110  +  130
= 3
May 110 88.3
June 50 95.0
July 75 78.3 =  110  orders
Aug 130 78.3
for  Nov
Sept 110 85.0
Oct 90 105.0
Nov -­‐ 110.0

35
Example  :  MA(5)

ORDERS MOVING  
5
MONTH PER  MONTH AVERAGE
Jan 120 –
S
i =  1
Di

Feb 90 – MA5 =  
Mar 100 –
5
Apr 75 –
90  +  110  +  130+75+50
May 110 – =
June 50 99.0
5
July 75 85.0
Aug 130 82.0 =  91  orders
Sept 110 88.0 for  Nov
Oct 90 95.0
Nov -­‐ 91.0

36
 Smoothing  Effects

150  –

125  – 5-­month

100  –
Orders

75  –

50  – 3-­month

Actual
25  –

0  – | | | | | | | | | | |
Jan Feb Mar Apr May June July Aug Sept Oct Nov
Month
37
2.  Weighted  Moving  Average

Adjusts  moving  average  method  to  more  closely  

reflect  data  fluctuations

WMAn =  S Wi Di
i =  1

where
Wi =  the  weight  for  period  i,  
between  0  and  100  
percent

S Wi =  1.00
38
Weighted  Moving  Average  Example

MONTH   WEIGHT   DATA

August 17% 130
September 33% 110
October 50% 90
3
November  Forecast WMA3 =  S Wi Di
i =  1

=  (0.50)(90)  +  (0.33)(110)  +  (0.17)(130)

=  103.4  orders
39
 Disadvantages  of  M.A.  Methods

• Increasing  “n” makes  forecast  

less  sensitive  to  changes
• Do  not  forecast  trends  well
• Require  sufficient  historical  
data

Maka, dikembangkanlah metode

EXPONENTIAL SMOOTHING
40
 3.  Exponential  Smoothing  (ES)
• Moving  average  treats  all  period  similarly,  but  ES  considers  the  
newer  data  to  be  more  important  than  the  previous  ones.
• ES  has  a  constant  -­‐ α
• α represents  the  importance  weight  of  the  newest  actual  data,  
while  1  – α  (damping  factor)  is  the  weight  for  the  forecasted  
demand  of  the  current  period.  
• damping  factor  is  a  correction  factor  of  the  forecasting  value  to  
minimize  the  effect  of  data  instability/variation.
• α is  set  between  0  and  1.  
• The  selection  of  alpha  is  a  trade  off  between  stability  and  
responsiveness.
Ft =  a Dt-1 +  (1  -­‐ a)Ft-1
F  is  the  level  factor  of  demand.  Trend  and  seasonal  factors  are  not  considered.
41
Exponential  Smoothing  (cont.)

Ft =  a Dt-1 +  (1  -­ a)Ft-1

where:
Ft = forecast  for  next  period
Dt-1 = actual  demand  for  present  period
Ft-1 = previously  determined  forecast  for  
present  period
a = weighting  factor / smoothing  constant /  
level  smoothing  parameter

Ft =  a Dt-1 +  (1  -­ a) a Dt-1 +  (1  -­ a)2 a Ft-1
And  soon
42
 Exponential  Smoothing:  Concept
• Include  all  past  observations
• Weight  recent  observations  much  more  
heavily  than  very  old  observations:
0 < a < 1
weight
Decreasing weight given a
to older observations
a (1 - a )
a (1 - a ) 2
a (1 - a ) 3
today
!
43
Exponential  Smoothing  (α=0.30)
Ft =  a Dt-1 +  (1  -­‐ a)Ft-1
PERIOD MONTH DEMAND F2 =  aD1 +  (1  -­ a)F1
1 Jan 37 =  (0.30)(37)  +  (0.70)(37)
2 Feb 40 =  37
3 Mar 41
4 Apr 37 F3 =  aD2 +  (1  -­ a)F2
5 May   45 =  (0.30)(40)  +  (0.70)(37)
6 Jun 50
=  37.9
7 Jul   43
8 Aug   47
F13 =  aD12 +  (1  -­ a)F12
9 Sep   56
10 Oct 52 =  (0.30)(54)  +  (0.70)(50.84)
11 Nov 55 =  51.79
12 Dec   54

44
Exponential  Smoothing  (cont.)

FORECAST,  Ft +  1
PERIOD MONTH DEMAND (a =  0.3) (a =  0.5)
1 Jan 37 37 37
2 Feb 40 37.00 37.00
3 Mar 41 37.90 38.50
4 Apr 37 38.83 39.75
5 May   45 38.28 38.37
6 Jun 50 40.29 41.68
7 Jul   43 43.20 45.84
8 Aug   47 43.14 44.42
9 Sep   56 44.30 45.71
10 Oct 52 47.81 50.85
11 Nov 55 49.06 51.42
12 Dec   54 50.84 53.21
13 Jan – 51.79 53.61

45
 Exponential  Smoothing  (cont.)

70  –

60  – Actual a =  0.50

50  –

40  –
Orders

a =  0.30
30  –
Larger a,  more  responsive  forecast;;  Smaller  a,  
20  – smoother  forecast  
“Best”  a can  be  found  by  Solver  
10 – Suitable  for  relatively  stable  time  series

0  – | | | | | | | | | | | | |
1 2 3 4 5 6 7 8 9 10 11 12 13
Month
46
Similarities  between MA  and ES

• Both  techniques  assume  demand  pattern  to  be  stationary.

• Both  methods  rely  on a  single  parameter:  N  (number  of  
periods)  for MA,  and α (smooting  parameter)  for ES.
• Both  MA  and  ES  are  less sensitive  towards  trend,  so  both  of  
them  will  be late  in  identifying  trend.
• Both  methods  have  a  similar  error  distribution  when  the α =  
2  /  (N+1)

47
• What  happens  when  there  is  a  definite  trend?

A trendy clothing boutique has had the following sales

over the past 6 months:
1 2 3 4 5 6
510 512 528 530 542 552

560
550
Actual
540
Demand 530
Forecast
520
510
500
490
480
1 2 3 4 5Month
6 7 8 9 10

48
4.  Holt’s  – Trend  Corrected
(double  exponential  smoothing)
Level  à St+1 =  aDt+1 +  (1-­‐a)(St +  Gt)
Trendà Gt+1 =  b(St+1 – St)  +  (1-­‐b)Gt

Ft+τ =  St +  τGt

D  =  Actual  Demand
S  =  Intercept  (level)
G  =  Slope  (trend)
• More  stability  is  given  to  the  slope  estimate  which  
implies  β ≤  α.
• τ is  the  amount  of  step  ahead  forecast

49
The  larger  the  Beta  value  leads  to  a  more  progressive  
forecasting  results,  in  terms  of  modeling  trend.

350

300

250 Demand
Holt  B(0.1)
200 Holt  B(0.3)

150

100
1 2 3 4 5 6 7 8 9 10 11 12
50
Holt’s  – Trend  Corrected
• alfa  =  0,1  ;  beta  =  0,1

Year Quarter Black  Plastic Level Trend Forecast

2042,737 286,606
1998 I 2250 225 76,17173 301,1717
II 1737 444,7546 90,53001 535,2846
III 2412 722,9561 109,2972 832,2533
IV 7269 1475,928 173,6646 1649,593
1999 I 3514 1836,033 RAW  
192,3087 2028,342
DATA  /  
II 2143 2039,808 ACTUAL  
193,4553DATA
2233,263
III 3459 2355,837 205,7127 2561,549
IV 7056 3010,995 250,6572 3261,652
2000 I 4120 3347,487 259,2406 3606,727
II 2766 3522,654 250,8334 3773,488
III 2556 3651,739 238,6585 3890,398
IV 8253 4326,658 282,2845 4608,942 51
Holt’s  – Trend  Corrected

S0 =  intercept,  G0 =  slope

Please  note,  2  digits  after  decimal  points  should  be  appropriate

52
 Holt’s  – Trend  Corrected

St+1 =  aDt+1 +  (1-­a)(St +  Gt)  

S1 =  (0,1)(2250) +  [(0,9)(2042.74 +286.61)

Gt+1 =  b(St+1 – St)  +  (1-­b)Gt

G1 =  (0,1)(2321.41 – 2042.74)  +   Ft+τ = St + τGt
(0,9)286.61 F1 = 2042.74
+(1)286.61

53
Holt’s  – Trend  Corrected

54
5.  Accommodating  Seasonality (No  Trend)

w1 w2 w3 W4

M 162 173 146 161

T 122 115 131 118

W 142 150 130 129

TH 173 176 169 166

F 225 235 219 243

55
300

250

200

150

100

50

0
M T W T F M T W T F M T W T F M T W T F

56
Procedures

• Compute  the  sample  mean  of  all  the  data

• Calculate  seasonal  factor  for  each  period  by  dividing  
each  observation  by  the  sample  mean.  

57
From  the  Above  Data

• Average  of  all   w1 w2 w3 w4 AVR

observations  is   M 0.99 1.05 0.89 0.98 0.98

164.25.
T 0.74 0.70 0.80 0.72 0.74
• Divide  each  
W 0.86 0.91 0.79 0.79 0.84
observation  by  the  
mean  value
T 1.05 1.07 1.03 1.01 1.04

• Average  factors  
F 1.37 1.43 1.33 1.48 1.40
corresponding  to  the  
same  period  of  the  
season

58
Forecasts  for  Next  Week

• Average  multiplied  by  those  factors

M  =  0.98  x  164.25  =  161
T  =  0.74  x  164.25  =  121
W  =  0.84  x  164.25  =  138
T  =  1.04  x  164.25  =  171
F  =  1.40  x  164.25  =  230

59
6.  Winter’s  – Trend  and  Seasonality  Corrected

This  method  is  appropriate  when  the  systematic  

component  of  demand  has  a  level,  a  trend,  and  a  
seasonal  factor.  

In  this  case  we  have  

or

Ft+τ =  (St+ τ Gt)(ct +1)

60
Winter’s  -­‐-­‐ Static  Methods

Systematic component = (level + trend) × seasonal factor

Ft+τ =  (S+ τ G)(ct +1)

where

S = estimate  of  level  at  t =  0  

G = estimate  of  trend
ct = estimate  of  seasonal  factor  for  Period  t
Dt = actual  demand  observed  in  Period  t
Ft = forecast  of  demand  for  Period  t
Winter’s  -­‐-­‐ Adaptive  Forecasting

• The  estimates  of  level,  trend,  and  seasonality  are  

adjusted  after  each  demand  observation
• Estimates  incorporate  all  new  data  that  are  
observed

Ft+τ =  (St+ τ Gt)(ct +1)

 St+1 =  a(Dt+1/ct+1)  +  (1-­‐a)(St+Gt)
Gt+1 =  b(St+1 – St)  +  (1-­‐b)Gt
ct+p+1  =  g(Dt+1/St+1)  +  (1-­‐g)ct+1
Ft+τ =  (St+ τ Gt)(ct +1)

• S  =  time  series  (level)

• G  =  trend
• c  =  seasonal  factor
• p  =  the  number  of  periods  after  which  the  seasonal  cycle  
repeats

63
 Winter’s  – Trend  and  Seasonality  Corrected
The  three  steps:

• Step  1:  Deseasonalize demand  and  run  linear  regression  to  

estimate  level  and  trend
• Step  2:  Estimate  seasonal  factors
• Step  3:  Forecast  the  demand

64
Winter’s  – Trend  and  Seasonality  Corrected

• Step  1:  Deseasonalize  the  data

p  is  the  number  of  periods  after  which  the  seasonal  cycle  repeats

65
 Deseasonalized  data  (for  Winter’s)
Year Quarter Black  Plastic Deseasonalized  
N  =  20,  p  =  4,  so  use  the  “even”  equation  
1998 I 2250
II 1737
III 2412 3575
IV 7269 3783,75

D
1999 I 3514 3965,375
II 2143 4069,625
ingat,  bahwa  
III 3459 4118,75
ada  5  cycle  
IV 7056 4272,375 kejadian  
2000 I 4120 4237,375 seasonal  (lihat:  
1 II 2766 4274,125 sheet  Raw  
III 2556 4595,125 Data)
IV 8253 4968,5
2001 I [2250  
5491 +  3514  +   5390,375
(2*(1737+2412+7269))]/(2*4)
II 4382 6083
III 4315 6575,375
IV 12035 6509,25
2002 I 5648 6489,5
II 3696 6688,25
III 4843
IV 13097 66
67
 Deseasonalized  data  (for  Winter’s)
Deseasonalized   Pooled  Seasonal  
Year Quarter Black  Plastic Deseasonalized   Seasonal  Factor
Regression Factor

1998 I 2250 2819,603493 0,797984541 0,899279159

II 1737 3046,459375 0,570170085 0,598875319
III 2412 3575 3273,315257 0,736867613 0,697570004
IV 7269 3783,75 3500,17114 2,076755596 1,800289582

D
1999 I 3514 3965,375 3727,027022 0,942842641
II 2143 4069,625 3953,882904 0,541998853
III 3459 4118,75 4180,738787 0,82736573
IV 7056 4272,375 4407,594669 1,600873159
2000 I 4120 4237,375 4634,450551 0,888994273
II 2766 4274,125 4861,306434 0,568982852
III 2556 4595,125
3 5088,162316 0,502342465
IV 8253 4968,5 5315,018199 1,552769848
Y  =  A  +  B*t
2001 I 5491 5390,375 5541,874081 0,990820058
II 4382 6083
Y t   =  2592,75  +  (226,85)  1
5768,729963 0,759612606
III 4315 6575,375 (where  t  =  the  sequence  number  of  
5995,585846 0,719696142
IV 12035 6509,25 data  ;;  A  =  intercept  ;;  B  =  slope)
6222,441728 1,934128197
2002 I 5648 6489,5 6449,29761 0,875754282
II 3696 6688,25 6676,153493 0,553612197
III 2
4843 6903,009375 0,701578071
IV 13097 7129,865257 1,83692111
intercept 2592,74761
slope 226,8558824 68
 Deseasonalized  data  (for  Winter’s)
Deseasonalized   Pooled  Seasonal  
Year Quarter Black  Plastic Deseasonalized   Seasonal  Factor
Regression Factor

1998 I 2250 2819,603493 0,797984541 0,899279159

II 1737 3046,459375 0,570170085 0,598875319
III 2412 3575 3273,315257 0,736867613 0,697570004
IV4 7269 3783,75 3500,17114 2,076755596 1,800289582
Seasonal  Factor  =  
D
1999 I 3514 3965,375 3727,027022 0,942842641
II 2143 4069,625 3953,882904
Actual  Data/Deseasonalized  Regression 0,541998853
III 3459 4118,75 4180,738787 0,82736573
Ex.  SF 1   =  2250/2819,6
IV 7056 4272,375 4407,594669 1,600873159
2000 I 4120 4237,375 4634,450551 0,888994273
II 2766 4274,125 4861,306434 0,568982852
5
III 2556 4595,125 5088,162316 0,502342465
Pooled  Seasonal  Factor  =  
IV 8253 4968,5 5315,018199 1,552769848
2001 I 5491 5390,375 The  average of seasonal
5541,874081 factor for
0,990820058 the
II 4382 6083 similar quarter
5768,729963 0,759612606
III 4315 6575,375 5995,585846 0,719696142
IV 12035 6509,25 6222,441728 1,934128197
2002 I
II
5648
3696
6489,5
6688,25
Step  2:  Estimate  seasonal  factors
6449,29761
6676,153493
0,875754282
0,553612197
III 4843 6903,009375 0,701578071
IV 13097 7129,865257 1,83692111
intercept 2592,74761
slope 226,8558824 69
 Winter’s  – Trend  and  Seasonality  
Corrected  (using  Seasonal  Factor)
Step  3:  Forecast  the  demand

These  
constants  
are  from  
the  
We  obtain  the  initial  estimates  of  level  and  trend  exactly   previous  
as  in  the  static  case  when  we  deseasonalized the   step  
demand (deseaso
(S0 =  intercept,  G0 =  slope) nalized  
demand-­
step)

Note:  the  Winter’s  method  can  be  started  from  this  step,  if  S0,  G0,  c1,  c2,  c3,  
and  c4 are  known  in  advance 70
 Winter’s  – Trend  and  Seasonality  
Corrected  (using  Seasonal  Factor)

These  
constants  
are  from  
the  
previous  
St+1 =  a(Dt+1/ct+1)  +  (1-­a)(St+Gt) step  
S1 =  (0,1)(2250/0,899) +   (deseaso
(0,9)(2592.75+226.86) nalized  
demand-­
step)

Gt+1 =  b(St+1 – St)  +  (1-­b)Gt

G1 =  (0,1)(2787.84–2592.75)  
+  (0,9)226.86

71
Winter’s  – Trend  and  Seasonality  
Corrected  (using  Seasonal  Factor)

Ft+τ = (St+ τ Gt)(ct +τ-p)

F2 = (2787.84+223.68)0.599

72
Winter’s  – Trend  and  Seasonality  
Corrected  (using  Seasonal  Factor)

ct+p+1  =  g(Dt+1/St+1)  +  (1-­g)ct+1

c0+4+1  =  g(D0+1/S0+1)  +  (1-­g)c0+1
c5 =  (0,1(2250/2787.84)  +  (0,9)0,899
73
 Symbol  adjustment  – a  little  note  …

St =      Lt     à represents  level

Gt =      Tt     à represents  trend
Ct =      St     à represents  seasonal
τ =      n      à the  amount  of  step  ahead  forecast

In  these  slides In  Chopra’s  book

74
 The  Role  of  IT  in  Forecasting

• Forecasting  module  is  core  supply  chain  software

• Can  be  used  to  best  determine  forecasting  methods  for  
the  firm  and  by  product  categories  and  markets
• Real  time  updates  help  firms  respond  quickly  to  
changes  in  marketplace
• Facilitate  demand  planning
 Risk  Management
• Errors  in  forecasting  can  cause  significant  
misallocation  of  resources  in  inventory,  facilities,  
transportation,  sourcing,  pricing,  and  information  
management
• Common  factors  are  long  lead  times,  seasonality,  
short  product  life  cycles,  few  customers  and  lumpy  
demand,  and  when  orders  placed  by  intermediaries  in  
a  supply  chain
• Mitigation  strategies  – increasing  the  responsiveness  
of  the  supply  chain  and  utilizing  opportunities  for  
pooling  of  demand
 Forecasting  In  Practice

• Collaborate  in  building  forecasts

• Share  only  the  data  that  truly  provide  value
• Be  sure  to  distinguish  between  demand  and  sales
Exercise
Quarterly  demand  for  flowers  at  a  
wholesaler  are  as  shown.  
Forecast  quarterly  demand  for  year  5  
using  :
a. Moving  average  N  =  4.
b. Simple  exponential  smoothing  
with  a  =  0.1
c. Holt’s  model  with  a  =  0.1  and  b  =  
0.1.
d. Winter’s  model  (with  
deseasonalized-­‐demand  
approach)  with  a  =  0.1,  b  =  0.1,  
and  c  =  0.1.  
Evaluate  the  MAD,  MAPE,  MSE,  bias,  
and  TS  in  each  case.  
Which  of  the  three  methods  do  you   S0  =  102.27  ;;  G0 =  3.46
prefer?  Why?  
 Discussion
1. What  role  does  forecasting  play  in  the  supply  chain  of  an  assemble  
to  order  manufacturer  such  as  Dell?
2. What  are  the  systematic  and  random  components  you  would  expect  
in  demand  for:
• Rice  in  Indonesia?
• Salt  in  Indonesia?
• Food  and  beverages  such  as  Coca-­‐Cola?
• Automobile  such  as  Toyota?
3. Why  should  a  manager  should  be  suspicious  if  a  forecaster  claims  to  
forecast  historical  demand  without  any  forecast  error?
4. Give  examples  of  products  that  display  seasonality  of  demand.
5. What  is  the  problem  if  a  manager  uses  last  year’s  sales  data  instead  
of  last  year’s  demand  to  forecast  demand  for  the  coming  year?
6. How  can  the  manager  use  information  from  MAD,  MAPE,  and  TS?
79