You are on page 1of 17

Demand Forecasting

in a Supply Chain
Dr RAVI SHANKAR
Dr.
Professor
Department of Management Studies
Indian Institute of Technology Delhi
Hauz Khas, New Delhi 110 016, India
Phone: +91-11-26596421 (O); 2659-1991(H); (0)-+91-9811033937 (m)
Fax: (+91)-(11) 26862620

Email: ravi1@dms.iitd.ac.in, r.s.research@gmail.com,


http://web.iitd.ac.in/~ravi1

D
Demand
dF
Forecasting
ti

Forecasting
Predict the next number in the pattern:

a) 3.7,

3.7,

3.7,

3.7,

3.7,

b) 2
2.5,
5

4
4.5,
5

6
6.5,
5

8
8.5,
5

10
10.5,
5

c) 5.0, 7.5, 6.0, 4.5, 7.0, 9.5, 8.0, 6.5, ?

Forecasting
Predict the next number in the pattern:

a) 3.7,

3.7,

3.7,

3.7,

3.7, 3.7

b) 2.5,
25

45
4.5,

65
6.5,

85
8.5,

10 5 12.5
10.5,
12 5

c) 5.0, 7.5, 6.0, 4.5, 7.0, 9.5, 8.0, 6.5, 9.0

BALANCE OF FORECASTING
EFFORTS
Reference:
Ravi Shankar, Industrial
Engineering & Management
(2010)

Types of forecasting methods


Qualitative methods Quantitative methods
Rely on subjective
opinions from one
or more experts.

Rely on data and


analytical
techniques.

Qualitative Methods
Executive Judgment
Grass Roots
Assessment
Historical analogy
gy

Qualitative
Methods

Market Research
Panel Consensus

Delphi Method

Qualitative forecasting methods


Executive Judgement: banking on the experience of executives,
who have dealt with similar situations.
Historical Analogy: identifying another similar market.
Market Research: trying to identify customer habits; new product
ideas.
Grass Roots: deriving future demand by asking the person
closest to the customer.
Panel Consensus: deriving future estimations from the synergy
of a panel of experts in the area.
Delphi Method: similar to the panel consensus but with
concealed identities.

Quantitative forecasting methods


Time Series: models that predict future demand based
on past history trends
Causal Relationship: models that use statistical
q
to establish relationships
p between various
techniques
items and demand

Components of Demand
Average demand for a period of time
Trend
Seasonal element
Cyclical elements
Random variation

Product Demand over Time


Cycle

Demand

Demand

Demand

SeasonalPattern

Average

Time

Time

Time

TrendwithSeasonalPattern

Demand

Demand

Trend

Random
movement
Time

Time
11

Time Series Models


Try to predict the future based on past data

Assume that factors influencing the past will


continue to influence the future

Naive Approach
Demand in next period is the same as
demand in most recent period
9 August Sales = 120

SEPTEMBER FORECAST = 120

Usually not good

Simple Moving Average


Assumes an average is a good estimator of

future behavior

Used when trend is lesser or absent

Ft +1 =
Ft+1
n
At

A t + A t -1 + A t -2 + ... + A t -n +1
n
= Forecast for the upcoming period, t+1
= Number of periods to be averaged
= Actual occurrence in period t

Simple Moving Average


Ft +1 =

A t + A t -1 + A t -2 + ... + A t -n +1
n

Forecast sales for months 4-6 using a 3-period


moving average.

Month
1
2
3
4
5
6

Sales
(000)
4
6
5
?
?
?

2a. Simple Moving Average


Ft +1 =

A t + A t -1 + A t -2 + ... + A t -n +1
n

Forecast sales for months 4-6 using a 3-period


moving average.

Month
1
2
3
4
5
6

Sales
(000)
4
6
5
?
?
?

Movingg Average
g
(n=3)
NA
NA
NA
(4+6+5)/3=5

Simple Moving Average

Month
M
th
1
2
3
4
5
6

Sales
(000)
4
6
5
3?
?
?

Moving Average
(n=3)
NA
NA
NA
5

Simple Moving Average

2a. Simple Moving Average

Month
M
th
1
2
3
4
5
6

Sales
(000)
4
6
5
3
?
?

Moving Average
(n=3)
NA
NA
NA
5
(6+5+3)/3=4.667

Simple Moving Average

Sales
(000)
4
6
5
3
?7
?

Month
M
th
1
2
3
4
5
6

Moving Average
(n=3)
NA
NA
NA
5
4.667

Simple
2a. Simple
MovingMoving
Average

Sales
(000)
4
6
5
3
7
?

Month
M
th
1
2
3
4
5
6

Average

Moving Average
(n=3)
NA
NA
NA
5
4.667
(5+3+7)/3=5

Weighted Moving Average


Gives more emphasis to recent data

Ft +1 = w1A t + w 2 A t -1 + w 3A t -2 + ... + w n A t -n +1
Weights
decrease
sum

for older data

to 1.0

Weighted Moving Average: 3/6, 2/6, 1/6


Ft +1 = w1A t + w 2 A t -1 + w 3A t -2 + ... + w n A t -n +1

Month
1
2
3
4
5
6

Sales
(000)
4
6
5
?
?
?

Weighted
Moving
Average
NA
NA
NA
31/6 = 5.167

Weighted Moving Average: 3/6, 2/6, 1/6


Ft +1 = w1A t + w 2 A t -1 + w 3A t -2 + ... + w n A t -n +1

Month
1
2
3
4
5
6

Sales
(000)
4
6
5
3
7

Weighted
Moving
Average
NA
NA
NA
31/6 = 5.167
25/6 = 4.167
32/6 = 5.333

Exponential Smoothing
Assumes the most recent observations have

the highest predictive value

gives more weight to recent time periods

Ft+11 = Ft + (At - Ft)


et
Ft+1 = Forecast value for time t+1
= Actual value at time t
At

= Smoothing constant

Exponential Smoothing

Ft+1 = Ft + (At - Ft)


Week (i) Demand (Ai)
1
820
2
775
3
680
4
655
5
750
6
802
7
798
8
689
9
775
10

Given the weekly demand


data what are the exponential
smoothing forecasts for
periods 2-10 using =0.10?
Assume F1=D1

Simple Moving Average

Ft+1 = Ft + (At - Ft)


i

Ai

Week
1
2
3
4
5
6
7
8
9
10

Demand
820
775
680
655
750
802
798
689
775

Fi

= 0.1

0.6
820.00
820.00
820.00
820.00
F2 = F1+ (A1F1)
815.50
793.00=820+0.1(820820)
801.95
725.20=820
787.26
683.08
783.53
723.23
785.38
770.49
786.64
787.00
776.88
728.20
776.69
756.28

Simple Moving Average

Ft+1 = Ft + (At - Ft)


i

Week
1
2
3
4
5
6
7
8
9
10

Ai

Demand
820
775
680
655
750
802
798
689
775

Fi

= 0.1

0.6
820.00
820.00
820.00
820.00
815.50
F3 =
F2+ (A2F2)793.00=820+.1(775820)
801.95
725.20
787.26
683.08=815.5
783.53
723.23
785.38
770.49
786.64
787.00
776.88
728.20
776.69
756.28

Simple Moving Average

Ft+1 = Ft + (At - Ft)


i

Ai

Week
1
2
3
4
5
6
7
8
9
10

Fi

Demand
820
775
680
655
750
802
798
689
775

= 0.1

0.6
820.00
820.00
793.00
725.20
683.08
723.23
770.49
787.00
728.20
756.28

820.00
820.00
815.50
801.95
787.26
783.53
785.38
786.64
776.88
776.69

Exponential Smoothing

Ft+1 = Ft + (At - Ft)


i

Week
1
2
3
4
5
6
7
8
9
10

Ai

Fi

Demand
820
775
680
655
750
802
798
689
775

= 0.1

= 0.6

820.00
820.00
815.50
801.95
787.26
783.53
785.38
786.64
776.88
776.69

820.00
820.00
793.00
725.20
683.08
723.23
770.49
787.00
728.20
756.28

To Use a Forecasting Method


1.Identifythepurpose
2.Collecthistoricaldata
3.Examinedata(plot)
4.Selectappropriatemodels
5.Computeforecastsforhistorical
data and check forecast accuracy
dataandcheckforecastaccuracy
6.Isaccuracyacceptable?
Yes

No

7b.Adjust
parameters
orselectnew
model

7a.Forecastoverplanninghorizon
8.Includequalitativeinformation
9.Monitorresults
30

10

Measures of Forecast Error


et
n

a. MAD = Mean Absolute Deviation

MAD =

(A

- Ft )

t =1

MSE =

c. RMSE = Root Mean Squared Error

- Ft

n
n

b. MSE = Mean Squared Error

t=1

RMSE =

MSE

Ideal values =0 (i.e., no forecasting error)

MEAN ABSOLUTE DEVIATION


(MAD)
n

MAD =

- Ft

= 40
4

t=1

At
Ft
Sales Forecast
220
n/a
250
255
210
205
300
320
325
315

Month
1
2
3
4
5

=10

|At Ft|
5
5
20
10
= 40

What are the Mean Squared Error (MSE) and


Root Mean Squared Error (RMSE) values?
n

MSE =

(A

- Ft )

t =1

= 550 =137.5
4

At

Month
1
2
3
4
5

Ft

Sales Forecast
220
n/a
250
255
210
205
300
320
325
315

RMSE = 137.5

=11.73
|At Ft| (At Ft)2
5
5
20
10

25
25
400
100
= 550

11

Measuring Bias in Forecast: Tracking signal


The tracking signal is a measure of how often our
estimations have been above or below the actual value. It
is used to decide when to re-evaluate using a model.
n

RSFE= (At Ft )

TS =

i=1

RSFE
MAD

Positive tracking signal: most of the time actual


values are above our forecasted values

Negative tracking signal: most of the time actual


values are below our forecasted values
If TS > 4 or < -4, investigate!

Example of Tracking Signal

2/7/2014

35

Linear regression for Forecasting


Linear regression is based on
1. Fitting a straight line to data
2. Explaining the change in one variable through changes
in other variables.

dependent variable = a + b (independent variable)


By using linear regression, we try to explore which
independent variables affect the dependent variable

12

Example: Do people buy more Coke when its


weather is hot?
Coke Sales

Average
Monthly
Temperature

The best line is the one that minimizes the


error
The predicted line is

Y = a + bX
So, the error is

i = y i - Yi
Where: is the error
y is the observed value
Y is the predicted value

Least Squares Method of Linear Regression

The goal of LSM is to minimize the sum of squared


errors

Min

2
i

13

What does that mean?

Coke Sales

Average
Monthly
Temperature

Least Squares Method of Linear Regression

Then the line is defined by

Y = a + bX

a = y bx
b=

xy nxy
x nx
2

Regression Example

y = a+ b X
Month
January
Februaryy
March
April
May
June
July
TOTAL

Advertising
3
4
2
5
4
2

20

b=

xy n x y
x nx
2

a = y bx

1
2
1
3
2
1

X2
9.00
16.00
4.00
25.00
16.00
4.00

XY
3.00
8.00
2.00
15.00
8.00
2.00

10

74

38

Sales

14

Case Study#1: HP desktop


Is it always possible to use it?
Only if the power supply can be assembled in small lead time
Power supply assembly should be at the end of the
manufacturing process

Board
assembly

Hard disk
Assembly

Board
assembly

Power
supply
110 V

Testing

Power
supply
220 V

Testing

Delayed product
differentiation
Product
postponement

Power
supply
110 V

Hard disk
assembly

Testing
Power
supply
220 V

43

Case Study#1: HP desktop

Product
Month

Power
supply
110 V
Board
assembly

Hard disk
assembly

Testing
Power
supply
220 V

110 V PC

Product
220 V PC

10000

8000

14000

4000

16000

2500

12000

6500

18000

2000

15000

4000

14000

3000

11000

7000

13000

5000

10

11000

6000

44

Forecast accuracy improves at different levels


110 V
Months

Demand

MA(4)

220 V
Error

Demand

MA(4)

Total
Error

Demand

MA(4)

Error

10000

8000

18000

14000

4000
(10000+14000+16000+12000)/4)

18000

16000

12000

13000-18000

6500

18000

13000

-5000

2000

5250

3250

20000

18250

15000

15000

4000

3750

-250

19000

18750

-250

14000

15250

1250

3000

3750

750

17000

19000

2000

11000

14750

3750

7000

3875

-3125

18000

18625

13000

14500

1500

5000

4000

-1000

18000

18500

500

10

11000

13250

2250

6000

4750

-1250

17000

18000

1000

2500

18500
18500
-1750

625

MAD

2291.67

1604.17

1020.83

Forecast
Accuracy

83.23%

64.35%

94.38%

(5000+1250+3750+1500+2250) / 6

100-[(5+1.25+3.75+1.5+2.25)/(18+15+14+11+13+11)]100
45

15

Case Study#2: Aggregate Forecast

2/7/2014

46

(c) Dr. Ravi Shankar, AIT


(2008-09)

Learning Lesson of Case 1

What are the Learning


Lessons of this case
Study?

47

General Guiding Principles for Forecasting

G 40: Forecasts are more accurate for larger


groups of items.
G 41: Forecasts are more accurate for shorter
periods
i d off ti
time.
G 42: Every forecast should include an
estimate of error.
G 43: Before applying any forecasting method,
the method should be tested and evaluated.

16

Product Redesign Helps Supply Chain


Competitiveness
G 44: Delayed product differentiation is the key to
this redesign
G 45: Similarly, forecast at the most upstream of the
supply chain (if possible)
G 46:
46 If possible,
ibl never use fforecastt information
i f
ti att
the lower levels. At the lower levels, decisions
should be based on actual demand

49

17

You might also like