Demand

Demand Forecasting
Forecasting in
in aa
Supply
Supply Chain
Chain

Presented by
Prof. M. K. Tiwari
At the end of session you will…
• Understand the role of forecasting
for both an enterprise and a Supply
Chain (SC)
• Identify the components of a
demand forecasts.
• Forecast demand in a SC given
historical demand data using time
series methodologies.
• Analyze demand forecasts to
estimate forecast error.
 Forecasting!......why?
• Push system requires planning about:
– Level of production
• Pull system requires planning about:
– Level of available capacity
– Level of inventory
• Both require future demand of customers.
• Either Pull or push, both processes are
driven by customer demand.
 Example of Dell Computer:
Mastering Pull and Push
• Dell orders components anticipating
customers order (Push)
• It determines capacity of assembly
plants on customer demand basis.
(Pull)
• For both purposes it requires demand
forecasting.
Forecasting: Definition
and its role
Definition: In its simplest form “It is
estimation of expected demand over a
specified future period.”
– If each SC stage makes own demand
forecast variation is unavoidable.
– Collaborative forecasts tend to be more
accurate.
• Role:
– This accuracy enables SC to be more
responsible and efficient in serving their
customers.
Forecasting makes decisions
about:
1. Production: scheduling, inventory control,
aggregate planning, purchasing.

2. Marketing: sales-force allocation,

promotions, new product introduction.

3. Finance: plant/equipment investment,

budgetary planning.

4. Personal: workforce planning, hiring, layoffs.

Characteristics of forecasts
1. Should include both expected and measure of
forecast error (demand uncertainty). High
Uncertaint
y
• Consider, two car dealers
• One expects sales between 100 and 1900
• Other expects sales between 900 and 1100.
Low
Uncertainty

• Even though for both average sales is 1000,

sourcing strategy will be different.
• First dealer will have to arrange more resources
due to higher forecasting error.
Characteristics of forecasts

2. Long term forecasts are usually less

accurate than short term forecasts.

3. For same percentage error, aggregate

forecasts (e.g. GDP of a country) are
usually more accurate than
disaggregate forecasts (e.g. yearly
revenue of company or product wise
details).
Characteristics of forecasts
• The classic example of summing up the
forecast error is bullwhip effect. Here
order variation is amplified as they
move up in SC from the end customers.
• Mature products with stable demand
are usually easiest to forecast.
• Forecasting is difficult when either the
supply of raw materials or the demand
for the finished products is highly
variable.
Factors related to demand
forecast
• Past demand
• Lead time of products
• Planned advertising or
marketing efforts
• State of the economy
• Planned price discounts
• Actions competitors have taken.
Classification of forecasting
methods
• Qualitative:
– Methods are subjective and rely on
human judgment.
– Appropriate when there is little
historical data available or experts
have market intelligence.
– Used to forecast demand several
years into the future in a new
industry.
Classification of forecasting
methods…
• Time series:
– Uses historical demand to make
forecasts.
– Based on assumption that past
demand history is a good indicator of
future demand.
– Appropriate when the basic demand
pattern does not vary significantly
from one period to next.
– Simple to use and can serve as a good
starting point.
Classification of forecasting
methods…
• Causal
– Assumes that demand forecast is highly
correlated with certain factors in the
environment (e.g. state of economy,
interest rates etc.).

– This method find the correlation

between demand and environment and
use estimates of environment factors to
forecast future demand.
Classification of forecasting
methods…
• Simulation
– These methods imitate consumer choices that
give rise to demand to arrive at a forecast.

– Using it a firm can combine time series and

causal method to answer:

1. What will be impact of price promotion?

2. What will be the impact of a competitor opening a
store nearby?
3. Airlines simulate customers buying behavior to
forecast demand for higher fare seats.
Appropriate method
• Several studies have indicated that using
multiple forecasting method is more
effective than any individual method.

• Deal with time series method when

future forecast is expected to follow
historical method.

• Historical demand, growth pattern, any

seasonal pattern influence the forecast.
Components of demand
• Observation demand can be broken
into two components.
– Observed demand (0)=systematic (S)
+random (R)
• Systematic component measures the
expected value of demand.
• Random component is that part of
forecast that deviate from the
systematic part.
Company
can not
forecast
this value.
 Basic approach
• Step 1: Understand the objective of
forecasting

• Step 2: Integrate demand planning

and forecasting through the SC

• Step 3: Understand and identify

customers segments
 Basic approach…
• Step 4: Identify the major factors
that influence the demand forecast

• Step 5: Determine the appropriate

forecasting technique

• Step 6: Establish performance and

error measure for the forecast.
Step 1: Understand the objective
of forecasting
• Clearly identify the decisions such as:
– How much of a particular product to make?
– How much to inventory?
– How much to order?
• It is important that all parties must come
up with a common forecast demand.
• Failure to make such decisions jointly may
results either too much or too little
product in various stages of supply chain.
Step 2: Integrated demand
planning and forecasting
through SC
• A company should link its forecast to all
planning activities throughout SC.

• Link should exist at both the information

system and human resource system.

• As a variety of functions are affected by the

outcomes of the planning process, it is
important that all of them are integrated into
the forecasting process.
 Step 3: Understand and
identify customers segments
• Customers may be grouped by similarities
in service requirements, demand volumes,
order frequency, demand volatility,
seasonality.

• Companies may use different forecasting

methods for different segments.

• Clear understanding facilitates an accurate

and simplified approach for forecasting.
Step 4: Identify major factors that
influence the demand forecasts

• A proper analysis of major factors is

central to developing an appropriate
forecasting technique.
• The main factors are demand, supply,
and product-related phenomena.
• On the demand side, a company must
ascertain whether demand is growing,
declining or has a seasonal pattern.
• Must be based on demand– not sales
data.
Step 5: Supply side Vs Product
side
• On the supply side, a company must
consider available supply sources to
decide on accuracy of forecast desired.
– If alternative supply sources with short lead
time is available, a highly accurate forecast
may not be specially important.
• On the product side, firm must know the
number of variants of a product being
sold.
– If demand for a product influences or is
influenced by demand of another product,
two forecasts are made jointly.
Step 6: Determine appropriate
forecasting technique
• A company should first understand the
dimensions that will be relevant to
forecasts.

• These dimensions include geographical

area, product groups, and customers
groups.

• A firm should be wise enough to have

different forecasts and techniques for
each dimension.
Step 7: Establish performance and
error measure of forecast
• These measures should evaluate accuracy and
timeliness of forecast.
• Measure should correlate with the objective of
the business decisions based on the forecasts.
• Example:
– A mail order company uses forecast to place orders to
suppliers.
– Orders are send to the suppliers with two months lead
time
– Orders are to provide company with a quantity
minimizing both extra product left over at the end of
sale season and any lost sale due to unavailability.
Step 7: Establish performance and
error measure of forecast…
– At the end of season the company
must compare actual demand to
forecasted demand to estimate the
accuracy of forecast.

– The observed accuracy should be

compared with the desired and
resulting gap should be used to
identify corrective action that
company needs to take.
Time series forecast
methods

• Goal of forecasting is to predict

systematic component demand and
estimate the random component.

• In general, systematic component

data contains 3 factors:
– level factor (L)
– trend factor (T) , and
– seasonal factor (S).
Forms of seasonal component
• Multiplicative:
systematic component=level * trend *
seasonal factor

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

• Mixed:
systematic component=(level + trend)*
seasonal factor
 Static methods

For static methods, the level, trend, and

seasonality within systematic component
is estimated on the basis of historical
data and then same values are used for
future forecasts.
Mathematical model

• The forecast in period t for demand

in period t+l
Ft+1=[L+(t+l)T]St+1
where,
L=estimate of level at t=0
T=estimate of trend
St=estimate of seasonal factor for period t
Dt= actual demand observed for period t
Ft=forecast for demand for period t
Example problem to estimate L, T, and S.
• Tahoe , a producer of salt noticed that his
retailers always overestimated the demand. This
lead to his excess inventory holding costs of rock
salt used in the production of salt. To reduce his
inventory costs. Tahoe decided to produce a
collaborative demand forecast.

• The demand is measured on quarterly basis and

demand pattern repeats every year (i.e. p =4). p is
the periodicity defined as the number of periods
after which seasonal cycle repeats itself.
Quarterly demand for Tahoe salt
Year Quarter Period Demand Dt
2000 2 1 8,000
2000 3 2 13,000
2000 4 3 23,000
2001 1 4 34,000
2001 2 5 10,000
2001 3 6 18,000
2001 4 7 23,000
2002 1 8 38,000
2002 2 9 12,000
2002 3 10 13,000
2002 4 11 32,000
2003 1 12 41,000
Estimation of three
parameters
• Two steps
1. Deseasonalize demand and run
linear regression to estimate level
and trend.

1. Estimate seasonal factors.

Estimating level at period 0
and trend

• First, deseasonalize the demand data.

• Deseasonalized demand represents the

demand that would have been observed
in the absence of a seasonal
fluctuations.
 Method
• To ensure that each season is given equal
weight when deseasonalizing the demand,
take average of p consecutive periods.
• Average of demand from period l+1 to l +
p provides deseasonalized demand for
period l+(1 + p)/2.
• This method provides deseasonalized
demand
– for existing period, if p is odd,
– at a point between period l+(p)/2 and l +1+
( p)/2 , if p is even,
 Method…
• By taking the average of deseasonalized demand
provided by periods l+1 to (l + p) and l+2 to
(l+ p+1) we obtain deseasonalized demand for
period l+1+(p/2) .
 i  t 1 (p / 2)

D t   D t p / 2  D t  p / 2   2Di  / 2p if p is even
 i  t 1 (p / 2) 
t   p / 2 


i  t   p / 2 
Di / p, for p odd

Contd...
For example, in the case of Tahoe salt
where p=4, for t=3 the decentralized
demand is given by

  4
D3   D1  D5   2Di  / 8
 i2 
Deseasolized demand for Tahoe
Demand
Period Demand Dt Deseasonalized
demand
1 8,000
2 13,000
3 23,000 19,750
4 34,000 20,625
5 10,000 21,250
6 18,000 21,750
7 23,000 22,500
8 38,000 22,125
9 12,000 22,625
10 13,000 24,125
11 32,000
12 41,000
• Once the demand is deseasonalized it is
either growing or declining at a steady
rate. It can be expressed as follows:

D t  L  Tt
Where,
L= level or deseasonalized demand at period t.
T=rate of growth of deseasonalized demand or trend
 L and T estimation
• For previous formula, need to estimate the
values of L and T.

• Use linear regression with deseasonalized

demand( using excel sheet).

• For the example of salt, L=18439, T=524

D t  18, 439  524t
 Deseasonalized demand and
seasonal factor for Tahoe Salt
Period Demand Dt Deseasonali Seasonal
zed demand Factor
1 8,000 18,963 0.42
2 13,000 19,487 0.67
3 23,000 20,011 1.15
4 34,000 20,535 1.66
5 10,000 21,059 0.47
6 18,000 21,583 0.83
7 23,000 22,107 1.04
8 38,000 22,631 1.68
9 12,000 23,155 0.52
10 13,000 23,679 0.55
11 32,000 24,203 1.32
12 41,000 24,727 1.66
Estimating seasonal factors
_

• The seasonal factor St for period t is the ratio of actual

demand to deseasonalized demand and is given as follows:

St  D t / D t

• For a given periodicity, p, we can obtain the

seasonal factor for a given period by averaging
seasonal factors corresponding to similar periods.

• Example ,if periodicity p=4, the periods 1,5,and 9

will have similar seasonal factors. The seasonal
factor for these periods is obtained as the average
of the 3 seasonal factors.
Estimating seasonal factors

• Given r seasonal cycles in data, for all periods of

the form pt  i,1  i  p we obtain the seasonal
factor as:
 r 1 
Si    S jp i  / r
 j0 

• For the Tahoe salt example, a total of 12 periods and

periodicity of 4 implies r=3 therefore we get
_ _ _
S1  ( S1  S5  S9 ) / 3  (0.42  0.47  0.52) / 3  0.47
Adaptive forecasting
• In this, the estimates of level, trend,
and seasonality are updated after each
demand observation.
• In most general setting, systematic
component of demand data contains a
level, a trend, and a seasonal factor.
• Can be easily modified for other cases
also.
• We have historical data for n periods
and demand is periodic with periodicity p
 Mathematical model
• In adaptive methods the forecast for period t+1is
given as follows:

Ft l  ( Lt  lTt ) St l
where
– Lt=estimate of level at the end of period t
– Tt=estimate of trend at the end of period t
– St=estimate of seasonal factor at the end of period t
– Dt= Actual demand observed at the end of period t
– Ft=forecast for demand at the end of period t
– Et=forecast for demand at the end of period t
 Steps in adaptive
forecasting framework
1. Initialize:
1. Compute initial estimates of level (L0),
trend (T0), seasonal factor (S1,…, SP)
from given data.

2. Forecast:
1. Given the estimates in period t,
forecast demand for t+1.
2. First forecast if for period 1 and is
made with the estimates of level,
trend, and seasonal factor at period 0.
Steps…
3. Estimate error

Record the actual demand Dt+1 for

period t+1 and compute the error Et+1
in the forecast for period t+1 as the
difference between the forecast
and actual demand. The error for
period t+1 is stated as follows:

Et+1= Ft+1- Dt+1

Steps…
4. Modifying estimates
o Modify estimates of level (Lt+1), trend (Tt+1), and
seasonal factor (St+p+1) for a given error Et+1 .
o Desirable modifications can be such that
• If the demand is lower than forecast, estimates
are revised downward.
• If demand is higher than forecast, estimates are
revised upward.

• Revised estimates in period t+1 are then

used to make forecast for period t+2.

• Steps 2, 3, 4 are repeated until all

historical data upto period n have been
covered.
 Moving Average
• Used when demand has no observable trend
or seasonality.
i.e. Systematic components of demand=level
– Estimate level in period t is given as the
average demand over most recent N
periods.

Lt  ( Dt  Dt 1  ...  Dt  N 1 ) / N
Moving average…

• The current forecast of all future periods is

same and is based on current estimate of level.
The forecast is stated as follows:
Ft 1  Lt and Ft  n  Lt

• To compute new moving average, simply add the

latest observation and drop the oldest one.
Revised moving average serves as the next
forecast.
Lt 1  ( Dt 1  Dt  ...  Dt  N  2 ) / N , Ft  2  Lt 1
Simple exponential smoothing
• Appropriate when demand has no
observable trend or seasonality.
i.e. Systematic component=level
• Initial estimate of level, L0, is taken to be
the average of all historical data.
1 n
L0   Di
n i 1

• The current forecast for all the future

periods is equal to the current estimate
of level is as follows Ft 1  Lt and Ft  n  Lt
Simple exponential smoothing
• After observing demand Dt+1 for period t+1,
estimate the level as
Lt 1   Dt 1  (1   ) Lt
 is the smoothing constant for the level,0< <1
• Revised value of level is a weighted average of
observed value of level Dt+1 and old estimates of
level (Lt) in period t.
1 t 1
L t 1   ()(1  ) n D t 1 n
n i 0

• Higher value of  corresponds to a forecast that

is more responsive to recent observation and vice-
versa.
Trend corrected exponential
smoothing (Holt’s method)
• Appropriate when demand is assumed to have a
level and a trend in systematic component but
no seasonality.
i.e. Systematic component of demand=level + trend
– Initial estimation of level and trend by
running a linear regression between demand
Dt and time period t
Dt=at+b
where ‘b’ mesures the estimate of the
demand at period t=0 and is our estimate of
initial level L0and
slope ‘a’ measure the rate of change in
demand per period and is our intial estimate
of trend T0.
 Holt’s model
• Running a linear regression
between demand and time periods
is appropriate since demand has a
trend but no seasonality.

• Forecast for future periods is

Ft+1=Lt+Tt and Ft+n=Lt +nTt
Holt’s Method…
• After observing the demand for
period t, we estimate the level and
trend as follows

Lt+1=αDt+1+(1-α)(Lt +Tt)
Tt+1=β(Lt+1-Lt)+(1- β)Tt

α is the smoothing constant for the

level 0<α<1

β is the smoothing constant for the

trend 0< β <1
Trend and seasonality exponential
smoothing (winter’s model)

• Appropriate when systematic

component of demand assumed to be
have a level, a trend, seasonal factor.
• Systematic component of
demand=(level + trend) X seasonal
factor
– Initiate with the estimation of level and
trend, and seasonal factor.
– Forecast for future periods
Ft+1=(Lt +Tt)St and Ft+1= (Lt +lTt)St+1
Winters model

• Lt+1=α(Dt+1/St+1)St +(1-α) (Lt+Tt)

• Tt+1= β(Lt+1 - Lt)St +(1-β)Tt

• St+p+1= γ(Dt+1/Lt+1)St +(1- γ)St+1

α is the smoothing constant for the level 0<α<1

β is the smoothing constant for the trend 0<β<1
γ is the smoothing constant for the level 0< γ <1
 Forecast error
• Every demand has a random component. A good
forecast method should capture the systematic
component of the demand but not the random
component. The random component manifests
itself in the form of forecast error.
• Reasons for the error analysis of forecast.
– Use error analysis to determine Whether the current
forecasting method is accurately predicting the systematic
component of demand. E.g. If a forecasting method continues
to give positive error appropriate measures can be taken by
the manager.
– Estimate forecast error because any contingency plan must
account for such an error
Measures of forecast error
• Observed error are within historical error estimates,
firms can continue to use their current forecasting
method.

• In the other case, finding may indicate forecasting

method is no more appropriate.

• Forecast error Et for tth time period is given as the

difference between the forecast for period t and actual
demand.
Et 1  Ft 1  Dt 1
 Mean Squared Error
(MSE)

• One measure of forecast error is the mean

squared error (MSE) and is given by :

1 n 2
MSE n   E t
n i 1
Absolute deviation
• Absolute deviation is the absolute value of
error in period t
At  Et

• Mean absolute deviation (MAD) to be average

of absolute deviation over all periods
1 n
MAD n   A t
n t 1

• MAD can be used to estimate the SD of

random component assuming it to be normally
distributed.
Mean percentage of
error
• Is the average absolute error as a
percentage of demand
n
Et

t 1 Dt
100
MAPE n 
n

• We can use the sum of forecast errors

to evaluate the bias bias  n E
n 
t 1
t

• The bias will fluctuate around 0 if the

error is truly random and not biased one
way or the other.
 Tracking signal

• The tracking signal is the ratio of the

bias and the MAD and is given as
follows:
biast
TSt 
MAD t
 Tracking signal

• If the TS at any period is outside the range

+6, this signal that the forecast is biased and
is either under-forecasting (TS below -6) or
over forecasting (TS below -6) .

• In such a case, a firm will choose a new fore

casting method.
Specialty Packaging Corporation: A Case Study

Company Profile
• Manufacturer of disposable containers
• Major customers are from the food industry
• Main raw material is Polystyrene resin
• Manages inventory through a make-to-stock
policy
 Specialty Packaging Corporation: A Case Study

Manufacturing Process
• Polystyrene is stored in the form of resin pallets
• Extruder generates rolled sheets, which may be
stored or further processed
• Thermo-setting press trims the rolls into
containers

Thermo-
Resin Extruder Roll Setting
Storage Storage Press

Fig. Manufacturing Process at SPC

 Specialty Packaging Corporation: A Case Study

Market Scenario
• Steady growth in demand, which will stabilize
after 2005
• Unable to meet the peak demand, extrusion
becomes a bottleneck
• Lost sales occur frequently

Plastic
Plastic
Peak demand in summer
Peak demand in fall

Clear
Clear Black
Black

Grocery
Grocery Grocery
Grocery
Bakery
Bakery Restaurants
Restaurants Catering
Catering
Store
Store Store
Store

Fig. Customer Base

Specialty Packaging Corporation: A Case Study

Goal
Synergizing marketing and customer feedbacks to
improve supply chain performance by adequate demand
matching
Objective
Forecasting quarterly demand during 2003-2005 for
both types of containers
Results
• Method of forecasting
• Likely forecast errors