Demand Forecasting

Content

The role of forecasting in a supply chain

Characteristics of forecasts
Components of forecasts and forecasting methods
Basic approach to demand forecasting
Time series forecasting methods
Measures of forecast error
Forecasting demand at Tahoe Salt
Demand forecasting in Indian Retail Industry

Role of forecasting
The basis for all strategic and planning decisions in a
supply chain
Used for both push and pull processes
Examples:
Production: scheduling, inventory, aggregate
planning
Marketing: sales force allocation, promotions, new
production introduction
Finance: plant/equipment investment, budgetary
planning
Personnel: workforce planning, hiring, layoffs
All of these decisions are interrelated

Characteristics of forecasting
Forecasts are always wrong. Should include expected
value and measure of error.
Long-term forecasts are less accurate than short-term
forecasts (forecast horizon is important)
Aggregate forecasts are more accurate than
disaggregate forecasts

Forecasting methods
Qualitative: primarily subjective; rely on judgment and
opinion
Time Series: use historical demand only
Static
Adaptive
Causal: use the relationship between demand and some
other factor to develop forecast
Simulation

Imitate consumer choices that give rise to demand

Can combine time series and causal methods

Components of a demand
O=S+R+E
O= Observed demand
S = Systematic component: expected value of demand
R = Random component: part of the forecast that deviates from
the systematic component
E = Forecast Error: difference between forecast and actual
demand
Systematic component

Level

Trend

Seasonality

(current deseasonalized demand) (growth or decline in the demand) (predictable seasonable fluctuation)

Basic approach for forecasting

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

Time Series Forecasting Methods

Goal is to predict systematic component of demand

Multiplicative: (level)(trend)(seasonal factor)

Additive: level + trend + seasonal factor
Mixed: (level + trend)(seasonal factor)
Static methods
Adaptive forecasting

Static Methods
Assume a mixed model:
Systematic component = (level + trend)(seasonal factor)
Ft+l = [L + (t + l)T]St+l
= forecast in period t for demand in period t + l
L = estimate of level for period 0
T = estimate of trend
St = estimate of seasonal factor for period t
Dt = actual demand in period t
Ft = forecast of demand in period t

7-9

Static Methods

Estimating level and trend

Estimating seasonal factors

7-10

Estimating Level and Trend

Before estimating level and trend, demand data must be
deseasonalized
Deseasonalized demand = demand that would have
been observed in the absence of seasonal fluctuations
Periodicity (p)

the number of periods after which the seasonal cycle

repeats itself

for demand at Tahoe Salt p = 4

Time Series Forecasting

Quarter
II, 1998
III, 1998
IV, 1998
I, 1999
II, 1999
III, 1999
IV, 1999
I, 2000
II, 2000
III, 2000
IV, 2000
I, 2001

Demand Dt
8000
13000
23000
34000
10000
18000
23000
38000
12000
13000
32000
41000

Forecast demand for the

next four quarters.

Time Series Forecasting

(Figure 7.1)
50,000
40,000
30,000
20,000
10,000
0

Deseasonalizing Demand

[Dt-(p/2) + Dt+(p/2) + S 2Di] / 2p

for p even (sum is from i = t+1-(p/2) to t+1+(p/2))

Dt =

S Di / p
for p odd (sum is from i = t-(p/2) to t+(p/2)), p/2 truncated to
lower integer

Deseasonalizing Demand
For the example, p = 4 is even
For t = 3:
D3 = {D1 + D5 + Sum(i=2 to 4) [2Di]}/8
= {8000+10000+[(2)(13000)+(2)(23000)+(2)(34000)]}/8
= 19750
D4 = {D2 + D6 + Sum(i=3 to 5) [2Di]}/8
= {13000+18000+[(2)(23000)+(2)(34000)+(2)(10000)]/8
= 20625
7-15

Deseasonalizing Demand
Then include trend
Dt = L + tT
Where
Dt = deseasonalized demand in period t
L = level (deseasonalized demand at period 0)
T = trend (rate of growth of deseasonalized demand)
Trend is determined by linear regression using deseasonalized
demand as the dependent variable and period as the
independent variable
In the example, L = 18,439 and T = 524
7-16

Time Series of Demand

50000

Demand

40000
30000

Dt
Dt-bar

20000
10000
0
1 2 3 4 5 6 7 8 9 10 11 12
Period

7-17

Estimating Seasonal Factors

Use the previous equation to calculate deseasonalized
demand for each period
St = Dt / Dt = seasonal factor for period t
In the example,
D2 = 18439 + (524)(2) = 19487

D2 = 13000

S2 = 13000/19487 = 0.67
The seasonal factors for the other periods are calculated
in the same manner
7-18

Estimating Seasonal Factors

t
1
2
3
4
5
6
7
8
9
10
11
12

Dt Dt-bar S-bar
8000 18963 0.42 = 8000/18963
13000 19487 0.67 = 13000/19487
23000 20011 1.15 = 23000/20011
34000 20535 1.66 = 34000/20535
10000 21059 0.47 = 10000/21059
18000 21583 0.83 = 18000/21583
23000 22107 1.04 = 23000/22107
38000 22631 1.68 = 38000/22631
12000 23155 0.52 = 12000/23155
13000 23679 0.55 = 13000/23679
32000 24203 1.32 = 32000/24203
41000 24727 1.66 = 41000/24727
7-19

Estimating Seasonal Factors

The overall seasonal factor for a season is then obtained
by averaging all of the factors for a season
If there are r seasonal cycles, for all periods of the form pt+i,
1<i<p, the seasonal factor for season i is
Si = [Sum(j=0 to r-1) Sjp+i]/r
In the example, there are 3 seasonal cycles in the data and
p=4, so
S1 = (0.42+0.47+0.52)/3 = 0.47
S2 = (0.67+0.83+0.55)/3 = 0.68
S3 = (1.15+1.04+1.32)/3 = 1.17
S4 = (1.66+1.68+1.66)/3 = 1.67
7-20

Estimating the Forecast

Using the original equation, we can forecast the next four
periods of demand:

F13 = (L+13T)S1 = [18439+(13)(524)](0.47) = 11868

F14 = (L+14T)S2 = [18439+(14)(524)](0.68) = 17527
F15 = (L+15T)S3 = [18439+(15)(524)](1.17) = 30770
F16 = (L+16T)S4 = [18439+(16)(524)](1.67) = 44794

7-21

Adaptive Forecasting
The estimates of level, trend, and seasonality are
adjusted after each demand observation
General steps in adaptive forecasting
Moving average
Simple exponential smoothing
Trend-corrected exponential smoothing (Holts model)
Trend- and seasonality-corrected exponential smoothing
(Winters model)
7-22

Basic Formula for Adaptive Forecasting

Ft+1 = (Lt + lT)St+1 = forecast for period t+l in period t
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 for period t
Ft = Forecast of demand for period t (made period t-1 or
earlier)
Dt = Actual demand observed in period t
Et = Forecast error in period t
At = Absolute deviation for period t = |Et|
MAD = Mean Absolute Deviation = average value of At

General Steps in Adaptive Forecasting

1. Initialize: Compute initial estimates of level (L 0), trend (T0),
and seasonal factors (S1,,Sp). This is done as in static
forecasting.
2. Forecast: Forecast demand for period t+1 using the
general equation
3. Estimate error: Compute error Et+1 = Ft+1- Dt+1
4. Modify estimates: Modify the estimates of level (L t+1),
trend (Tt+1), and seasonal factor (St+p+1), given the error
Et+1 in the forecast
Repeat steps 2, 3, and 4 for each subsequent period

Moving Average
Used when demand has no observable trend or
seasonality
Systematic component of demand = level
The level in period t is the average demand over the last N
periods (the N-period moving average)
Current forecast for all future periods is the same and is
based on the current estimate of the level
Lt = (Dt + Dt-1 + + Dt-N+1) / N
Ft+1 = Lt and Ft+n = Lt
After observing the demand for period t+1, revise the
estimates as follows:
Lt+1 = (Dt+1 + Dt + + Dt-N+2) / N
Ft+2 = Lt+1

Moving Average Example

From Tahoe Salt example (Table 7.1)
At the end of period 4, what is the forecast demand for periods
5 through 8 using a 4-period moving average?
L4 = (D4+D3+D2+D1)/4 = (34000+23000+13000+8000)/4 =
19500
F5 = 19500 = F6 = F7 = F8
Observe demand in period 5 to be D5 = 10000
Forecast error in period 5, E5 = F5 - D5 = 19500 - 10000 =
9500
Revise estimate of level in period 5:
L5 = (D5+D4+D3+D2)/4 = (10000+34000+23000+13000)/4 =
20000
F6 = L5 = 20000

Simple Exponential Smoothing

Used when demand has no observable trend or seasonality
Systematic component of demand = level
Initial estimate of level, L0, assumed to be the average of all
historical data
L0 = [Sum(i=1 to n)Di]/n
Current forecast for all future periods is equal to the current
estimate of the level and is given as follows:
Ft+1 = Lt and Ft+n = Lt
After observing demand Dt+1, revise the estimate of the
level:
Lt+1 = aDt+1 + (1-a)Lt
Lt+1 = Sum(n=0 to t+1)[a(1-a)nDt+1-n ]

Simple Exponential Smoothing Example

From Tahoe Salt data, forecast demand for period 1 using
exponential smoothing
L0 = average of all 12 periods of data
= Sum(i=1 to 12)[Di]/12 = 22083
F1 = L0 = 22083
Observed demand for period 1 = D1 = 8000
Forecast error for period 1, E1, is as follows:
E1 = F1 - D1 = 22083 - 8000 = 14083
Assuming a = 0.1, revised estimate of level for period 1:
L1 = aD1 + (1-a)L0 = (0.1)(8000) + (0.9)(22083) = 20675
F2 = L1 = 20675
Note that the estimate of level for period 1 is lower than in
period 0

Trend-Corrected Exponential Smoothing (Holts Model)

Appropriate when the demand is assumed to have a level
and trend in the systematic component of demand but no
seasonality
Obtain initial estimate of level and trend by running a linear
regression of the following form:
Dt = at + b
T0 = a
L0 = b
In period t, the forecast for future periods is expressed as
follows:
Ft+1 = Lt + Tt
Ft+n = Lt + nTt
7-29

Trend-Corrected Exponential Smoothing (Holts Model)

After observing demand for period t, revise the estimates for
level and trend as follows:
Lt+1 = aDt+1 + (1-a)(Lt + Tt)
Tt+1 = b(Lt+1 - Lt) + (1-b)Tt
a = smoothing constant for level
b = smoothing constant for trend
Example: Tahoe Salt demand data. Forecast demand for
period 1 using Holts model (trend corrected exponential
smoothing)
Using linear regression,
L0 = 12015 (linear intercept)
T0 = 1549 (linear slope)
7-30

Holts Model Example (continued)

Forecast for period 1:
F1 = L0 + T0 = 12015 + 1549 = 13564
Observed demand for period 1 = D1 = 8000
E1 = F1 - D1 = 13564 - 8000 = 5564
Assume a = 0.1, b = 0.2
L1 = aD1 + (1-a)(L0+T0) = (0.1)(8000) + (0.9)(13564) =
13008
T1 = b(L1 - L0) + (1-b)T0 = (0.2)(13008 - 12015) + (0.8)
(1549)
= 1438
F2 = L1 + T1 = 13008 + 1438 = 14446
F5 = L1 + 4T1 = 13008 + (4)(1438) = 18760
7-31

Trend- and Seasonality-Corrected Exponential

Smoothing
Appropriate when the systematic component of demand is
assumed to have a level, trend, and seasonal factor
Systematic component = (level+trend)(seasonal factor)
Assume periodicity p
Obtain initial estimates of level (L0), trend (T0), seasonal
factors (S1,,Sp) using procedure for static forecasting
In period t, the forecast for future periods is given by:
Ft+1 = (Lt+Tt)(St+1) and Ft+n = (Lt + nTt)St+n
7-32

Trend- and Seasonality-Corrected Exponential

Smoothing (continued)
After observing demand for period t+1, revise estimates for
level, trend, and seasonal factors as follows:
Lt+1 = a(Dt+1/St+1) + (1-a)(Lt+Tt)
Tt+1 = b(Lt+1 - Lt) + (1-b)Tt
St+p+1 = g(Dt+1/Lt+1) + (1-g)St+1
a = smoothing constant for level
b = smoothing constant for trend
g = smoothing constant for seasonal factor
Example: Tahoe Salt data. Forecast demand for period 1
using Winters model.
Initial estimates of level, trend, and seasonal factors are
obtained as in the static forecasting case
7-33

Trend- and Seasonality-Corrected Exponential Smoothing

Example
L0 = 18439 T0 = 524

S1=0.47, S2=0.68, S3=1.17, S4=1.67

F1 = (L0 + T0)S1 = (18439+524)(0.47) = 8913

The observed demand for period 1 = D1 = 8000
Forecast error for period 1 = E1 = F1-D1 = 8913 - 8000 = 913
Assume a = 0.1, b=0.2, g=0.1; revise estimates for level and
trend for period 1 and for seasonal factor for period 5
L1 = a(D1/S1)+(1-a)(L0+T0) = (0.1)(8000/0.47)+(0.9)
(18439+524)=18769
T1 = b(L1-L0)+(1-b)T0 = (0.2)(18769-18439)+(0.8)(524) = 485
S5 = g(D1/L1)+(1-g)S1 = (0.1)(8000/18769)+(0.9)(0.47) = 0.47
F2 = (L1+T1)S2 = (18769 + 485)(0.68) = 13093
7-34

Measures of Forecast Error

Forecast error = Et = Ft - Dt
Mean squared error (MSE)
MSEn = (Sum(t=1 to n)[Et2])/n
Absolute deviation = At = |Et|
Mean absolute deviation (MAD)
MADn = (Sum(t=1 to n)[At])/n
s = 1.25MAD

7-35

Measures of Forecast Error

Mean absolute percentage error (MAPE)

MAPEn = (Sum(t=1 to n)[|Et/ Dt|100])/n
Bias
Shows whether the forecast consistently under- or
overestimates demand; should fluctuate around 0
biasn = Sum(t=1 to n)[Et]
Tracking signal
Should be within the range of +6
Otherwise, possibly use a new forecasting method
TSt = bias / MADt

Forecasting Demand at Tahoe Salt

Moving average
Simple exponential smoothing
Trend-corrected exponential smoothing
Trend- and seasonality-corrected exponential
smoothing

7-37

Indian Retail Industry

Retail - The sale of goods or commodities in small
quantities directly to consumers.
Retailers in India

Shoppers Stop
Westside (Trent)
Pantaloon (Big Bazaar)
Lifestyle
RPG Retail (Foodworld, Musicworld)
Crossword

What are the key aspects of a forecasting exercise?

1. Beginning Early
2. Taking expert inputs
3. The tools used for demand forecasting cannot replace
domain expertise
4. Thinking of multiple solutions
5. Being on the field

Questionnaire
How is the forecasting done?
Past data and Targeted growth
Tastes and trends are subjective terms

What are some tips for developing a forecasting model for new
product?
try out different forecasting techniques

What is the period of time in the future for which you do forecasting?
a six monthly basis, place orders at 2 months in advance

Could you share with us an example story?

Questionnaire
Does the volume of sale have a bearing on the rack
space you allocate to a good?
They have a fixed rack space and that dictates how much they can stock.

For all the analysis do you use software?

Could you relate a failure story?

Conclusion
The increasing competition in Indian Retail Sector is
compelling retailers to use demand forecasting tools.
Retailers who have their own brand labels use the
forecasting techniques of the kind we study in theory.
On the other hand, small scale retailers can employ
qualitative techniques on the historical data and
considering the behavior trends in the market.

References
Supply Chain Management by Sunil Chopra, Peter
Meindl, Dharam Vir Kalra
Images retail. http://www.imagesretail.com/india retail
report.htm.
www.sciencedirect.com