Demand Forecasting

Section Objectives
After completing this section, you should be able to:

1. List the features of a good forecast.

2. Outline the steps in the forecasting process.
3. Compare and contrast qualitative and quantitative approaches to
forecasting.
4. Identify three qualitative forecasting methods.
5. Briefly describe averaging techniques, trend and seasonal techniques
and regression analysis, and solve typical problems.
6. Describe two measures of forecast accuracy.
7. Describe two ways of evaluating and controlling forecasts.

5-1

Demand Forecasting

Features of Forecasts
1. Causal System. Forecast techniques generally assume that the

same underlying causal system that existed in the past will

continue to exist in the future.
2. Forecast Error. Forecasts are rarely perfect; actual results usually
differ from predicted values.
3. Group Forecasts. Forecasts for groups of items tend to be more
accurate than forecasts for individual items because forecasting
errors among items in a group usually have a canceling effect.
4. Accuracy and Time. Forecast accuracy decreases as the time
period covered by the forecast (i.e. the time horizon) increases.
Generally, short-term forecasts must deal with fewer uncertainties
than long-term forecasts.

5-2

Demand Forecasting

Forecast Process
The process of forecasting has four clearly definable steps:
1. Determine the purpose of the forecast. The use to which the forecast will
be used will determine both the technique to be used and the frequency
with which the forecast has to be updated.
2. Establish a time horizon. How far forward are we interested in
forecasting? Next week? Next month? Next year? Next 20 years? The
choice of horizon affects the choice of technique and this, in turn,
determines the amount of data and effort needed to prepare the forecast.
3. Prepare the forecast. This involves four steps:

a. Identify the assumptions in the forecast model you propose to use.

b. Gather the data.
c. Analyze the data.
d. Forecast.
4. Monitor the results. It is necessary to monitor forecast results to
determine whether certain underlying factors in the model have
undergone change. Has the trend weakened? Strengthened? Is the
seasonal variation the same as in prior periods?
5-3

Demand Forecasting

Types of Forecasts
1. Qualitative - consists mainly of subjective inputs such as human
factors, personal opinions or hunches which may be difficult or
impossible to quantify.
2. Quantitative - involve the extension of historical data or
development of associative models.
Time Series - extension of historical data by identifying
patterns in the past that might reasonably be expected to
continue in the future.
Causal models - development of an association between the
variable we are interested in forecasting and one or more
variables that might explain the variable of interest.

5-4

Demand Forecasting

Classification of Forecasting Models

Forecasting
Statistical
Time Series
Models
Trend
Projection

Trend and
Seasonal

Causal or
Regression Models

Weighted Moving
Averages

Expert
Opinion

Market
Surveys

Delphi
Method

Smoothing
Methods

Moving
Averages

Simple Moving
Averages

Judgemental

Exponential
Smoothing

Simple Exponential
Smoothing

Exponential Smoothing
with Trend

Exponential Smoothing
with Seasonal Variation

Exponential Smoothing
with Trend and Seasonal

5-5

Demand Forecasting

Operations Forecasting: Uses and Methods

Uses of
Forecasting for
Operations

Time
Horizon

Accuracy
Required

Management
Level

Forecasting
Methods

Location

Long

Medium

Top

Qualitative
and Causal

Long

Medium

Top

Qual. & Causal

Medium

High

Middle

Causal & T.S.

Scheduling / MRP

Short

Highest

Lower

Time Series

Order Processing

Short

Highest

Lower

Time Series

Capacity Planning:
Facilities
Equipment

5-6

Demand Forecasting

Qualitative Forecasting Methods

Executive Opinion. Forecasts that are based on the judgment and
experience of managers.
Sales Force Composite. Forecasts compiled from estimates of
demand made by members of a companys sales force
Consumer Surveys. A forecasting method that seeks input from
customers regarding future purchasing plans for existing products
or services.
Market Research. This method tests hypothesis about new products
or services or new markets for existing products or services.
Delphi Method. A forecasting technique using a group process that
allows experts to make forecasts.

5-7

Demand Forecasting

Quantitative Forecasting Methods

These can be broken into two main categories:
1. Time Series (TS) Models. A forecasting approach in which future
values of a series can be estimated from past values of the series.
Driving forward by looking at the rear view mirror. Types of TS
models include:

Simple Average / Moving Average / Weighted Moving Average

Exponential Smoothing: Single, Trend, Seasonal, and Trend

and Seasonal

Trend Projection

2. Associative (Causal) Models. A forecasting method which

identifies related variables that can be used to predict values of the
variable of interest. The essential element is the development of
an equation that summarizes the effects of predictor variables.
The primary method of analysis is known as regression.
5-8

Demand Forecasting

Time Series Models

A time series is a time-ordered sequence of observations taken at regular
intervals over a period of time. Analysis of a time series requires an
identification of the underlying behaviour of the series. This behaviour
may have four patterns:
1. Trend refers to a gradual, long-term movement in the data.
Population shifts, changing incomes and cultural changes often
account for such movements.
2. Seasonality refers to short-term, fairly regular variations that are
generally related to weather factors or to human-made factors such
as holidays.
3. Cycles are wavelike variations of more than one years duration.
These are often related to a variety of economic and political factors.
4. Irregular variations are due to unusual circumstances such as severe
weather conditions, strikes or a major change in a product or service.
They do not reflect typical behaviour, and they should be removed
from the data before any analysis is done.
5. Random variations are the residual variations that remain after al the
other behaviours have been accounted for.
5-9

Demand Forecasting
Application of Forecasting Methods
Combination of Components
in the Series

Objectives

Models Often
Appropriate

No trend (horizontal trend), no seasonal variation;

i.e. a stable variation with random fluctuation

To average out
randomness and
find average

Simple moving average

Weighted moving average
Single exponential smoothing
Simple moving average

No trend, but seasonal variation

To determine seasonal
pattern and project it or
average out seasonality

Time series decomposition

Short term projection

Double exponential smoothing

Long term projection

Time series decomposition

To project trend and

seasonal variation
around it

Triple exponential smoothing

To identify variables
that explain level
of demand

Simple linear regression

Curvilinear regression
Multiple regression

Trend, but no seasonal variation

Trend and seasonal variation

Patterns of changes not related to time

Time series decomposition

5 - 10

Demand Forecasting
2000

Seasonal Variations

2001
2002

700

2003
2004

600

2005
500
400
300
200
Jan

Feb

Mar

Apr

May

June

July

Aug

Sept

Oct

Nov

Dec

The data in the graph is monthly sales for a six-year period. Each year is
graphed on top of the preceding one.
Question: What time series patterns exist in this data?
5 - 11

Demand Forecasting

Time Series: Averaging Techniques

1. Naive Forecasts - a naive forecast for any period equals the previous
periods actual value. Although it appears simplistic, it is a legitimate
forecasting technique:
it has virtually no cost
forecasts are quick and easy to prepare
easy to understand
can be used for seasonal data (e.g. sales for this December equal
sales for preceding December)
2. Moving Average - a forecasting technique that use a number of the most
recent actual data values in generating a forecast. There are two types:
a. Simple moving average = SMA = Si Ai / n
where i = the age of the data
n = the number of periods in the moving average
Ai = actual value with age i
Note that each data value has the same importance (i.e. weight)
5 - 12

Demand Forecasting
b. Weighted moving average = WMA = Si Ai Wi
where Wi = the relative weight of each data point in the moving
average
Note that the sum of all weights , SWi , must equal 1.
For both the SMA and the WMA, a key issue is how many data points will
be used to calculate the average. A large number of data points results in
a smooth average: a small number of data points means the the model
responds very quickly to the most recent changes.
If responsiveness in important, a simple moving average with relatively
few data points, or a weighted moving average with a heavy weight on
recent data, should be used.
A decision maker must weigh the risk of responding quickly to what might
be random fluctuations in the data against the risk of responding slowly to
real changes.

5 - 13

Demand Forecasting
3. Exponential Smoothing - This is a special case of a weighted moving
average in which the weights are determined by mathematical formula,
rather than assigned by the decision maker.
Each new forecast is based on a percentage of the previous periods
demand and a percentage of the previous periods forecast. That is:
Ft + 1 = aDt + (1-a)Ft
where Ft+1 = forecast of the time series for period t + 1
Dt = actual value of the time series for period t
Ft = forecast value for the time series for period t
a = smoothing constant (0 1)

5 - 14

Demand Forecasting
Alpha () is a weighting factor with values between zero and one.
The sensitivity of forecast adjustments is determined by this
smoothing constant.
The closer is to zero, the slower the forecast will be to adjust to
forecast errors (i.e. the greater the smoothing). Conversely, the
closer the value of is to 1.00, the greater the sensitivity and the
less the smoothing.
Commonly used values of range from .05 to .50.

Impact of a Values on the Weight Attached to Observations in a Time Series

Values

Weight
Attached To

.1

.2

.3

.4

.5

Dt
Dt-1
Dt-2
Dt-3
Dt-4
Dt-5
Dt-6
Dt-7

.1000
.0900
.0810
.0729
.0656
.0590
.0531
.0478

.2000
.1600
.1280
.1024
.0819
.0655
.0524
.0419

.3000
.2100
.1470
.1029
.0720
.0504
.0353
.0247

.4000
.2400
.1440
.0864
.0518
.0311
.0187
.0112

.5000
.2500
.1250
.0625
.0313
.0156
.0078
.0039

5 - 15

Demand Forecasting
Simple Moving Average - Illustration
Compute a three-period simple moving average forecast given demand for gizmos for the last five periods:
Period
1
2
3
4
5

Age
5
4
3
2
1

Demand
42
40
43
40
41

Solution = Forecast for Period 6

MA3 = (43 + 40 + 41) / 3 = 41.33

If actual demand in period 6 is 39, the forecast for period 7 will be: MA3 = (40 + 41 + 39) / 3 = 40.00
Note that in a moving average, as each new actual value becomes available, the forecast is updated by
adding the newest value and dropping the oldest and then recomputing the average. Therefore, the
forecast moves by reflecting only the most recent values.
3-Period Moving Average

MA3
Demand

50
45
40
Demand
35
30
25
2

10

12

14

16

18

20

22

24

26

28

30

Period

5 - 16

Demand Forecasting

Exponential Smoothing Models

1. Simple Model - assumes the time series is flat with no trend or seasonality.
Ft + 1 = Dt + (1-)Ft
2. Exponential Smoothing for Trend - assumes the time series has a long term
linear trend. Trend may exhibit growth or decline.
At = Dt + (1-)(At-1 + Tt-1)
Tt = b(At - At-1) + (1 - b)Tt-1
Ft + 1 = At + Tt
3. Exponential Smoothing for Trend and Seasonal - assumes the time series
has both a long-term trend and seasonal variation. Seasonal variation
should occur at approximately the same time each year and be of the
same degree.
At = (Dt / It-L) + (1-)(At-1 + Tt-1)
Tt = b(At - At-1) + (1 - b)Tt-1
It = g(Dt/At) + (1-g)Rt-L
Ft + 1 = (At + Tt)(It-L + K )
5 - 17

Demand Forecasting
Simple Exponential Smoothing: An Illustration
t

Actual
Demand

Forecast
a = .1

Error
(Dt - Ft)

Error 2
(Dt - Ft)2

Forecast
a = .3

Error
(Dt - Ft)

Error 2
(Dt - Ft)2

1
2
3
4
5
6
7
8
9
10
11
12
13

170
210
190
230
180
160
200
180
220
200
180
190
200

170.0
170.0
174.0
175.6
181.0
180.9
178.8
181.0
180.9
184.8
186.1
185.7
186.1
187.5

0.0
40.0
16.0
54.4
-1.0
-20.9
21.2
-1.0
39.1
15.2
-6.1
4.3
13.9

0.0
1600.0
256.0
2959.4
1.1
438.3
447.6
0.9
1531.8
231.8
39.7
18.8
193.2

170.0
170.0
182.0
184.4
198.1
192.7
182.9
188.0
185.6
195.9
197.1
192.0
191.4
194.0

0.0
40.0
8.0
45.6
-18.1
-32.7
17.1
-8.0
34.4
4.1
-17.1
-2.0
8.6

0.0
1600.0
64.0
2079.4
326.9
1066.4
293.8
64.0
1183.3
16.6
293.9
4.0
73.9

Sum of the errors

= S(Dt-Ft) = 174.9

Sum of the errors

= S(Dt-Ft) = 79.9

Sum of the absolute errors

= S|Dt-Ft| = 233.4

Sum of the absolute errors

= S|Dt-Ft| = 235.7
5 - 18

Demand Forecasting

SIMPLE EXPONENTIAL SMOOTHING MODEL

ACTUAL vs. FORECAST
240

Original Demand

220

200

a = .3
a = .1

180

160

140

120
1

10

11

12

13

5 - 19

Demand Forecasting
Exponential Smoothing With Trend: An Illustration

t
0
1
2
3
4
5
6
7
8
9
10

Dt
(in '000s
of tons)
216.00
229.00
255.00
219.00
239.00
275.00
315.00
297.00
286.00
314.00

At
(average)

Tt
(trend)

Ft
(forecast)
At+Tt

205.00
216.00
227.40
241.75
246.30
253.39
265.98
284.22
295.84
302.95
313.91

11.00
11.00
11.04
11.37
10.69
10.33
10.55
11.32
11.35
10.93
10.93

216.00
227.00
238.44
253.12
256.99
263.72
276.53
295.55
307.19
313.88

Dt-Ft
(error)

|Dt-Ft|
(absolute
error)

0.00
2.00
16.56
-34.12
-17.99
11.28
38.47
1.45
-21.19
0.12

0.00
2.00
16.56
34.12
17.99
11.28
38.47
1.45
21.19
0.12

(Dt-Ft)^2
(squared
error)
0.00
4.00
274.23
1164.39
323.54
127.26
1479.97
2.11
449.07
0.01

A(t) + T(t) = F(t)

205 + 11 = 216
216 + 11 = 227
Sum of Forecast Errors
Sum of Absolute Forecast Errors
Sum of Squared Forecast Errors

- 3.42
143.18
3824.59

Assume A0 = 205; T0 = 11; = .2; b = .1

5 - 20

Demand Forecasting
TREND EXPONENTIAL SMOOTHING MODEL
D(t)

ACTUAL vs FORECAST

F(t)
350

300

Demand 250

200

150
1

10

Time Period

5 - 21

Demand Forecasting
Exponential Smoothing With Trend and Seasonal: An Illustration
t
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

Dt
(in
At
Tt
units) (average) (trend)
4800
4100
6000
6500
5800
5200
6800
7400
6000
5600
7500
7800
6300
5900
8000
8400

5500
5689
5889
6016
6123
6227
6393
6552
6639
6715
6832
6969
7080

It
(seasonal
ratio)

0
47
85
96
99
100
116
127
117
107
109
116
115

[A(t) + T(t)] x I(t-3) = F(t)

(5500 + 0) x 0.90 = 4950
(5689 + 47) x 0.80 = 4589

0.90
0.80
1.10
1.20
0.96
0.84
1.12
1.20
0.96
0.86
1.13
1.19
0.95
0.86
1.14
1.19

Ft
(forecast)
[At+Tt]*It-L+K

4950
4589
6572
7334
5971
5324
7259
8044
6497
5858
7843
8428
6835
6296
8457
8958

Dt-Ft
(error)

|Dt-Ft|
(absolute
error)

(Dt-Ft)^2
(squared
error)

850
611
228
66
29
276
241
-244
-197
42
157
-28

850
611
228
66
29
276
241
244
197
42
157
28

722500
373457
52104
4367
849
75911
58115
59764
38811
1727
24772
804

2030

2970

1413181

Assume A0 = 5500; T0 = 0; L = 4; I0 = 1.20; I-1 = 1.10; I-2 = 0.80; I-3= 0.90; a = .20; b = .25; g = .50

5 - 22

Demand Forecasting

TREND AND SEASONAL EXPONENTIAL SMOOTHING MODEL

ACTUAL vs FORECAST
D(t)
10000

F(t)

9000
8000
7000

Orders
6000
5000
4000
3000
1

10

11

12

13

14

15

16

17

18

19

20

Time periods

5 - 23

Demand Forecasting

Trend Projection: An Alternative to Exponential Smoothing

Whazzit? A method of taking time series data and separating (decomposing) it into
one or more components of trend, seasonal, cyclical, and random variation. Once
the data has been decomposed, we can estimate the values of the individual
components and use these estimates to predict future values of the time series.

Steps:
1. Calculate an annual moving average.
2. Centre the moving average.
3. Divide the centered moving average into the demand values. This is the
seasonal-random component.
4. Average the seasonal-random component for the same time period in
successive years. This average is the seasonal factor for the time period.
5. Divide each actual demand value by its seasonal factor. This produces
deseasonalized demand.
6. Regress deseasonalized demand against time and calculate the trend value
and the constant term.
7. Develop a trend forecast.
8. Multiply the trend forecast by the seasonal factor. This is the actual forecast.

5 - 24

Demand Forecasting
Trend Projection: An Illustration

Year
Year 1

Year 2

Year 3

Year 4

Time
Period

Sales

1
2
3
4
1
2
3
4
1
2
3
4
1
2
3

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

4800
4100
6000
6500
5800
5200
6800
7400
6000
5600
7500
7800
6300
5900
8000

16

8400

Quarter

Centered Seasonal
Moving
Moving
Random Seasonal
Average Average Component Factor
5350
5600
5875
6075
6300
6350
6450
6625
6725
6800
6875
7000
7150

(Step 1 )

5475
5738
5975
6188
6325
6400
6538
6675
6763
6838
6938
7075

( Step 2 )

1.096
1.133
0.971
0.840
1.075
1.156
0.918
0.839
1.109
1.141
0.908
0.834

( Step 3 )

Deseasonalized
Sales

0.932
0.838
1.093
1.143
0.932
0.838
1.093
1.143
0.932
0.838
1.093
1.143
0.932
0.838
1.093

5149
4894
5488
5685
6222
6207
6219
6472
6436
6684
6860
6822
6758
7043
7317

1.143

7347

( Step 4 )

( Step 5 )
5 - 25

Demand Forecasting
Regression Output:

Trend Forecast:

Quarterly Forecast:

Constant = 5099.5
Std Err of Est = 212.6531
R Squared = 0.920804
No. of Observations = 16
Degrees of Freedom = 14

T(17) = 5100 + 147(17) = 7601

F(17) = 7601 x .932 = 7084

T(18) = 5100 + 147(18) = 7748

F(18) = 7748 x .838 = 6493

T(19) = 5100 + 147(19) = 7895

F(19) = 7895 x 1.093 = 8629

X Coefficient(s)
Std Err of Coef.

T(20) = 5100 + 147(20) = 8042

F(20) = 8042 x 1.143 = 9192

( Step 6 )

147.1397
11.53273

( Step 7 )

( Step 8 )

5 - 26

Demand Forecasting
TREND PROJECTION MODEL
DESEASONALIZED DATA
Sales
Deseasonalized Sales
9000

8000

7000

Quantity 6000
5000

4000

3000
1

10

11

12

13

14

15

16

Time Periods

5 - 27

Demand Forecasting
TREND PROJECTION MODEL
Actual

ACTUAL vs FORECAST

Forecast
10000

9000

8000

7000

Quantity
6000

5000

4000

3000
1

11

13

15

17

19

Quarters

5 - 28

Demand Forecasting
Demand Forecasting- Additional Illustration # 1
National Mixer Inc. sells can openers. Monthly sales for a seven-month period were as follows:

Month
Feb
Mar
Apr
May
Jun
Jul
Aug

Sales
20
18
15
20
18
22
20

a. Plot the monthly data on a sheet of graph paper.

b. Forecast September sales volume using each of the following:

(1) A linear trend equation.
(2) A five-month moving average.
(3) Exponential smoothing with a smoothing constant (a) equal to .20, and a March forecast of 19.
(4) The naive approach
c. Which method seems least appropriate? Why?
d. What does the use of the term sales rather than demand presume?

5 - 29

Demand Forecasting
Demand Forecasting- Additional Illustration # 2
a. Develop a linear trend equation for the following data on freight car loadings, and use it to predict
loadings for periods 11 through 14.

Year
1
2
3
4
5
6

Number(00)
220
245
280
275
300
310

Year
7
8
9
10
11
12

Number(00)
350
360
400
380
420
450

b. Use trend-adjusted exponential smoothing with a = .3 and b = .2 to smooth the data. Forecast periods
11 through 14.

5 - 30