1/13/2014

MGMT 2012: Introduction to Quantitative Methods

FORECASTING METHODS
Lecture 11

Lecturer : Dr. Dwayne Devonish

Learning Objectives
Students should be able to:

Define the concept of a forecast and
forecast error

Calculate different measures of forecast
error

Apply, evaluate and compare different
methods of forecasting in QM based on
time series data.

Introduction to Forecasting

Forecasting Benefits

A forecast is a statement about the future

value of a particular variable (e.g. demand,
sales, etc).
Forecasts provide managers with essential
information about the future, even in the
midst of uncertainty.
Managers must rely on forecasts in QM to
help them make more informed decisions
about meeting future demand.

Forecasting helps managers:

Know what capacity will be needed to make
staffing and equipment decisions,

Prepare sound budgets for future decisions,

Determine optimal production and inventory
levels

Anticipate, plan, and respond to future changes
or events
However, managers or quantitative analysts
must be able to select accurate forecasting
methods in order to enjoy desired benefits of
these tools.

Forecast Accuracy
Selecting a forecasting method depends on the
accuracy of the method compared to other methods.
Forecasters need to choose methods which generate
the least amount of error - this represents the
methods which are most accurate (better
representation of actual data)
Forecast error is simply the difference between actual
data values or demand and forecast values.
Forecast Error = Actual (A) - Forecast (F).
If an actual demand value was 19 and a
corresponding forecast value in the same period
was 23, the forecast error is - 4.
Forecast error can be converted to an absolute
error which is simply the removal of the negative
sign from the value thus creating an absolute
(non-negative number) which is 4.
This is called absolute error or [error].

Measures of Forecast Accuracy

Measures of forecast accuracy include:
 Mean Absolute Deviation (MAD)
= |absolute forecast errors|
n

Mean Squared Error (MSE)

= (errors) 2
n

MAD = add absolute errors for periods and divide by

number of available periods used in the forecast.
MSE = square each error value for each period, then
sum all squared error values and divide by no. of
available periods.

1/13/2014

CALCULATING MAD AND MSE IN TIME

SERIES EXAMPLE
JAN
FEB
MAR
APR
MAY
JUN

Actual
243
315
286
256
241
298

Forecast
250
320
275
260
250
275

|error|

error2

7
5

49
25

11
4
9
23

121
16
81
529

MAD =
(9.83)
MSE =

(136.8)

Forecasting Methods
There are three categories of forecasting
methods:
 Time series methods: Highlight patterns in
historical data and then extrapolate this
pattern into the future.
 Causal methods: Focus on the relationships
between variables to forecast future values.
 Qualitative methods: Focus on subjective
judgements and expert opinions to develop
relevant forecasts.

Time Series Methods

Nave Method

A time series is a time-ordered sequence of

observations
taken
at
regular
intervals
(Stevenson, 2009, p.78)
Time series can be expressed daily, monthly,
yearly, etc.
Specific time series forecasting methods include:

Nave method

Moving averages method

Exponential method

Linear trend projection method

This method uses a single previous

value of a time series as the basis of
the forecast.
It is useful for stable time series.
Most simple and straightforward
forecasting method available.

Yearly Time Series:Nave Method

Years

Actual (A) Demand

(No. of customer
orders)

Forecast (F)

2000
2001

5
9

2002

2003

12

2004

12

2005

10

2006

10

Forecast Accuracy for Nave Method

Yrs
2000
2001
2002
2003
2004
2005
2006
MAD
MSE

A.
5
9
6
12
7
10
8

F.
5
9
6
12
7
10

[Error]
[Error]2
MAD
MSE
4 SUM
16
SUM
AND
AND
3 AVERAGE 9 AVERAGE
6
36
5
25
3
9
2
4
3.83
16.5

[ERROR] is the absolute error which is the difference

between actual and forecast. Minus signs are removed.

1/13/2014

Moving Averages Method

This method consists of computing an average of
the most recent n data values in a time series
and using this average for the forecast of the next
period.
Moving averages methods use more historical
data to forecast future values. These methods are
referred to smoothing methods.
Good for stable time series and short-range
forecasts.
You can use different numbers of past periods
(years, months, etc) in the forecasts.
Consider a 3 year moving average on these
time series data:

Yearly Time Series: 3yr MA

Years

Actual (A) Demand

Forecast (F)
(No. of customer 3 Year Moving Average
orders)

2000
2001

5
9

2002

2003

12

(5 + 9 + 6)/3 = 6.67

2004

(9 +6 +12)/3 = 9

2005

10

(6 +12+7)/3 = 8.33

2006

(12+7+10)/3 =9.67

Forecast Accuracy for 3 yr MA

Yrs
2000
2001
2002
2003
2004
2005
2006
MAD
MSE

A.
5
9
6
12
7
10
8

F.
6.67
9
8.33
9.67

[Error]
5.33
2
1.67
1.67
2.67

[Error]2
28.44
4
2.78
2.78

LETS TRY A 2 YEAR MOVING

AVERAGE FORECAST
NB:
Whereas a 3 year moving average begins on the
4th year given that it uses data from the 3 most
recent years, the 2 year moving average forecast
begins on the 3rd year (it uses the 2 most recent
years)

9.50

[ERROR] is the absolute error which is the difference

between actual and forecast: Minus signs removed.

Forecast Accuracy for 2 yr MA

Yrs
2000
2001
2002
2003
2004
2005
2006
MAD
MSE

A.
5
9
6
12
7
10
8

F.
7
7.5
9
9.5
8.5

[Error]
1
4.5
2
0.5
0.5
1.70

[Error]2
1
20.25
4
.25
.25
5.15

[ERROR] is the absolute error which is the difference

between actual and forecast: Minus signs removed.

Exponential Smoothing
Another smoothing method which represents a
weighted average method.
It uses a weighted average of past data values as
a forecast.
Each new forecast is based on the previous
forecast plus a percentage of the difference
between the forecast and the actual value in that
previous period (i.e. percentage of forecast error
not absolute error).
Next Forecast = Previous forecast + (Previous
Actual Previous Forecast), where:
(Actual Forecast) = Error in last period, and is
percentage of error (i.e. smoothing constant).

1/13/2014

Smoothing Constant in Exponential

Smoothing

Exponential Smoothing

The smoothing constant provides the weight for the

next forecast. It is seen as the percentage of error
from last period forecast which you are planning to
incorporate into a future forecast.
It ranges from 0 to 1. Higher values make greater
adjustments to the next forecast based on past
forecast errors. That is, higher values suggest that
the forecast will be more sensitive to recent changes
to actual demand values.
If the constant was zero then,
Next Forecast = Previous forecast + 0 (A - F)
We consider no percentage of error from last forecast
period, so the next forecast equals previous forecast.

Commonly used smoothing constants range

from .05 to .50.
Lower values work well when the underlying
average (or series) tends to be stable, whereas
higher values work well when the underlying
average is susceptible to change (i.e. volatile).
Lets use exponential smoothing forecast, with a
smoothing constant of .30 (i.e. 30% of prior
forecast error is used for future forecasts).
NB: For the 1st period, no forecast value is
generated. However, the forecast value in the
2nd period in the time series always reflects the
actual demand value in the past period. So if
period 1 has an actual demand of 5, then the
forecast for period 2 is 5. See next slide:

Exponential Smoothing at 0.3

Forecast Accuracy for Exp. Smoot. (0.3)

Years

2000
2001

Actual (A)
Demand
(No. of
customer
orders)

Forecast (F)
(0.3 = )
Formula: Next forecast =
Previous forecast + (A F)

5
9

2002

6.2 = 5 + .3 (9 - 5)

2003

12

6.14 = 6.2 + .3 (6 - 6.2)

2004

7.90 = 6.14 +.3 (12 - 6.14)

2005

10

7.63 = 7.90 +.3 (7 - 7.90)

2006

8.34 = 7.63 +.3 (10 7.63)

Linear Trend Forecast Method

This involves the use of linear equation to describe
or assess a trend in a time series.
A trend is a long-term upward or downward
movement in a time series.
The linear trend equation has the form:
Forecast () = a + bt, where a = intercept or the
value of forecast () when t is 0; b is the slope of
the forecast line, and t = number of periods from t
= 0 (i.e. 1999).
Actual values/orders = y,
forecasted values/orders = .
One can must calculate the b and then a to be
able to compute a forecast for a particular period (t
=1,2,3,etc)

Yrs
2000
2001
2002
2003
2004
2005
2006
MAD
MSE

A.
5
9
6
12
7
10
8

F.
5
6.2
6.14
7.9
7.63
8.34

[Error]
4
.2
5.86
0.9
2.37
.34
2.28

[Error]2
16
.04
34.34
.81
5.62
.12
9.49

[ERROR] is the absolute error which is the difference

between actual and forecast: Minus signs removed.

Formula for b (slope)

Formula for a (intercept)
for Linear Trend Equation

1/13/2014

LINEAR TREND PROJECTION

Using the table data, compute t, ty, and
t2
Yrs
2000
2001
2002
2003
2004
2005
2006

Actual
y
5
9
6
12
7
10
8

ty
5
18
18
48
35
60
56

1
2
3
4
5
6
7

LINEAR TREND PROJECTION

Using the table data, compute t, ty, and t2
Yrs
2000
2001
2002
2003
2004
2005
2006

t2
1
4
9
16
25
36
49

Formula for b (slope)

Formula for a (intercept)
for Linear Trend Equation

Actual
y
5
9
6
12
7
10
8
57

28

ty
5
18
18
48
35
60
56
240

t2
1
4
9
16
25
36
49
140

ty

t2

1
2
3
4
5
6
7

Formula for b (slope)

Formula for a (intercept)
for Linear Trend Equation
7 x 240

28

57

282

7 x 140

57

28
.43*

Formula for b (slope)

Formula for a (intercept)
for Linear Trend Equation
1680
980

1596
0.43

6.42

784

57

Linear Trend Equation

Forecast: = a + bt
a (intercept) = 6.42
b (slope) = .43
So Linear trend equation =
= 6.42 + .43t

28
.43*
.43

1/13/2014

LINEAR TREND PROJECTION

Calculations of Forecasts
Yrs
2000
2001
2002
2003
2004
2005
2006

Actual
y
5
9
6
12
7
10
8

T
1
2
3
4
5
6
7

Forecast
Y = a + bt
6.85 = 6.42 + .43(1)
7.28 = 6.42 +.43 (2)
7.71 = 6.42 + .43 (3)
8.14 = 6.42 + .43 (4)
8.57 = 6.42 + .43 (5)
9.00 = 6.42 + .43 (6)
9.43 = 6.42 + .43 (7)

Forecast Accuracy for Linear Trend

Yrs
2000

A.
5

F.
6.85

[Error]
1.85

[Error]2
3.42

2001
2002
2003

9
6
12

7.28
7.71
8.14

1.72
1.71
3.86

2.96
2.92
14.90

2004
2005
2006
MAD
MSE

7
10
8

8.57
9.00
9.43

1.57
1
1.43
1.87

2.46
1
2.04
3.52

[ERROR] is the absolute error which is the difference

between actual and forecast: Minus signs removed.

COMPARE FORECASTING METHODS

Forecast
Method
Naive
2 yr MA

MAD

MSE

3.83
1.70

16.5
5.15

3 yr MA
Exp. Smoothing (0.3)
Linear Trend
Projection

2.67
2.28

9.50
9.49

1.87

3.52

Lowest
MAD

Lowest
MSE

2 yr M.A Forecast (based on MAD)

Yrs
Actual
Forecast
2000
5
2001
9
2002
6
7
2003
12
7.5
2004
7
9
2005
10
9.5
2006
8
8.5
2007
??
(10 + 8)/2 = 9
Hence, forecast for 2007 is 9000 orders
based on 2 yr MA.

Final Decision
The measure of forecast accuracy chosen
influences the forecast method selection.
One would use the 2 year MA as the
forecasting method of choice if MAD is the
preferred criterion, or one may use linear trend
method if the MSE is the preferred criterion.
Lets say we used the MAD as the criterion, you
can use the 2 year forecast method to predict for
the next year (in this case, year 2007). To
calculate this forecast for 2007, average the data
values for 2005 and 2006.
If we relied on the MSE as the criterion, you will
use the linear trend equation to predict 2007: =
6.42 + .43(8).

Linear Trend Forecast (based on MSE)

Yrs
2000
2001
2002
2003
2004
2005
2006
2007

Actual
y
5
9
6
12
7
10
8
?

T
1
2
3
4
5
6
7
8

Forecast
Y = a + bt
6.85 = 6.42 + .43(1)
7.28 = 6.42 +.43 (2)
7.71 = 6.42 + .43 (3)
8.14 = 6.42 + .43 (4)
8.57 = 6.42 + .43 (5)
9.00 = 6.42 + .43 (6)
9.43 = 6.42 + .43 (7)
9.86 = 6.42 + .43 (8)

Hence, forecast for 2007 is 9860 orders based on linear

trend forecast.

1/13/2014

Qualitative Forecasting

 Delphi Method
interactive group process consisting of obtaining
information from a group of respondents through
questionnaires and surveys
 Jury of Executive Opinion
obtains opinions of a small group of high-level
managers in combination with statistical models
 Sales Force Composite
allows each sales person to estimate the sales for
his/her region and then compiles the data at a
district or national level
 Consumer Market Survey
solicits input from customers or potential
customers regarding their future purchasing plans