Beruflich Dokumente
Kultur Dokumente
Management
Forecasting Techniques
G. Pond
Outline
Background
Time-Series Models
Associative Models
Forecast Accuracy
Forecast Control
Background
Within a business context, forecasting methods can
be used to forecast:
Consumer demand
Utility prices
Network expansion (social media)
Labour requirements
Machine-time demands
Capacity growth
Market share growth.
Demand
Dependent Demand
Internal (e.g., four screws for every chair
leg)
- How can firms influence internal demand?
Independent Demand
External (e.g. , consumer demand for toys
at Christmas)
How can firms influence external demand?
Time-Series Models
Nave Method
Moving Averages
Weighted Moving Averages
Exponential Smoothing
Linear Trend Projection
Trend-Adjusted Exponential Smoothing
Cyclical/Seasonal Demands
June Forecast
Value
May Observed
Value
Feb
Mar
Apr
May
June
July
0
Nov Dec Jan
Feb Mar Apr May June July Aug Sept Oct Nov Dec Jan
Feb
675 + 75 =
750
675
800
600
675 600 =
75
600
400
200
0
Jan
Feb
Mar
Apr
May
June
July
Moving Average
Forecast
Moving Average
Consider the following temperatures for the first
week in March, 1996, in Kingston:
Date
Temperature
1
-2
2
2
3
-4
Moving Average
Moving Average
The scattergram below depicts the number of
students registered in
business/commerce/management programs in the
province of Ontario: 12,000
10,000
8,000
Student Registration
6,000
4,000
2,000
0
2004
2006
2008
Year
2010
2012
Moving Average
12,000
10,000
8,000
Student Registration
Raw Data
6,000
n=2
4,000
n=3
n=4
2,000
0
2000 2005 2010 2015
Year
1
-2
2
2
3
-4
ts
Exponential Smoothing
where:
Exponential Smoothing
Exponential Smoothing
Year
Actual
2005
9,228
2006
9,297
2007
10,573
2008
11,047
2009
10,326
2010
10,354
2011
10,940
Develop
= 0.6
To start, use the nave method to obtain a
forecast value for Year 2006
Exponential Smoothing
Foreca
Actual st
Year
2005
2006
1
2
9,228
9,297
2007
3 10,573
2008
4 11,047
2011
7 10,940
9,228
Exponential Smoothing
Year
Foreca
Actual st
2005
2006
2007
1 9,228
2 9,297
3 10,573
2008
4 11,047
2009
5 10,326
2010
6 10,354
2011
7 10,940
9,228
9,269
Exponential Smoothing
11,000
10,500
10,000
Student Registration
Raw Data
9,500
alpha = 0.6
alpha = 0.5
9,000
alpha = 0.4
alpha = 0.1
8,500
8,000
200420062008201020122014
Year
Exponential Smoothing
Note:
Be sure to start the forecast using one of the
nave methods
Smaller values of the smoothing constant () put
more emphasis on forecast values and trend
(consequently, the forecast is more smooth).
Typical values for are in the range (0.05,0.5)
Use as much of the data as possible to help
train your model.
9,297
10,573
11,047
10,326
10,354
10,940
Actual
9,228
2006 Actual
9,297
Year
2007
2005
10,573
9,228
2008
2006
11,047
9,297
2009 10,573
10,326
2007
2010 11,047
10,354
2008
2007
2008
2009
2010
2011
10,573
11,047
10,326
10,354
10,940
1
2
3
4
15
6
27
3
4
5
6
7
1
4
9
16
25
1
36
4
49
9
16
25
36
49
9,228
18,594
31,719
44,188
51,630
9,228
62,124
18,594
76,580
31,719
44,188
51,630
62,124
76,580
2007
2008
2009
2010
2011
10,573
71,765
11,047
10,326
10,354
10,940
1
2
3
4
15
26
7
3
28
4
5
6
7
1
4
9
16
25
1
36
4
49
9
140
16
25
36
49
9,228
18,594
31,719
44,188
51,630
9,228
62,124
18,594
76,580
31,719
294,063
44,188
51,630
62,124
76,580
Now we have
everything we
need:
12,000
10,000
8,000
Student Registration
6,000
4,000
2,000
0
2005 2006 2007 2008 2009 2010 2011
Year
Select Add
Trendline
12,000
10,000
8,000
Student Registration
6,000
4,000
2,000
0
2005 2006 2007 2008 2009 2010 2011
Year
Trend-Adjusted
Exponential Smoothing
where:
Trend-Adjusted
Exponential Smoothing
200
6
Actual
1
9,228
2
9,297
3
10,573
4
11,047
Actual
5
10,326
6
10,354
10,940
17 9,228
9,297
We
can obtain by averaging the
first three observed values to get
Trend-Adjusted
Exponential Smoothing
Example
Year
2005
2006
Year
2007
2008
200
2009
5
2010
200
2011
6
200
7
200
8
200
9
Actual
1
9,228
2
9,297
Actual
3
10,573
4
11,047
5
10,326
16 9,228
10,354
7
10,940
9,297
3 10,573
4 11,047
5 10,326
Trend-Adjusted
Exponential Smoothing
Example
Year
Actual
2005
9,228
2006
2007
2
3
9,297
10,573
2008
11,047
9,699
673
10,372
2009
5with 10,326
NOW, we can proceed
using our TA Exp. Sm. formula:
2010
6
10,354
1. Find
2. Find
2011
7
10,940
3. Find
Trend-Adjusted
Exponential Smoothing
Example
Year
Actual
2005
9,228
2006
2007
2
3
9,297
10,573
2008
11,047
10,326
2010
10,354
2011
10,940
2009
9,699
673
10,372
Trend-Adjusted
Exponential Smoothing
Example
Year
Actual
2005
9,228
2006
2007
2
3
9,297
10,573
9,699
2008
11,047
10,432
2009
10,326
2010
10,354
2011
10,940
673
10,372
Trend-Adjusted
Exponential Smoothing
Example
Year
Actual
2005
9,228
2006
2007
2
3
9,297
10,573
9,669
673
2008
11,047
10,432
697
2009
2010
2011
10,326
10,354
10,940
10,372
Trend-Adjusted
Exponential Smoothing
Example
Continuing to use the TA Exp Sm.
Forecasting method throughout the
remainder of the table permits obtaining a
future value for 2012:
Year
Actual
2005
9,228
2006
2007
2008
2009
2010
2011
2
3
4
5
6
7
9,297
10,573
11,047
10,326
10,354
10,940
2012
9,699
10,432
11,105
11,352
11,411
673
697.1
687.2
511.3
330.2
10,372
11,129
11,792
11,863
11,741
11,501
234.1
11,735
Seasonal/Cyclic Demand
Pearson Airport (Enplaned + Deplaned)
Passenger Volume
1,800,000
1,600,000
1,400,000
Number of
Domestic
Passenger
s
1,200,000
1,000,000
800,000
600,000
400,000
200,000
0
Month
Seasonal/Cyclic Demand
First, we need to de-seasonalize the data. We can consider each
month as a season that repeats yearly and then find a 12-period
moving average.
- we use 12 periods because, in total, there are 12 seasons that
Date
Actual
repeat annually
Jan
926,061
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Jan
870,271
980,311
979,293
1,123,837
1,174,270
1,342,358
1,394,984
1,153,668
1,129,786
967,043
1,036,631
980,596
1,089,876
Seasonal/Cyclic Demand
First, we need to de-seasonalize the data. We can consider each
month as a season that repeats yearly and then find a 12-period
moving average.
- we use 12 periods because, in total, there are 12 seasons that
Date
Actual
repeat annually
Jan
926,061
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Jan
870,271
980,311
979,293
1,123,837
1,174,270
1,342,358
1,394,984
1,153,668
1,129,786
967,043
1,036,631
980,596
1,089,876
1,094,421
Seasonal/Cyclic Demand
First, we need to de-seasonalize the data. We can consider each
month as a season that repeats yearly and then find a 12-period
moving average.
- we use 12 periods because, in total, there are 12 seasons that
Date
Actual
repeat annually
Jan
926,061
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Jan
870,271
980,311
979,293
1,123,837
1,174,270
1,342,358
1,394,984
1,153,668
1,129,786
967,043
1,036,631
980,596
1,089,876
1,094,421
1,092,148
Seasonal/Cyclic Demand
Date
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Jan
Actual
926,061
870,271
980,311
979,293
1,123,837
1,174,270
1,342,358
1,394,984
1,153,668
1,129,786
967,043
1,036,631
980,596
Centred
Moving
Average
1,089,876
1,094,421
Seasonal
Relative for
July
1,092,148
Seasonal/Cyclic Demand
Date
Jan-11
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Jan-12
Feb
Mar
Apr
May
Jun
Jul
Actual
926,061
870,271
980,311
979,293
1,123,837
1,174,270
1,342,358
1,394,984
1,153,668
1,129,786
967,043
1,036,631
980,596
941,046
1,003,134
1,029,917
1,149,068
1,197,496
1,373,150
12-Period Moving
Average
1,089,876
1,094,421
1,100,319
1,102,221
1,106,439
1,108,542
1,110,477
Centred Moving
Average
1,092,148
1,097,370
1,101,270
1,104,330
1,107,490
1,109,510
Seasonal
Relative
1.22909856
1.27120705
1.04758005
1.02305124
0.87318405
0.93431467
Seasonal/Cyclic Demand
Date
Actual
2011 Jan
926,061
Feb
870,271
Mar
980,311
Apr
979,293
Seasonal
Relative
May
1,123,837
Jun
1,174,270
Jul
1,342,358
1.229
Aug
1,394,984
1.271
Sep
1,153,668
1.048
Oct
1,129,786
1.023
Seasonal/Cyclic Demand
Season
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Average
Seasonal
Relative
0.859
0.815
0.905
0.935
1.049
1.100
1.233
1.294
1.046
1.042
0.877
0.934
Seasonal/Cyclic Demand
Now divide each of the observed (actual) values by its seasonal
Seasonal Deseasonalized
Year
Date
Actual
Relative Passenger Load
relative
2011 Jan
926,061 0.858985
1,078,087
Feb
870,271
0.814857
1,068,005
Mar
980,311
0.90535
1,082,798
Apr
979,293
0.935123
1,047,234
May
1,123,837
1.04872
1,071,628
Jun
1,174,270
1.099889
1,067,626
Jul
1,342,358
1.233022
1,088,673
Aug
1,394,984
1.294222
1,077,855
Sep
1,153,668
1.046004
1,102,929
Oct
1,129,786
1.042401
1,083,830
Seasonal/Cyclic Demand
1,800,000
1,600,000
1,400,000
1,200,000
1,000,000
800,000
600,000
400,000
200,000
0
Seasonal/Cyclic Demand
Suppose we want to forecast passenger load for December, 2015
(Period #62 in my data set),
Associative Models
Associate
models are those where the independent variable is
something other than time. For example, it may be forecasting your
term grade (the dependent variable, ) as a function of the number of
practice problems you complete (the independent variable, )
Linear regression can be done to obtain a trend, just as it was done
before. The only difference is that where appeared in the previous
equations, it is replaced with
where:
With
so many different forecasting
methods, how does one know which is
best?
Mean Absolute Deviation (MAD)
Mean Absolute Percentage Error
(MAPE)
Mean Squared Error (MSE)
Correlation coefficient ()
Correlation Coefficient
Correlation Coefficient
12,000
10,000
R
f(x)==1250.11x + 9251.71
8,000
Student Registration
6,000
4,000
2,000
0
2005 2006 2007 2008 2009 2010 2011
Year
Warning:
Higher values arent always better. A perfect fit to your
data can be made using a polynomial function but it might be
garbage when used in forecasting
-2
-4
-4
-1.3
-2
-2
-7
-10
Error ()
Date
2.7
4
0
5
Squared Error ()
-2
-4
-4
7.1
-2
0
-7
-10
1.3
2.7
-2
7.1
-2
-4
-4
-1.3
-2
-2
-7
-10
Error ()
Date
2.7
4
0
5
Squared Error ()
-2
-4
-4
7.1
-2
0
-7
-10
1.3
2.7
-2
7.1
-2
-4
-4
-1.3
-2
-2
-7
-10
Error ()
Date
2.7
4
0
5
Squared Error ()
-2
-4
-4
7.1
-2
0
-7
-10
-1.3
67.5%
-2
0%
2.7
7.1
67.5
%
0%
% Error
% Error
Tracking
Signal
Amount of Historical
Data
Data Pattern
Forecast
Horizon
Preparation time
Complexity
Simple
exponential
smoothing
5 to 10 observations
Data should be
stationary
Short
Short
Little sophistication
Trend-adjusted
exponential
smoothing
10 to 15
observations
Trend
Short to
medium
Short
Moderate
sophistication
Regression
Trend models
10 to 20
Trend
Short,
medium, long
Short
Moderate
sophistication
Seasonal
Enough to see 3
peaks and troughs
seasonal patterns
Short to
medium
Short to moderate
Moderate
sophistication
Causal
regression
models
10 obs per
independent
variable
Medium or
long
Long development
time, short time
implementation
Considerable
sophistication
Source: J. Holton Wilson and D. Allison-Koerber, Combining Subjective and Objective Forecasts Improves Results,
Journal of Business Forecasting Methods & Systems, 11(3) Fall 1992, p. 4.
62
Short Term
daily,
weekly
Medium Term
monthly,
quarterly
Long Term
annual
2. Level of
aggregation
Item
Product family
Total output
3. Type of model
Smoothing
Trend
4. Degree of
management
involvement
5. Cost per forecast
Low
Trend
Seasonal
Regression
Moderate
Managerial
Judgment
Trend Regression
High
Low
Moderate
High
Source: C. L. Jain, Benchmarking Forecasting Models, Journal of Business Forecasting Methods & Systems, Fall
2002, pp. 1820, 30.
63
Review Chapter 3
Try Problems 4, 10 17
Read Chapter 8