TM 745 Forecasting for Business

1
TM 745 Forecasting for Business TechnologyDr.
Frank Joseph Matejcik
7th Session 6/28/07 Chapter 8 Combining
Forecast Results Chapter 9 Forecast
Implementation

• South Dakota School of Mines and Technology,
  Rapid City

2
Tentative Schedule
Chapters Assigned 17-May 1 e-mail,
contact problems 1,4, 8 24-May 2
problems 4, 8, 9 31-May 3,4 problems
ch3(1,5,8,11) ch4(6,10) 07-June 5 problems
5,8 14-June 6 start Test (Covering chapters 1-4)
Study Guide is on the class website. problems 4,
7 21-June 6 finish, 7 problems 3, 4, 5(series A
only), and 7B 28-June 8, 9 problem 6 in chapter
8 05-July Final Traveling in July and
August Dates on board
3
Web Resources

• Class Web site on the HPCnet system
• http//sdmines.sdsmt.edu/sdsmt/directory/courses/2
  007su/tm745001
• Streaming video http//its.sdsmt.edu/Distance/
• Answers at http//www.hpcnet.org/what63
• The same class session that is on the DVD is on
  the stream in lower quality. http//www.flashget.c
  om/ will allow you to capture the stream more
  readily and review the lecture, anywhere you can
  get your computer to run.

4
Agenda New Assignment

• Chapter 8 problem 6, Chapter 9 no problems
• Final is next week
• Study guide is posted
• Chapter 8 Combining Forecast Results
• Chapter 9 Forecast Implementation

5
Combining Forecast Results

• Intro
• Bias
• Ex. What can be combined?
• How to get the weights?
• Three techniques
• Delfield
• About ForecastX comnbining

6
Introduction

• 83 of experts believe that combining forecasts
  will produce more accurate forecasts than
  originals Collopy Armstrong (1992)
• Bates Granger (1969) 1st idea
• Why the best forecast may not be
• 1) Some variables may be missing
• 2) Discarded forecast may use a type of relation
  ship ignored in the best forecast

7
Bias

• Unbiased here not used strictly as in Statistics
• Statistics term unbiased
• 1) a strong property of a statistics
• 2) excludes reasonable statistics
• Forecasters believes may influence forecasts
• Try to ignore preconceived ideas
• Fresh employees may help

8
An Example

• Output indexes of Gas, Electricity, Water
• Linear Fit

9
An Example Exponential fit

• Uses a transform to fit it

10
An Example Combined fit

• Combined Improvements
• 1) Optimal weights yield considerable
  improvements
• 2) combined forecasts statement 2)page 395 is
  not quite correct. It happens. See table 8-1
  (Elmo)

11
What Forecasts Are Combined?

• Actual Practice try very different models
• 1) Extract different predictive factors
• a) transforms
• b) model format
• 2) Models use different variables
• Air Travel Forecast
• 1) Judgmental (Expert survey)
• 2) Extrapolation (time series)
• 3) Segmentation (Customer survey)
• 4) Econometric (Causal regression)

12
What Forecasts Are Combined?
13
Choosing Weights for Combined Forecasts

• Armstrong likes equal weights (ex ante)
• MAPEs reduced 6.6
• Better if gt2 forecasts
• Bates Granger weight more accurate more heavily
• In general combined forecast have smaller errors
  (exs bw)
• Book suggests weight more accurate more heavily

14
3 Techniques for Selecting Weights

• 1) Minimize variances of forecasts
• 2) Adaptive weights based on each error
• 3) Use regression on the forecasts. (Optimal
  linear composite)

15
Minimize variances of forecasts
16
Adaptive weights based on each error
17
Optimal linear composite
18
Optimal linear composite procedure

• Constant term is found and tested, if tested to
  be in the model dont apply it.
• Comment b1 b2 about 1, b1, b2 gt 1
• Comment F(1) F(2) will likely haveconstant
  terms. So?

19
Regression for Combining Household Cleaner,
application

• Sales by Sales Force Composite
• Sales by Winters Method

20
Regression for Combining Household Cleaner,
application

• Run usual regression
• Force constant to be zero
• Improved

21
Forecasting THS with a Combined Forecast

• Time Series Decomposition chapter 6
• Multiple regression chapter 5THS106.31
  10.78(MR) - 0.45(ICS)
• THS -2.540.06(THSRF)0.97(THSDF) (-1.2)
  (1.53) (31.8)
• THS 0.03(THSRF) 0.97(THSDF)
• RMSE combined 3.54
• RMSE THSDF Winters 3.55

22
Forecasting THS with a Combined Forecast
23
Forecasting DCS with a Combined Forecast
24
Forecasting DCS with a Combined Forecast
25
Comment from the field

• Delfield Company Food Service Co.
• 4th edition table below

26
Integrative Case The Gap 4th
27
Integrative Case The Gap 4th
28
Integrative Case The Gap 4th
29
Using ForecastXTM to Combine Forecasts

• It is just the regression that was given, so just
  remember to check the no intercept box.

30
9 Forecast Implementation

• Keys (a list)
• Forecast Process (steps)
• Choosing the right forecast
• New Product
• Artificial Intelligence

31
Keys to Obtaining Better Forecasts

• 1. Understand what forecasting is is not
• Focus on management processes controls, not
  computers Establish forecasting group
• Implement management control systems before
  selecting forecasting software
• Derive plans from forecasts
• Distinguish between forecasts and goals
• Forecasting is acknowledged as a critical
• Accuracy emphasized not game-playing

32
Keys to Obtaining Better Forecasts

• 2. Forecast demand, plan supply
• Dont use shipments as actual demand
• Identify sources of demand information
• Build systems to capture key demand data
• Get improved customer service capital planning

33
Keys to Obtaining Better Forecasts

• 3. Communicate, cooperate, collaborate
• Avoids duplication Mistrust of "official
  forecast
• Creates understanding of impact throughout
• Establish a cross-functional approach to
  forecasting

34
Keys to Obtaining Better Forecasts

• 3. Communicate, cooperate, collaborate
• Establish an independent forecast group that
  sponsors cross-functional collaboration
• All relevant information used to generate
  forecasts
• Forecasts trusted by users
• More accurate relevant forecasts

35
Keys to Obtaining Better Forecasts

• 4. Eliminate islands of analysis
• Mistrust inadequate information leading
  different users to create their own forecasts
• Build 1 "forecasting infrastructure"
• More accurate, relevant, credible forecasts
• Provide training for both users developers of
  forecasts
• Optimized investments in information
  communication systems

36
Keys to Obtaining Better Forecasts

• 5. Use tools wisely
• Relying solely on qualitative or quantitative
• Integrate quantitative qualitative methods
• Identify sources of improved accuracy increased
  error
• Provide instruction
• Process improvement in efficiency effectiveness

37
Keys to Obtaining Better Forecasts

• 6. Make it important
• Have accountability for poor forecasts
• So developers can understand forecast uses
• Training developers to understand implications of
  poor forecasts
• Include forecast performance in performance
  plans reward systems
• Striving for accuracy credibility

38
Keys to Obtaining Better Forecasts

• 7. Measure, measure, measure
• Know if the firm is getting better
• Measure accuracy at relevant levels of
  aggregation
• Ability to isolate sources of forecast error
• Establish multidimensional metrics
• Incorporate multilevel measures
• Measure accuracy whenever wherever forecasts
  are adjusted

39
Keys to Obtaining Better Forecasts

• 7. Measure, measure, measure
• Forecast performance can be included in
  individual performance plans
• Sources of errors can be isolated and targeted
  for improvement
• Achieve greater confidence in forecast process

40
The Forecast Process

• 1. Specify objectives
• Articulate role of forecast in decisions
• If forecasts dont effect decisions, Why?
• 2. Determine what to forecast
• Sales revenue or units?
• weekly, annually, quarterly?
• Communicate with user

41
The Forecast Process

• 3. Identify time dimensions
• Horizon
• Frequency
• Urgency
• 4. Data considerations
• Internal needs database management
  disaggregation time, unit, region
• External govt, trade association

42
The Forecast Process

• 5. Model selection (next section)
• 6. Model evaluation
• Less important for subjective methods
• Use holdout method if quantitative
• Go back to step five if a problem
• 7. Forecast preparation
• Try for multiple multiple types

43
The Forecast Process

• 8. Forecast presentation
• Management must understand be confident
  (corporate culture)
• Oral written
• same time same level
• be generous with charts etc.
• 9. Tracking results
• process continues
• reviews open, objective, positive

44
Choosing the Right Forecasting Techniques

• Few hard and fast rules (guidelines)
• Focus on data, time, personnel
• Subjective Methods
• Sales force composite
• short to medium term
• Preparation time is quick once set up
• Customer surveys
• medium to long term, take 2-3 months
• survey research is a profession

45
Choosing the Right Forecasting Techniques

• Subjective Methods
• Jury of Executive Opinion
• Requires Expertise
• Is relatively quick to prepare
• Delphi
• long to medium term
• useful for new products
• can be slow computers help
• alternatives are better

46
Choosing the Right Forecasting Techniques

• Objective Methods
• Naive (little data, sometimes good)
• Moving Averages (easy, little data)
• Exponential Smoothing Simple
• Need to establish weight
• Easy to compute, quick
• Adaptive response ES
• short term, no seasonality
• Users need little background

47
Choosing the Right Forecasting Techniques

• Objective Methods
• Holt's ES
• short term, no seasonality, trend included
• Users need little background
• Winters ES
• short term, seasonality, trend included
• Need 4 or 5 observations per season
• Need computer for updates
• Users need little background (tell them about
  weighted dates)

48
Choosing the Right Forecasting Techniques

• Objective Methods
• Regression-Based
• Trend (10 observations, quick to develop, easy
  for users, modest developer skills)
• Trend with Seasonality (Need 4 or 5 observations
  per season, short to medium term, need a
  computer, usually little sophistication)
• Causal (10 observations per independent
  variable, short, medium, or long term,
  developers need regression skills.)

49
Choosing the Right Forecasting Techniques

• Objective Methods
• Time-Series Decomposition (two peaks and two
  troughs per cycle, 4 to 5 seasons of data, can
  use turning points, short to medium range, modest
  sophistication, managers like it.)
• ARIMA (managers dont like it, it takes a
  skillful developer, Need a computer to do ACF
  and PACF plots)

50
New Product Forecasting

• Product Life Cycle (PLC) curve

51
New Product Forecasting

• Analog forecasts
• Similar products
• Think Christmas movie toys
• Test marketing
• Pick a smaller representative place
• Ex. given is Indianapolis
• Product clinics (panel lab study)
• Type of Product Affects NPF

52
Artificial Intelligence and Forecasting

• Expert systems
• Neural Networks

Summary

• Difficult task many considerations
• New opportunities

53
Using ProCastTM in ForecastXTM to Make Forecasts

• It is okay now that you know what you are doing.
• You understand that a selection method is
  choosing the best of things that you already know.