32 views

Uploaded by Pragyan Nanda

- Noise Resilient Periodicity Mining In Time Series Data bases
- 06rr410308 Production Planning and Control
- Forecasting Sk
- Student Slides Chapter 3
- Forecating Questionnaire-Customer (Process)
- OM Assignment 2
- 210094081-Speciality-Packaging-Case-study.pdf
- 2inventory Management
- Forecasting Lecture
- Rao, Ambar G. and Gary Lilien - A System of Promotional Models
- UNIT II DF
- F-GD02 STLF
- Analytic
- Forecasting 1
- IAME 14
- PEC Productivity (Equipment)
- End Aut 2011
- Sales Objectives and Quotas
- Class Notes RM2
- OSCM Summary

You are on page 1of 161

In preparing the manuscript for the third edition of Forecasting: methods and applications,

one of our primary goals has been to make the book as complete and thorough as possible

in order that it might best meet its intended objectives. The same set of principles has

guided us in preparing this instructors manual. Our intent has been not only to provide

solutions to the exercises but to go beyond and suggest several other types of teaching

materials and suggestions to help those who teach forecasting. We hope that you will find

that this manual delivers on those objectives.

The instructors manual is in four parts. To avoid confusion with the chapters in the

textbook, we will refer to these as Parts A through D. When the word Chapter is used, it

refers to that chapter in the text itself.

Part A is aimed at providing different course outlines for a number of different settings

in which the text has been used. These range from short executive seminars to subsegments

of a required college course to full-length courses at the graduate level on the subject of

forecasting.

In Part B, we have provided some teaching suggestions as to how we would teach a

course based on the book.

When teaching, we always use a range of additional teaching materials as complements

to the text. In Part C, we discuss the use of case studies and provide suggestions for

projects and exam questions. There use will depend on the overall structure, teaching

style and design selected for the course.

Part D provides solutions to the end-of-chapter exercises. We have provided solutions

that can be used in teaching the course rather than just for grading student work. We hope

the graphs, tables and descriptions will be useful in preparing overhead transparencies or

handouts for students.

We have long found it useful to teach forecasting using a computer package for the

computational aspects of the subject. We have chosen in this edition not to emphasize

a particular package but to comment on the facilities available in a range of packages

(see Appendix I of the text). It is important to have a package that the students can

iii

iv Preface

learn relatively quickly and which provides as many of the statistical facilities as possible.

This will depend on the students background and the type of course being taught. In

preparing the solutions, we have mainly used Minitab version 11 and SAS version 6.12.

Be aware that other packages may give slightly different numerical results due to different

algorithms being used.

We would like to express special thanks to a number of instructors who have helped us

with this manual and given us their feedback from use of the second edition of Forecasting:

methods and applications. At Stanford University, the book has been used by Professor

Fred Shepardson and Professor Peter Reiss. Teaching materials were also provided by

two of our colleagues at the University of Virginia, Professor Jim Freeland and Professor

Bob Landel and by Dr Gary Grunwald who provided some ideas for the exercises while

teaching at the University of Melbourne.

Spyros Makridakis

Fontainebleau, France

Steven Wheelwright

Boston, Massachusetts

Rob Hyndman

Melbourne, Australia

November 1997

Contents

v

A/Planning a forecasting course

Since the late 1940s organizational forecasting has been directly affected by numerous de-

velopments in estimation and prediction. Today organizations of all sizes find it essential to

make forecasts in order to reduce the uncertainty of their environment and take advantage

of opportunities. We have been involved in the forecasting area for more than 25 years,

and this book is the culmination of our efforts and experiences at teaching forecasting to

both full-time students and practicing managers with forecasting responsibility.

This book grew out of our perception that a book was needed that would cover a

full range of forecasting methods, be accurate and complete in describing the essential

characteristics and steps for applying those methods in practice, and yet not get bogged

down in theoretical questions underlying the development of individual techniques. The

purpose of the book is to fill a gap in the literature by presenting in terms that are easily

understandable (but that accurately and rigorously describe the techniques) a wide range

of forecasting methods useful to students and practitioners of management, economics,

engineering, and other disciplines requiring effective forecasting.

The material included in the book has been particularly effective in a wide range of

teaching activities. These have included seminars for middle management, seminars for

practicing forecasters, and classes taught to both graduates and undergraduates in busi-

ness and in more specialized topic areas such as courses for statisticians, economists, and

management scientists.

Four objectives guided the structuring and development of the major sections of this

book. It is useful to understand these objectives before describing the way in which various

materials might be combined to develop an effective course or program on forecasting

and planning. Part A of this manual reviews these objectives and describes how the

materials in the text can be used to teach forecasting effectively. Alternative course

outlines that indicate how various parts of the book might be used for different audiences

and in programs of different length are also suggested. Parts B and C of this manual

provide additional information on the use of supplementary material and the teaching of

individual sessions in a forecasting course.

The various uses of the book can probably best be seen by considering specific course

1

2 Part A. Planning a forecasting course

outlines built around the text. Several such outlines are discussed and presented in Section

A/3.

A/1 Objectives

The major objectives we pursued in structuring and developing the four parts in this book

included the following:

detail and clarity that they could be applied easily.

nical background would be required to understand each technique, yet including

enough of the essential concepts and their theoretical basis that students could gain

a thorough understanding of each technique if desired.

sociated with each of the major forecasting methodologies so that criteria for select-

ing the most appropriate methods can be developed and applied.

4. Examining the important factors and issues in the application of forecasting and

planning and the effective use of forecasting resources in an ongoing organization.

It should be emphasized that while the body of each chapter is geared to the reader

with only a basic background in algebra, the aim has still been to present a complete

description of each technique and those factors that are relevant in deciding where and

how to use it.

In addition to having a class study individual chapters, an instructor has four other major

sources of material to use as building blocks in developing a forecasting course.

Exercises As indicated in the preceding section, many of the chapters in this book deal

with specific techniques for forecasting. At the end of these methodology-oriented

chapters are exercises that we have found particularly helpful in teaching forecast-

ing. The purpose of these exercises is to give students the opportunity to test their

A/3 Sample course outlines 3

ing of the strengths and limitations of those approaches. Complete solutions for

these exercises are provided in Part D of this manual. All data sets used in the

exercises are available from the books web page at

www.maths.monash.edu.au/hyndman/forecasting/

Case studies Case studies on forecasting situations can take a variety of forms including

basic exercises with a few paragraphs describing the situation surrounding that exer-

cise, providing organizational descriptions of a forecasting system, and even outlining

the management perspective of a decision maker who must use forecasts prepared

in accordance with a given methodology. Part C outlines procedures for using such

cases and suggests how those cases might be used to supplement the exercises at

the end of each chapter. Part C concludes with several additional exercises, project

suggestions and examination questions that might be useful in a forecasting course.

quantitative techniques described in the text are available and their facilities are

summarized in Appendix I of the book. We have found that it is essential to use a

computer package when teaching Forecasting to enable students to apply the tech-

niques themselves without unnecessarily detailed calculations. In addition, use of

such a package has proved a particularly appropriate means of making students

aware of the strengths and weakness of individual techniques on the basis of their

applicability and results in a range of situations.

Additional readings At the end of each chapter is a set of references that we have found

relevant to topics raised in the chapter. These references provide more depth on the

most important issues and suggest follow-up material for students with a special

interest in a given topic.

We have used the materials included in Forecasting: Methods and Applications, third

edition, for a variety of purposes. These have varied in terms of (1) number of class

sessions in the course, (2) the audience at which the course is directed, and (3) the mix of

lecture and case sessions.

In each course, we find it useful to use real case studies. These are most successful

when the subject matter is of relevance to those attending the course. For longer courses

designed for users, it is important to include computer laboratory sessions giving those

4 Part A. Planning a forecasting course

attending real experience in forecasting real data. Often the concepts that are covered in

lectures are only understood after a person has carried out the procedure and seen the

results on a computer.

Here we provide six course outlines that vary along these three dimensions. Chapter

references are given for background reading. In longer courses, we would cover all (or

almost all) of the material in each chapter. In shorter courses, we given only a brief

introduction to the material listed.

Session 1: Introduction (Chapter 1)

Evaluating the performance of a forecasting system, identifying issues in organiza-

tional forecasting, and providing an overview of forecasting methodology.

Exponential smoothing methods and ARIMA models.

Correlation analysis, simple regression, and multiple regression.

Introduction of a basic framework.

Auditing the status of forecasting, identifying key problems, designing, and imple-

menting effective action plans.

The first outline is for a one-day course that would serve as a primer for potential users

of forecasting in a business organization. Given the audience, the teaching approach is a

combination of lectures and case discussions that can draw into the class the experience

of individual participants.

It is assumed that those attending such a course have some basic mathematical abilities;

if they also have some forecasting experience, it is assumed that it is in a fairly limited area

or that it involves a fairly narrow set of methodologies. As outlined below, selected sections

of half a dozen chapters in this book can be used to illustrate various methodologies and

the ways in which forecasting management issues can be tackled. The references to the

textbook are given in italics.

A/3 Sample course outlines 5

In the second course outline, 12 sessions spread over two days are used to introduce

those with forecasting responsibility in a company to various methodologies and to an

interactive forecasting computer package. This program has been used with companies

who have have assigned forecasting responsibility to several people acting as business

managers for various product lines or as product/market managers for various segments

of the companys business.

The text chapters indicated serve as further reference on specific topics. The basic

information on those topics would be covered during lecture class sessions and would be

applied during the sessions that use the computer. Cases are used as the source of data

and as illustrations of various management issues.

Course C: 11-session (two-day) course for managers who will use forecasting

This is also a two-day program (11 sessions) aimed at managers who make use of forecasting

rather than at forecasters. These may be managers who must do their own forecasting or

managers who interface with a forecasting group, such as at a public utility or in a large

consumer products firm. Again, this program assumes that an interactive forecasting

package is available, and output from the package is used to illustrate methodologies.

However, the participants do not actually use the package during the course.

Much of the class time is spent describing the way in which alternative forecasting

techniques can tackle representative management problems. The objectives of such a

course would be to help managers identify situations in which a quantitative forecasting

approach would be appropriate and how to describe and define such situations so that

such a technique could be applied.

This course gives a nine-session program that could be provided as an elective in a regular

college or university curriculum or it might be compressed into a three-day program for

managers or practicing forecasters. It would go into more depth on the topics covered in the

outlines for one- or two-day courses and take a management orientation by concentrating

on case applications and their discussion in the classroom. As indicated, this course outline

includes specific time for group work after individual preparation of the cases.

6 Part A. Planning a forecasting course

Session 1: Introduction and defining the forecasting problem (Chapters 1 and 2)

Major issues in forecasting, the concept of a forecasting strategy, a framework for

classifying forecasting methodologies, and measuring forecasting error.

Single exponential smoothing, Holts method, and Holt-Winters seasonal method.

A laboratory session introducing a forecasting package. Inputting data, plotting data,

using exponential smoothing methods for tackling a given case study.

Autoregressive and moving average models for time series analysis, autocorrelation

analysis, and model selection.

Determining what to forecast, deciding how to forecast, and securing the required

data. Use examples from past experience.

Application of smoothing methods and time series analysis to given case studies.

Correlation analysis, simple linear regression, and statistical tests of significance.

Basics of time series multiple regression, causal factors in multiple regression, statis-

tical characteristics of this method.

Applying simple and multiple regression to given case studies.

Session 10: Selection of a forecasting methodology (Section 2/4 and Chapter 11)

Techniques for analyzing the characteristics of a given data set, criteria for select-

ing a forecasting methodology, and comparing the results obtained from alternative

techniques

Designing a forecasting strategy, measuring performance, and considering organiza-

tional aspects of the forecasting function

A/3 Sample course outlines 7

Session 1: Overview of forecasting for management (Chapter 1)

A management perspective on forecasting, introduction to forecasting methodologies,

accuracy as a performance criterion, additional criteria for selecting a forecasting

method.

Session 2: Seasonal indices and decomposition (Chapter 3/1, 3/2 and 3/4)

Classical decomposition, computation of seasonal indices, deseasonalizing a data se-

ries.

Basic smoothing models, performance on trend and seasonal patterns.

Philosophy and characteristics, overview of program structure and components, data

creation and handling, inputs and outputs to the program, sample runs, limitations,

and applications.

Time dependence, correlation error analysis. Apply to errors from exponential

smoothing method.

Concepts and theory, performance, and practice.

Linear time trend analysis, least squares estimation, performance, and practice.

Causal relationships, illustrative examples, performance, and practice.

Computer programs, business cycles, and judgmental inputs.

Selecting the right data, specifying the forecasting project, determining support re-

quirements, using available methodologies.

8 Part A. Planning a forecasting course

Session 1: Introduction and overview

(Chapter 1)

Classical decomposition, computation of seasonal indices, deseasonalizing a data se-

ries. Value of decomposition for forecasting.

Group work on exercises for smoothing time series. Discussion of time series smoothing

and evaluation of accuracy

Introduction to autoregressive and moving average models for time series analysis,

autocorrelation analysis, and model selection. Some exercises on model selection and

autocorrelation analysis of errors. Interpretation of forecasts and prediction intervals.

Correlation, least squares estimation, simple linear regression, statistical tests of sig-

nificance, forecasts.

Seasonal dummy variables, other explanatory variables, variable selection, estimation,

forecasts, collinearity.

Group discussion of a case study. Identification of appropriate variables, variable

selection, interpretation of computer output, use of model for forecasting, forecasting

explanatory variables.

A/3 Sample course outlines 9

Session 1: Introduction to short-range forecasting (Chapters 1, 2, and 12)

Plots, seasonality, autocorrelation.

Simple models, calculation of forecasts, autocorrelation of errors.

Overview of models, model selection, interpretation of forecasts, and prediction inter-

vals.

This short course on forecasting is actually a four session segment of an MBA elective

course on production and operations management that focuses on short-range forecasting

as an input to production planning and control. Students are expected to use an interac-

tive forecasting package in applying selected methods to a variety of production planning

situations.

This course is designed as a full-quarter elective course for MBA students who may accept

job assignments with forecasting responsibility immediately upon graduation. Thus, stu-

dents electing this course would tend to have a fairly good background in mathematics and

would be particularly interested in the knowledge needed to apply individual techniques.

The majority of the class sessions would involve lectures dealing with various topics related

to quantitative forecasting techniques and their application. However, at the end of each

of several major sections of the course, one or more cases would be used to show the prac-

tical application of the concepts being discussed and some of the difficulties encountered

in their implementation. A forecasting competition would be conducted during the course

so students could forecast a variable, obtain results, and then prepare further forecasts.

We have also used class visitors at the end of a long course of this type. These would

be practicing forecasters from business or industry who would provide a fresh perspective

on the practice of forecasting and would also discuss non-statistical problems they have

encountered in forecasting.

10 Part A. Planning a forecasting course

Session 1: Introduction to forecasting - Why forecast? (Chapter 1)

Summarizing data and relationships: a review of some useful concepts Explanatory

methods of forecasting. Forecasting and causality

A review of some simple quantitative concepts, time series, moving average smoothing.

Trend analysis and seasonality, detrending, classical decomposition, overview of other

decomposition methods.

Simple exponential smoothing, Holts method, Holt-Winters method.

Regression analysis, correlation, least squares estimation, tests of significance.

Multiple regression models, significance tests, variable selection,

More on estimating regressions, diagnostics, transformations, prediction with regres-

sion models.

Lagged variables, spurious regressions.

Econometrics and economic models.

tests for autocorrelation, analysis of errors from forecasting methods, autoregressive

models.

ARMA and ARIMA models

Box-Jenkins procedures, estimation, tests

A/3 Sample course outlines 11

Examples and applications.

Overview of some more advanced methods.

Cycle analysis and indexes, cycles and the forecast problem, index construction, lead-

ing indicators.

Subjective methods, estimation, tests, model selection criteria, estimation, and diag-

nostics, choosing a method to fit the problem.

Measures of accuracy. How do we compare forecasts? Canonical procedures, rules of

thumb.

Review and award presentation for the competition

A major project is also a worthwhile adjunct to such a course. One option would be

to give students a case and ask them to prepare the complete analysis and forecasts in

both written and oral form for the management of that company. Such a project allows

students not only to test their knowledge of various techniques but also to handle the

problems of deciding what to forecast and how, and determining what data would be

most appropriate.

A final variation of this forecasting project that we have used with some success is to

identify a forecasting situation in an existing company and to have the manager in that

situation come to class so that the students, acting as forecasters, can define the task and

determine what information is needed. Students can work on the project in teams and

make their final presentations to the manager involved. This type of project is perhaps

the best possible test of students knowledge of the subject area and their ability to handle

the practical aspects of forecasting in an organization. The final presentation could serve

as the final exam for the students.

Many other course outlines based on the material in this book are possible, of course.

Here we wish to introduce the range of such outlines and suggest how they might be

adapted to use the complementary materials described in Parts B and C of this manual.

B/Teaching suggestions

The fate of a course is often decided early. So motivating the subject is critical. This

chapter needs help to make it a live opening for the course, and here are some suggestions.

(b) Acquiring additional resources

(c) Determining what future resources are needed

If possible, invite a couple of business practitioners into the first class to make the

reality of these matters clear.

2. Consider Figure 1-1 and study it from the point of view of the group responsible for

making the sales forecast for a company. Sales forecasting is one of the most routine

and most fundamental business tasks. It directly impinges on budget policy at all

levels of the firm. Make the point that there are two directions to look:

Sales Forecast -

3. Following the previous point, try to establish (i) the context within which a fore-

cast is made, and (ii) the environment in which a forecast is received. There is

more to forecasting than learning a mathematical method, getting data, running a

computer program, and reporting on extrapolated future values. Sales of a product

group take place within the context of sales of competing products (both internal

to the company and externally). Sales depend on market demand. Market demand

12

Chapter 1: The forecasting perspective 13

there. When the sales forecast has been made it is passed on in the hierarchy, to

be examined, modified by expert judgment, repackaged, and passed on again. Fore-

casting is not a passive activity. The physical environment within which forecasts

are created is dynamic, and the organizational consumer of forecasts shares these

dynamics and adds its own human relations.

4. Table 1-1 offers opportunities for class participation. What forecasting scenarios

are hot topics right nowin the local environment (what will acid rain do to our

forests)? in the nation (will the national deficit ever decrease)? in any firm that you

know about (will our ski businesses survive in the uncertain weather)? And so on.

5. Sections 1/2 and 1/3 begin to introduce the jargon of the field of forecasting, and it

is important to establish some of the most commonly used terms. For example, a lot

will be said about time series methods and explanatory methods. Distinctions are

also made between quantitative methods and qualitative methods. And you can no

doubt think up other dichotomous categories. The purpose of these categorizations

is to aid in communication.

For example, with the two mentioned above, we can identify four cells:

time series causal

quantitative non-quantitative

time series causal

non-quantitative quantitative

As more and more dichotomous categories are invented we can generate many, many

possible (methodology) cellsbut some of them will not be possible. It might be

enough to try to think of examples for the four cells given above.

6. Be careful with the distinction between explanatory models and time series models.

There is some semantic confusion with this choice of words, because explanatory

models can deal with time series data. So the context has to help out. When we

say time series models or time series methods we usually mean to talk about one set

of (time-dependent) data and we will try to develop a model (an equation) which can

be though of as the generating process of these data. We specifically will not look

at its relationship with other variables. When we causal models we are specifically

looking for other variables (which may be time series) which offer explanations (or

linkages) to the main variable (which may also be a time series).

7. It is important to see forecasting as a process involving several steps as outlined in

Section 1/3. Many people focus on just Step 4: choosing and fitting models. To

emphasize the basic steps in a forecasting task, it is helpful to take a particular

example and lead a class discussion on what is involved in each of the five steps.

14 Part B. Teaching suggestions

Chapter 2 introduces some of the basic quantitative tools that will be used extensively

later. Therefore it is important the material covered here is well understood.

1. The two examples in section 2/1 will help to fix the distinction between explanatory

and time series methods.

2. The patterns discussed in Section 2/2/1 are important for later reference. In the

many time series to be dealt with in the book, it will become second nature to ask

(i) is there a trend? (ii) is there an annual pattern (seasonality)? (iii) is there a

longer (than one year) cycle?

It is important to emphasize that when we talk about business cycles we mean

something different from seasonality which refers to the annual pattern. Cycles

are almost always longer than seasonal patterns.

3. The plots in Section 2/2, particularly the time plot and the scatterplot, will be

used extensively in subsequent chapters. These also help students understand the

difference between explanatory and time series methods.

4. In dealing with the familiar descriptive statistics in Section 2/3, the following aspects

could be stressed.

(a) Cross-sectional versus time series data: Some students find it difficult to switch

their thinking from elementary descriptive statistics of cross-sectional data (e.g.,

the weights of 100 mice) to time series data which may have trend, seasonality

and longer term business cycles (e.g., housing starts). The mean of housing

starts does not convey much information because of the trend, seasonality and

cycle. Similarly, the variance of housing starts is hard to interpret. So in

section 2/3, as each descriptive statistic is defined, it is useful to ask: is there

any difference in meaning when we talk about cross-sectional data or time series

data?

(b) Absolute values: The sum of the deviations from the mean for any data set

always equals zero. The sum of the absolute deviations from the mean does

better. But it is an awkward statistic to deal with. (For those with a little

calculus background, ask them to differentiate the MAD equation (2.3) with

respect to the mean, and the point will be established.) It is easier to deal with

the mean of the squared values. (Again, ask those with calculus background to

differentiate the MS equation (2.4)).

Chapter 2: Basic forecasting tools 15

(c) Mean square versus variance: There is a tendency for current statistics texts

to define variance using (n 1) in the denominator without explanation. Then

there is the inevitable question, Why (n1) and not n? When we write down

formulas for computing summary numbers (statistics) the formulas dont know

anything about assumptions, so from their point of view it makes no difference.

Only when a statistical model has been defined (using random variables with

their accompanying probability distributions) does it make any sense to talk

about unbiased or biased estimates, and then a meaning can be given to

the issue. To avoid some of the problems in this connection, we have chosen to

talk about the mean square as the simple average of squares, and the variance

will use (n 1) in the denominator. This is a matter of convenience. Since

this issue is tied up with d.f. (degrees of freedom), it is useful to make a point

that d.f. are defined as (number of independent data points) minus (number of

parameters estimated). For example, if there are n = 20 students in the class

and each has a score on a test, then how many numbers must you sing out to a

stranger before they know all the scores? Twenty. So there are 20 d.f. for the

raw data. Now compute the mean. Find the deviation from the mean for all

20 students. How many of these deviations must be sung out before a stranger

can know them all? Nineteen. After hearing 19 of them, the last one must be

such that they sum to zero. One d.f. was lost in computing the mean.

(d) Bivariate statistics: Covariance and correlation are important concepts to un-

derstand. Note that variance is a special case of covariance (if you convert Y

to X). Understanding correlation is important in regression (Chapters 5 and

6) and in autocorrelation (Chapter 7).

(e) Autocorrelation: To give students a feel for the meaning of autocorrelation,

show some plots of a time series plotted against itself, lagged one period, then

lagged two periods, etc. The physical act of making one or two of these scat-

terplots is worth the time it takes.

Another worthwhile activity is to do Exercise 2.4 in class with students guessing

which plots align with which ACFs. Discussion is usually easy to stimulate with

students justifying their answers.

(f) Accuracy measurements: Dwell on the interpretation of these summary num-

bers, so that the distinctions between absolute statistical measures and relative

measures are clearly understood. For example, in dealing with MSE versus

MAPE, you could consider the X and F data in Table 2-9.

(i) What happens to the values of MSE and MAPE if the data were all mul-

tiplied by 2?

(ii) Similarly consider what happens to the values of MSE and MAPE if 100

was subtracted from all numbers.

16 Part B. Teaching suggestions

In case (i) the MSE is going to be 4 times larger but the MAPE wont change.

In case (ii) the MSE wont change but the MAPE is considerably changed.

What happens if 119 is subtracted from all data?

(g) Theils U statistic: This is worth understanding. It becomes easy to use in

practice because when U < 1 the forecasting method is better than the nave

one (using last periods observation as the forecast for the next period). When

U > 1 it is worse then the nave method. Unfortunately, not many computer

forecasting programs calculate Theils U .

(h) ACF of forecast error: It is good practice to look at the ACF of the errors ob-

tained from any forecasting method. (See Section 7/1) It is worth emphasizing

this throughout Chapters 48 whenever a forecast is calculated.

possible because it provides a sense of how dependable a forecast is. Section 2/5

allows some simple forecast intervals to be computed using only the MSE. This

helps students understand the MSE better and introduces the basic concept of a

forecast interval.

6. The reason for giving a special section (2/6) to least squares is simply because it is

almost all-pervasive in the model building world. It has been found to be extremely

valuable in statistics (e.g., in the fitting of regression models to data).

some way. It is helpful to ask students how they would transform or adjust a few

given data sets. This can lead to some interesting discussion and emphasizes the

need to understand the data before doing any forecasting.

Decomposition methods are among the oldest of all the forecasting procedures. They are

easy to understand, at least in principle, because most people dealing with time series

data assume the presence of trend, the influence of a business cycle, seasonality (if were

dealing with monthly or quarterly data, for instance), and the ever-present noise (the error

term, the disturbance term, or the random shock). The following points should be noted

in teaching this chapter.

seasonality and noise (or other words to describe this uncontrollable part). when

we speak of trend it seems easy to understand, but in fact it is not all that clear. it

Chapter 4: Exponential smoothing methods 17

is often inextricably mixed up with the so-called cycle (which itself is not a math-

ematical cyclesuch as a sine wavebut rather a irregularly shaped up and down

movement associated with general business conditions) and the only way trend can

be separated from cycle is by arbitrary definition of trend. Because of this complex-

ity many decomposition methods (e.g., the Census II method) identify trend/cycle

as one component. Seasonality is less ambiguous and it refers to systematic patterns

that occur within the calendar year.

Suggestion: Have students come up with a written definition of these four compo-

nents.

2. The ratio-to-moving averages method is easy to compute (see Table 3-6) and it is

good to see plots of the original data, the moving average and the ratio-to-moving

average. When the ratio-to-moving average values are portrayed as in Table 3-6 they

can be visualized in a seasonal plot (Section 2/2/2) which allows for the stability

of the seasonal pattern to be assessed. Students should be clear on the meaning of

each of the columns in Table 3-6.

3. The Wall Street Journal, Business Week, Fortune and other business magazines all

make repeated reference to seasonally adjusted time series and it is important that

all students know exactly what this means. The ratio-moving averages are in fact

seasonal indices plus the random noise component. By averaging these seasonal

indices for each month (or quarter as the case may be), the random component is

reduced, and the resulting seasonal index is a measure of the impact of the season.

By dividing the original data by this seasonal index we are left with seasonally

adjusted datawhich has in it, trend and cycle and noise. That is, the influence of

the season has been removed. The Census II method talks about preliminary and

final seasonal adjustment factors (same as seasonal indices), and preliminary

and final seasonally adjusted series.

ing for students to compare the results. In particular, investigate what the methods

produce when the series contains some unusual behavior such as a level shift or

some outliers. The classical decomposition method is not designed to cope with

such behaviors, but the Census II and STL methods both contain some robustness

facilities.

time series forecasting, particularly the concept of forecasts being weighted averages of

18 Part B. Teaching suggestions

time-lagged observations. They are also useful forecasting methods in their own right.

because it is dynamic. It moves with time.

3. Whereas moving averages involve equal weights over a set of observations, the simple

exponential smoothing (SES) method is fundamentally different in that it implies

unequal (exponentially decaying) weights.

Aside: You can engage the students in a discussion on how to weight past data in

making a forecast. Should the latest data count more than earlier data? When is

this true? (E.g., when older data was based on a different manufacturing process.)

When might it not be true? (E.g., when current data occurs during a strike.)

4. In order to appreciate the fact that all methodologies have built-in limitations, it is

useful to do what engineers typically do, namely, test the methods on some standard

types of input series. This can be demonstrated by constructing a simple series of 20

observations containing a level shift part way through the series. Then apply both

a moving average and SES to the data to see how each method accommodates the

step. This can be a most enlightening experience for students and a valuable base

for latter work. Similar test series might contain a single outlier, or a trend.

5. Following the previous point, you can ask if SES and moving averages keep pace with

trend. And then discuss seasonality as a complicating factor. Can moving averages

and SES take care of seasonal indices?

6. Discuss Pegels two-way classification with the students to emphasize the difference

between linear and multiplicative trend and seasonality. The flexibility of Pegels

classification has yet to be fully appreciated, and it is worth discussing some of the

cells other than those corresponding to SES (A-1), Holts method (B-1) and Holt-

Winters method (B-2 and B-3). For example, consider cell C-3 which will often

outperform Holt-Winters method.

7. A very important point to establish for exponential smoothing methods is the fact

that an initialization process (for getting a method going) has to be defined. Since

the initialization procedure has an influence on all subsequent smoothed values it

has to be handled with care. Therefore, in deciding how well an SES model fits, for

example, it is wise to define a test period which excludes that early part of the

series which is still settling down during the initializing phase. Contrast this with

regression models where we can define errors of fit for all data points at once. For

this reason, we have given explicit statements about each strategy and some general

Chapter 5: Simple regression 19

are not the only ways of going about it. Students may be able to come up with there

own suggestions.

8. We have designed this chapter to be complete in the sense that the equations are

all given and fully worked examples are provided. Table 4-11 is something that you

might work toward. It gives in one place a comparison of how all the methods do on

one set of seasonal data. If your students have come to the course with their own

data sets they should be encouraged to work toward a table similar to this.

described in Chapter 11.

10. A good exercise in discussing ARRSES would be to ask students to generate a time

series using a slowly changing value, and then do an ARRSES analysis to see if

the method comes close to what was simulated.

Many students will have already have had a first course in statistics and will have done

some simple regressionmostly in the context of cross-sectional data. Chapters 5 and 6

of this text should accomplish the following:

sectional data.

(b) Discuss the importance and limitations of the correlation coefficient.

(c) Discuss the use of regression in a forecasting (time series) context.

(d) Deal with the practical application of simple regression, multiple regres-

sion and econometric models.

1. Discuss the data setup for simple regression, multiple regression and econometric

models. Mention that in econometric modelling the dependent and independent

variables become mixed upin the sense that Y variables appear on the right hand

side of econometric equations as well as on the left hand side.

2. Review the details of simple regression of Y on X, make sure everyone knows the fun-

damental statistics (mean, variance, covariance, correlation, regression coefficients),

20 Part B. Teaching suggestions

and then deal with the definition and role of the overall F test, the t-tests for indi-

vidual coefficients and the sampling fluctuation of the coefficients.

3. The correlation coefficient is a very widely used statistic and therefore should be un-

derstood well. Mention that it is a measure of linear association, that it is therefore

unaffected by any linear transformation, that its sampling fluctuation is large for

small sample sizes (so beware of those regressions based on 10 observations!), and

that it can be severely affected by skewness (or outliers).

4. Emphasize the difference between regression for cross-sectional data and for time

series data. Cross-sectional regression can be useful in a forecasting context (e.g., the

automobile data in chapter 2) but time-indexed data and time series regression pose

special problems. The error terms in cross-sectional regression are usually assumed

independent, but in time series regression this independence is often suspect, and

in some cases (e.g., in dynamic regression models) the errors are carefully defined

not to be independent. The autocorrelation coefficient is simple to define and to

compute, but its sampling distribution is more difficult to handle than the sampling

distribution of the correlation coefficient.

5. Many textbooks talk about regression as a forecasting tool but very few actually do

forecasting with regression. For example, if we regress Y t on Xt1 ,

Yt = 3 + 5Xt1 + (error),

and we want to forecast Yt+1 , then the equation allows us to do that:

Yt+1 = 3 + 5Xt .

We already know Xt , and so can obtain Yt+1 . However, if we regress Yt on Xt :

Yt = 3 + 5Xt + (error),

then in order to forecast Yt+1 we will need to know Xt+1 , and we do not know this.

So we will have to forecast Xt+1 before we can forecast Yt+1 .

6. Discuss the meaning of equations (5.3) and (5.4) for slope and intercept. The slope

is actually (covariance of X and Y ) divided by (variance of X) and the intercept is

the mean of Y minus the slope times the mean of X.

7. There are no assumptions involved in calculating a correlation coefficient. Equation

(5.10) for r is merely a formula. When it comes to regressing Y on X and some

statistical regression model is defined, then the correlation between X and Y , when

squared, has another useful interpretation. It is the proportion of variance explained

by the linear relationship between X and Y . What this means is that, knowing the

X values we will be able to recover a certain proportion of the variance of Y , and

this proportion is r 2 .

Chapter 6: Multiple regression 21

8. The r value is a measure of linear association so point out the message in Figure 5-7

when a strong nonlinear association cannot be picked up by r. Note also that for

small samples the correlation coefficient is notoriously unstable (a phrase Kendall

and Stuart use in the Advanced Theory of Statistics, Vol. 1). Finally, emphasise how

skewness can have a profound effect on r. The King Kong example (see Figure 5-8)

is a useful illustration of this effect and the answers to exercise 5.6 present further

evidence.

9. It is a good idea to contrast equations (5.13) and (5.14) and ask the question: Where

are the random variables in each equation? In (5.13) there is only one random

variable. In (5.14) there are three random variablesa, b and e. The values of

a and b are estimates for the unknown parameters, and in (5.13). This is why

we can define a standard error for the slope and a standard error for the intercept

standard errors which are needed to define t tests for the slope and intercept.

10. Emphasise that the F statistic involves a numerator degree of freedom and a de-

nominator degree of freedom, and make sure that students know how to read the

F tables (Appendix III, Table C). A pragmatic point is that the F test should be

done first when appraising a regression analysis, and afterward the individual t tests

can be examined. In the case of simple regression there is no difference, because the

t test is a special case of the F test, but in multiple regression (chapter 6) this is

important.

11. In simple regression there is an intimate connection between the slope and the in-

tercept. Since the least squares regression line always goes through the mean of X

and the mean of Y , it stands to reason that if the slope is changed the intercept is

changed and vice versa. If the mean of X and Y are both positive then an increase

in the slope will cause a decrease in the intercept, and vice versa.

Equations (5.17) and (5.18) should be studied. Note in (5.17) that the second term

under the square root sign will be small if the denominator (representing the spread

of the Xs) is large relative to the numerator (which is the mean of X). In (5.18)

the standard error of the slope depends on how spread out the Xs are. If they are

well spread out, the standard error is small.

1. As for many other chapters, it will be helpful here to have a readily available regres-

sion package for students to work on, so that they can check various things in the

chapter and can run their own data through various regression models.

22 Part B. Teaching suggestions

2. If at all possible, students should run their own data through the various analyses

for maximum understanding. We sometimes adapt the exercises at the end of the

chapter to use the data sets of interests to our students.

3. In multiple regression for cross-sectional data it is important to point out that the

significance of individual coefficients is contingent upon the other regressors present

in the regression. We emphasise this and say: the coefficient for X 3 is significant

in the presence of the other regressors.

independent variables and coming up with a reasonable model specification. To

illustrate this we have used a mutual savings bank data set and have done a detailed

analysis of most of the stages that led to a model which was actually used by a large

metropolitan bank. The example is sometimes a little complicated, but we feel it is

worth the effort to get into it in detail.

forecasting methodology, but seldom actually used explicitly in a forecasting context.

In this chapter we carry the bank study through to its conclusion by forecasting

with a final regression model. We explore the difficulties of having to forecast the

independent variables before we can forecast the dependent variable.

thoroughly, and then consider where the time series regression applications violate

certain assumptions. Since this text is not addressed to formal statisticians, it is

enough to discuss the implications of correlated errors, improper specification (e.g.,

linear when it might be curvilinear, or two regressors when it should be four regres-

sors), multicollinearity, etc., and refer interested students to texts such as Draper

and Smith (1981) for more details.

7. Table 6-1 and Table 6-2: Take time to get to know these data sets because they will

be used a lot during the chapter. Have students graph the data sets and keep them

handy for class discussion.

8. The Durbin-Watson statistic is described in equation (6.9) and Table 6-7. Students

dont have too much difficulty learning how this statistic is computed. However,

learning to use the D-W tables (Appendix III, Table F) is not so straightforward.

Please spend time going through a couple of examples in using the tables.

9. Selecting variables for inclusion in regression is a meaty subject and we give only an

introduction to the major ideas. In the context of the bank example we go through

some of the procedures without explaining all the details (for example, we talk about

Chapter 7: The Box-Jenkins methodology for ARIMA models 23

using principal components to get a short list of variables). However, any serious

multiple regression analysis will need to consider variable selection carefully.

10. Section 6/4 (Multicollinearity) gives some information that is not often given about

multicollinearity. We hear too often that multicollinearity is present when the

highest correlation among any pair of regressors is only .7, say. And we hear too

often that multicollinearity is not a problem when there are no large correlations

(i.e., not larger than say 0.5). Both of these statements are incorrect. Table 6-12

shows quite clearly that even when the correlations among regressors never get bigger

than 0.333 we can have perfect multicollinearity.

11. Standard error formula (6.13) is a multivariate equivalents of (5.19). It is a little

harder to interpret because it is written in matrix notation, but it should be part of

a regression package so that confidence intervals can be determined.

12. Table 6-14 should be studied very carefully to ensure students understand how these

forecasts are obtained. It takes a little while to get the time intervals straight, but

its a real issue.

13. In discussing econometric models (Section 6/6) we have only given an introduction to

how econometric models are related to the multiple regression models which are the

subject of this chapter. Our aims are to give students an appreciation for econometric

models, their breadth and depth, and the need for specialized skills to develop and

use them effectively.

The topic of econometric modelling is itself an extensive field and we have not chosen

to cover it in this book. An instructor may choose to include other materials on

econometric methods (such as Johnston, 1984; Judge et al., 1988; or Pindyck and

Rubenfeld, 1991) to complement the materials in this chapter. A useful introductory

perspective is provided by Aykac and Borges Econometric methods for managerial

applications in the Handbook of Forecasting, Makridakis and Wheelwright (editors),

(New York: Wiley and Sons, 1982).

1. We have chosen to present this material in a different order from that found in most

other textbooks. Students always find this material a little difficult at first, and we

have found the order given in the textbook the most successful approach in leading

students through ARIMA modeling.

2. Section 7/1 allows students to firmly grasp the idea of white noise and the use of

the ACF and PACF before considering ARIMA models. The white noise tests can

24 Part B. Teaching suggestions

Introducing residual analysis in the context of these earlier forecasting methods

emphasises that these ideas are not only applicable to ARIMA models, but to any

forecasting methodology. It also allows students to become familiar with some of

the tools used in ARIMA modelling before having to learn about ARIMA models

themselves.

3. Next we introduce the ideas of stationarity and differencing in Section 7/2. We find

it better to introduce these ideas before ARMA models (rather than after as is often

done) because it allows ARMA models to be applied to a much wider range of time

series from the start. A common approach is to consider only stationary series at

first and non-stationary series later, but this gives students the initial impression

that ARMA models are not very widely applicable. We find that the approach given

in the book leads students to be more positive about these very useful models.

4. You can anticipate that we will be wanting to use the backward shift operator nota-

tion later (particularly in chapter 8) and should make a decision whether to introduce

it at this time or not. Students do not seem to find this difficult to digest.

5. Students usually find autoregressive models relatively easy to understand as an ex-

tension of multiple regression. This is the way we normally introduce them. Next

consider the connection between exponential smoothing and AR processes. An ex-

ponential smoothing process also involves weighted past values and so is a special

case of an AR process.

6. Moving average models are usually more difficult to understand at first. Some in-

structors try to connect them with previous forecasting methods such as exponential

smoothing. We have not found that students find this particularly helpful. Instead

we just say they are like a multiple regression, but with past errors as the explana-

tory variables. Get the point across that linear functions of past values of the error

series are called MA processes, and linear functions of past values of the observations

are called AR processes.

7. Note the potential confusion between moving average models, moving average

smoothing and moving average forecasting. Students find this unfortunate dupli-

cation of terminology difficult and it needs to be explained very carefully.

8. If you have the facilities it is recommended that you have students generate time

series using some of the simpler models so that they really know what is implied.

Students will learn more about the models if they have to generate data using a

spreadsheet package, than if they use a package with built-in data generation facil-

ities. Generating data with known properties and then studying the shape of the

theoretical ACF and PACF gives a valuable insight into the BJ models.

Chapter 7: The Box-Jenkins methodology for ARIMA models 25

Make a point about learning in one direction and analyzing real data in the opposite

direction. In schematic form, this is:

(b) study the theoretical ACF and PACF

(c) store the simulated data and then analyze it.

See if the empirical ACF and PACF match the

properties that we started with.

(b) study the empirical ACF and PACF

(c) try to identify a theoretical underlying model

that could have given rise to the observed data

9. Note that the constant term c in these models is not the same as the mean of the

time series if there is an AR component. For example, if the mean of Y t is , then

Yt = (Yt1 ) + et

c = (1 1 2 p ).

10. It will become increasingly important for students to be able to write out the equation

for any ARIMA model, so practice at this time is important.

11. The ACF and PACF are used repeatedly later Sections of this chapter, so they

should be learned thoroughly. Give students some sample ACFs and PACFs and

have them guess the type of series they come from.

12. Before proceeding to Section 7/5, it would be wise to have students very comfortable

with the ideas of stationarity, ACFs, PACFs, ARIMA models and how they are all

related. They should be able to write the equations for simple ARIMA models.

13. In trying to generate time series that are from an ARIMA(p, d, q) process, students

will need to be aware of the restrictions on the values that the AR and MA coefficients

can take on.

26 Part B. Teaching suggestions

14. Students should not be fooled into believing that it will be easy identifying ARIMA

models for real data series. As soon as the model becomes a mixed modeleven

the very simplest ARIMA(1, 0, 1)the shape of the ACF and PACF can become

confusing. It is good to lean the properties of the elementary models, and it is good

to remember that they will seldom make themselves known unequivocally in real

data series.

15. Estimation (Section 7/4) is often taught in two parts: preliminary estimation and

final estimation (using some iterative process such as Marquardts algorithm). Com-

putationally, this is the way it must be done, but from the point of view of a forecaster

using a computer package, the computational details are not relevant. Therefore, we

have focussed on the results which are obtained from a computer package as these

are of most relevance to practising forecasters.

16. The use of the AIC in Section 7/6 is not common in introductory forecasting books,

but we have found it extremely useful in practice and many computer packages are

not giving it as part of the standard output.

17. The conversion of an ARIMA equation into a form suitable for forecasting (Section

7/8/1) takes a little bit of algebraic multiplication and rearrangement of terms.

However, since all forecasting packages will provide forecasts from an ARIMA model

automatically, students will probably never need to do the calculation themselves.

The purpose of including the details in this section is to show how the equations give

forecasts, something which may not be immediately obvious to students, particularly

when there is an MA component.

18. The material in Section 7/8/3 is very poorly understood, even by some experienced

forecasters. The effect of differencing on the forecasts is worth understanding. Too

often differencing is carried out without thought for its implications later on.

19. Students learn most from working through an analysis of a time series from start

to finish. Exercises 7.8 and 7.9 are useful for this purpose. See also Section C/2/7

which can be used as a student project in longer forecasting courses as it enables

each student to choose a different set of data to analyze.

1. There is a lot of material in this chapter, and the instructor may wish to select

only a few topics to cover. In a shorter course, we suggest omitting Section 8/5

(Multivariate autoregressive models) and Section 8/6 (State space models). Please

note that Sections 8/1, 8/2 and 8/3 are sequential. Therefore it is important to

Chapter 8: Advanced forecasting models 27

cover Section 8/1 (Regression with ARIMA errors) well before going on to Sections

8/2 (Dynamic regression models) and 8/3 (Intervention analysis).

2. The approach we have adopted for Sections 8/1 and 8/2 is very different from that

found in most books. We have followed the Pankratz approach (and terminology)

to modelling rather than the more traditional Box-Jenkins approach. We have

taught using both approaches many times and have found students find the Pankratz

approach very much easier to follow and use. In our own consulting work, we have

also found it a much simpler methodology when fitting dynamic regression (transfer

function) models.

3. It is essential when considering Sections 8/2 and 8/3 that students are comfortable

with the backshift operator notation.

4. Unfortunately, there are not many software packages which allow the range of models

covered in this chapter to be fitted. We have taught the material to a range of

students using click-and-point interface available in the SAS Forecasting system. It

provides particularly good facilities for dynamic regression and intervention models.

We have had most success in teaching this material through case studies with the

students spending most of the class time doing the analysis on PCs.

5. After studying Sections 8/1 and 8/2, have the students find examples in the real

world where one input variable influences another variable dynamically over future

time. One of our students came up with three series that seemed to go round robin:

monthly gas prices, domestic autos produced and autos sold.

6. For intervention analysis, a good class project is to ask students to read an article

involving the application of intervention analysis, then prepare their own report or

oral presentation on what was done. We have done this with the Ledolter and Chan

(1996) article with good results.

7. For Sections 8/4 through 8/7, we only provide a brief introduction to the ideas in-

volved, plus some applications. Our aim here is to provide students with enough

information to know when these models might be applicable and in what circum-

stances they might be useful. If students wish to use these models, they will need

to learn much more about them than is described in our book.

which make use of one or more of the methods covered in this chapter. Each student

can give a brief presentation based on one application and lead a discussion on

whether the model was appropriate to the problem.

28 Part B. Teaching suggestions

1. The first thing which it is important to bring across to the students is long-term

mega-trends. A good way for doing so is to present Figure 9-1 and ask if the data

presents a trend (most students will say No), then Figure 9-2 asking the same

question. At this point several students would say that 14 years is enough to establish

a long-term trend. One can then show Figure 9-3, which indicates that the 14 years

of Figure 9-2, shown as a shaded region of Figure 9-3, are merely a small part of

Figure 9-3. Finally, one can show Figure 9-4 and discuss why mega-trends can only

be established by going back to the beginning of the Industrial Revolution (that is

around 1800). Another figure which can be used to help illustrate this starting date

as well as the persistence of mega-trends is Figure 9-6 which shows wheat prices in

constant and goes back to the middle of the 13th century.

2. Once long-term mega-trends have been identified they can be extrapolated unless we

believe that they will change due to some other revolution similar to the industrial

one. If that is the case then we have to make our predictions not by extrapolation

but by using analogies or by making various scenarios about the implications of large

changes like those of the forthcoming Information Revolution.

3. When forecasting for the long term, deviations around the long-term mega-trends are

of critical importance as cycles can last for many years or even decades. Moreover,

since cycles are mostly random walks we have to go beyond pure quantitative models

to predict them. This is a point worth making and can be illustrated by generating

random numbers, cumulating their effects, and showing the result on a graph. Such

graphs show that predicting turning points is impossible quantitatively since they

present random walks.

4. Chapter 9 has a lot of figures that usually generate a great deal of interest from

students. The way to present them is by discussing the implications if they are

extrapolated in the long run, ending up with the question of what will happen when

our buying power increases (at a double rate since real prices drop exponentially and

real income increases exponentially) and we get a situation of over-abundance, while

at the same time huge inequalities between rich and poor nations, and rich and poor

citizens in single nations.

5. The discussion about such implications, as well as those that would come up by

talking about various analogies and scenarios between the Industrial and forthcoming

Information Revolution, generate great interest and strong opinions which provide

the basis for a lively debate.

Chapter 10: Judgmental forecasting and adjustments 29

1. One way of introducing the topic of judgmental forecasting is to give Figure 10-4 to

the class and then ask them to make forecasts after consulting the figure. Tell one

third of the people that the product shown in the figure is mature, the second third

that it is old, and the final third that it is new. The results of their forecasts can

be summarised and presented. They usually are similar to those shown in Figure

10-5 which indicates how pre-conceived ideas are being used and how they can bias

the forecasting process (after all it can be indicated that the great majority of new

products fail after a couple of years).

2. Large errors in judgmental forecasts can also be illustrated by comparing the per-

formance of professional investment managers to those of randomly selected stocks.

The consistent under-performance of expert managers is remarkable and can be used

as the basis to discuss what is wrong as well as illustrating how one can improve

investment returns without having to pay any fees to professional experts by simply

selecting bond, stocks and other investments randomly. In addition one can discuss

why people prefer experts to manage their investments (obviously, they feel more

secure by doing so, or alternatively they think that they reduce their uncertainty)

while clearly such a choice results in smaller returns and extra fees.

3. Another interesting topic in Chapter 10 is the use of decision rules instead of intuitive,

global judgment when the judgmental inputs can be quantified. Again there is

a lot of material for interesting discussion starting with the finding that decision

models in the form of multiple regression equations can predict more accurately

the performance (their average GPA) of candidates for universities than admissions

officers. This and similar types of decision rules can therefore be discussed as ways

of improving future oriented decision-making.

4. The last part of this chapter deals with ways of debiasing decision making so that

the advantages of both quantitative models and judgment can be exploited while

avoiding their disadvantages.

5. The following is a list of judgmental exercises (there are two versions: one to be

given to half the class and the second to the other half). These exercises provide

an excellent way to show the students their biases as their answers from the two

versions vary considerably.

30 Part B. Teaching suggestions

Judgmental exercises

1. What is the percentage of countries in the UN that are African? To make your

estimate, I would suggest that you start with a value of 65% (this percentage was

found in the computer by generating a random number between 0 and 100). First

decide whether this value is too high or too lowthen move upward or downward

from that value to what you feel is the true value.

Your final estimate as to the true percentage of African nations in the UN is

.

2. A psychological test was administered to a group of 100 people. The group consisted

of 30 engineers and 70 lawyers. The following descriptions were obtained for Peter

Jones:

Peter Jones is of high intelligence and exhibits a strong drive for compe-

tence. He has a need for order and clarity and for neat and tidy systems in

which every detail finds its appropriate place. His writing is enlivened by

somewhat corny puns and by flashes of imagination. He seems to have lit-

tle feel and little sympathy for other people and does not enjoy interacting

with others. Self-centred, he nonetheless has a deep moral sense.

If you had to place a bet on whether a participant in the test named Peter Jones

was an engineer or a lawyer, what would you say?

Peter Jones is an engineer .

Peter Jones is a lawyer .

Please put a cross on the appropriate line.

3. You are the chief executive officer of a company faced with a difficult choice. Because

of worsening economic conditions, 12,000 people will need to be fired to reduce the

payroll costs and avoid serious financial problems. Two alternative programs to

combat the firings have been proposed to you. The estimates of the consequences of

the programs are as follows:

If program B is adopted, there is a two-thirds probability that no jobs will

be saved and a one-third probability that 12,000 jobs will be saved.

Which of the two research and development projects would you select?

A B

Please tick the appropriate box.

Chapter 10: Judgmental forecasting and adjustments 31

4. The figure below shows the sales of Electrack, a video game produced by Jeu-

tronics, a medium sized French toy company. Provide optimistic, most likely and

pessimistic forecasts for the year 2001.

250000

200000

150000

Year Sales

1993 4,433

100000

1994 60,298

1995 67,884

50000

1996 89,512

0

Pessimistic .

Most Likely .

Optimistic .

5. You are in a store about to buy a new watch which will cost 350FF. As you wait for

the sales clerk, a friend comes by and tells you that an identical watch is available in

another store two blocks away for 200FF. You know that the service and reliability

of the other store are just as good as this one. Will you travel two blocks to save

150FF?

COMBINED WITH THE EXPERIENCE OF YEARS

Please indicate the number of Fs which appear in the above sentence. .

How confident are you of your above answer? Indicate your confidence on a scale of

0 to 100 with 0 indicating no confidence and 100 indicating full confidence. .

How many times did you read the sentence FINISHED . . . OF YEARS? .

32 Part B. Teaching suggestions

Judgmental exercises

1. What is the percentage of countries in the UN that are African? To make your

estimate, I would suggest that you start with a value of 10% (this percentage was

found in the computer by generating a random number between 0 and 100). First

decide whether this value is too high or too lowthen move upward or downward

from that value to what you feel is the true value.

Your final estimate as to the true percentage of African nations in the UN is

.

2. A psychological test was administered to a group of 100 people. The group consisted

of 30 engineers and 70 lawyers. If you had to place a bet on whether a participant

in the test named Peter Jones was an engineer or a lawyer, what would you say?

Peter Jones is a lawyer .

3. You are the chief executive officer of a company faced with a difficult choice. Because

of worsening economic conditions, 12,000 people will need to be fired to reduce the

payroll costs and avoid serious financial problems. Two alternative programs to

combat the firings have been proposed to you. The estimates of the consequences of

the programs are as follows:

If program B is adopted, there is a one-third probablity that no nobody will

be fired and a two-thirds probability that 12,000 people will be fired.

Which of the two research and development projects would you select?

A B

Please tick the appropriate box.

4. The figure below shows the sales of Electrack, a video game produced by Jeutron-

ics, a medium sized French toy company. Figure 2 shows the most likely predictions

from a widely-used computerized mathematical model for new products. After hav-

ing looked at Figures 1 and 2, provide an optimistic, most likely and pessimistic

forecast for the year 2001.

Chapter 10: Judgmental forecasting and adjustments 33

Year Sales

1993 4,433

1994 60,298

1995 67,884

1996 89,512

Figure 1: Actual sales of Electrack Figure 2: Actual and predicted sales of Electrack

250000

150000

200000

100000

150000

100000

50000

50000

1993 1995 1997 1999 2001 1993 1995 1997 1999 2001

Pessimistic .

Most Likely .

Optimistic .

5. You are in a store about to buy a new video camera that costs 4000FF. As you

wait for the sales clerk, a friend comes by and tells you that an identical camera is

available in another store two blocks away for 3850FF. You know that the service

and reliability of the other store are just as good as this one. Will you travel two

blocks to save 150FF?

COMBINED WITH THE EXPERIENCE OF YEARS

(Please do not read the above sentence again)

Please indicate the number of Fs which appear in the above sentence. .

How confident are you of your above answer? Indicate your confidence on a scale of

0 to 100 with 0 indicating no confidence and 100 indicating full confidence. .

34 Part B. Teaching suggestions

1. The material of Chapter 11 relates to that of Chapter 10. The major question which

arises at the beginning of the chapter is the choice between judgmental and statistical

forecasting methods. As the quote by Sanders and Manrodt explains

Like past investigations (surveys) we found that judgmental methods are the

dominant forecasting procedure used in practice.

This means that there is a great deal of potential improvement as managers now

realize the potential for higher forecasting accuracy (and therefore reduced costs)

and they are continuously pushed, at the same time, to operate more efficiently and

effectively in order to reduce their operational cost. Thus the time is right to per-

suade them of the considerable benefits they can obtain by using the knowledge and

experience we have accumulated which clearly indicates the benefits from statistical

forecasting and how it can be best integrated (this is also the major topic for the

next chapter) with judgmental predictions.

2. The findings in Section 11/2 are very useful and relevant in putting forecasting

on a practical grounding. In this section, the major empirical evidence is being

summarized and various tables and figures are available for backing up the findings.

The major question then for forecasting users is to select the most appropriate

method for their specific situation. This topic is discussed in Section 11/3. It is

worthwhile for the instructor to present each one of these factors and discuss them

in class. Obviously a critical factor is the last one (the number and frequency

of forecasts). Such factor signifies the necessity for simple methods that can be

completely automated when the number of forecasts required is very large and they

are needed frequently.

3. In the absence of clear factors, when guidelines are not obvious or in case of doubt

as to what method to select, the best alternative is to combine three or four simple

methods and use their average as a way of predicting the future. As is well known,

through many empirical studies, such a simple average of the combined forecasting

methods is both more accurate than the individual methods being combined while

at the same time variance of the forecasting errors of combining is smaller than that

of the individual methods involved. Combining can, therefore, be presented as a

practical alternative which improves forecasting accuracy and reduces the chances

of errors (in particular large ones).

Chapter 12: Implementing forecasting: its uses, advantages, and limitations 35

limitations

in other words present and discuss the limits of predictability. This can be done in

terms of short-, medium- and long-term predictions as the horizon of our forecasts

presents different challenges and problems as far as predictability is concerned. Crit-

ical in such a discussion is the medium-term which must predict the ups and downs

of business cycles. Such predictions are extremely difficult and present a major chal-

lenge for businesses when they attempt to make budget estimates. The same is true

for the longer run (18 months to 5 years) when predictions for the five years Business

Plan are made and when longer term cyclical deviations around the long term trend

must be dealt with. The topic of predictability, or the lack of it, can be related to

the introductory chapters of the book and to our experience from the forecasting

practice (including the findings from the surveys among forecasting users presented

in the previous chapter).

2. The second topic of Chapter 12 deals with the organisational aspects of forecasting

and the need to deal with the various forecasting problems that are encountered in

organisations which are using forecasting methods and the possible solutions to such

problems. There is enough information in the corresponding part of Chapter 12 to

describe these problems and discuss suggestions for solving them satisfactorily.

discusses the role and value of forecasting beyond its operational applications. As the

title implies, its greatest value is when the forecasts are creative in nature, which

by definition means that they cannot be based on simply extrapolating historical

information. On the contrary it may be necessary to go against conventional wisdom

in order to come up with creative insights about future changes or what the future

might hold.

4. In the last part of this chapter the instructor can discuss and possibly develop

his/her own ideas about how forecasting is going to evolve in the future. Central

to such a conception will have to be the creation of a learning process which will

result in organisational learning (rather than the experience of each individual person

concerning forecasting resting with such person and disappearing when he or she

changes jobs or company). Creating learning about forecasting is more practical

and cost efficient these days through the use of groupware (or intranets) which

allow the people in organisations working on forecasting not only to exchange ideas,

information and inside knowledge, but also to record the forecasting process they

have been using as well as their successes and failures so that they can be reviewed in

36 Part B. Teaching suggestions

the future by themselves or others in ways that can enhance learning. In other words

ways must be developed which can help organisations to improve their forecasting

process by knowing and avoiding past mistakes while using practices that they have

been found to be successful in the past.

C/Additional materials for teaching

forecasting

This chapter suggests additional materials that might be used to complement the contents

of Forecasting: Methods and Applications, 3rd ed., in a teaching situation.

Since cases can be a valuable addition to a forecasting course and yet often represent

a very different style of teaching from straight lectures and problem sets, Section C/1

outlines ways in which cases can be effectively integrated into the teaching of forecasting.

Section C/2 provides some suggestions for special project assignments. These provide a

context for forecasting but are shorter than field-based cases. Lastly, Section C/3 consists

of exam questions.

Many different types of cases can be used to meet very different purposes in teaching.

Briefly, these can be grouped into three categories. The first would be case exercises in

which the case is simply an expanded problem (that is, what was traditionally described

as a story problem) providing data and their context for the students to use in applying

a specific tool or technique being covered in a forecasting course. The second type would

be the management process case in which the case describes the forecasting process in

an organization and allows students to consider the process itself as opposed to specific

techniques applied to specific data sets. The final category would be a mix of these two

but generally oriented toward applying a technique to specific data and then looking at the

management decision-making implications of the resulting forecasts. This third category

is closely tied to the implementation of what the forecaster recommends to management.

A case course is taught using a problem-centered, participant-involved method of in-

struction. For most class sessions in such a course, a case is assigned to be read and

prepared for discussion and analysis in the classroom. Each case describes a specific man-

agement problem and seeks to describe selected aspects of an everyday situation that

either the forecasting specialist or the management user of forecasting might encounter.

37

38 Part C. Additional materials for teaching forecasting

In spite of the realism that an instructor seeks to build into all cases, they cannot be

completely true-to-life management situations for the following reasons.

managers and forecasters must gather facts in ongoing situations from memos, con-

versations, statistical reports, and the public press in a much less organized fashion.

2. A case is designed to fit a particular unit of class time and to focus on a certain

category of problem, such as a forecasting technique, forecasting system, or the

management use of a given forecast. Consequently, it may omit elements of the real

situationpeople or organizational issues for examplein order to focus attention

on what the instructor would like the class to see.

form a continuum requiring some action today, further consideration and action

tomorrow, and so on. It is very seldom that a manager can wrap up his or her

problems, put them away and go on to the next case as is done in a course.

4. While students studying cases are required to make decisions, they do not have the

responsibility for implementing those decisions and do not have to bear the burden

of ineffective implementation. This can be a particularly important shortcoming

when training forecasting specialists, who will have to interface with managers, who

in turn must make decisions based on their forecasts.

Cases can help forecasters and managers sharpen their analytical skills by exposing them

to facts and figures which must be evaluated and used to produce both quantitative and

qualitative evidence to support recommendations and decisions. In case discussions stu-

dents are typically challenged by instructors and peers to defend their arguments and

analyses. This can have the cumulative effect on the students of helping to develop a

problem-solving methodology and heightened ability to think, reason, and apply specific

techniques in a rigorous fashion.

Case studies cut across a range of organizational situations and provide exposure to

a far greater number of situations than would be likely on a single job involving normal

day-to-day routine. Thus, cases permit building knowledge across a range of subjects and

situations by dealing selectively and intensively with problems in each field. Students

come to recognize that the problems they face as a manager or forecaster are not unique

C/1 Using cases in teaching forecasting 39

to one organization or even a system of organizations. This helps them to develop a more

professional sense of their tasks and the way in which they can be handled most effectively.

Cases and the related class discussions can provide the focal point around which the

students past experience, expertise, observations, and rules of thumb can be brought

together in a framework for effectively tackling new situations. What each class member

brings to identifying the central problems in a case, analyzing them and proposing solutions

to them, is as important as the content of the case itself. The lessons of experience

can be tested as students present and defend their analyses against those of participants

with different experiences and attitudes. This is one place where common problems,

interdependencies, differences of perception, and organizational needs can be highlighted

and resolved in a systematic fashion.

An important benefit of using cases is that they help students learn how to ask the

right questions. It has often been said that 90% of the task of a good manager is to ask

useful questions. Answers can be relatively easy to find once the appropriate questions are

asked. Even when assignment questions are used with individual cases, students should

be pressed to ask themselves, What are the real problems that the individual forecaster

or manager must resolve in this situation?

One final benefit that an instructor often seeks to achieve through the use of cases is

to transfer to the students the excitement and challenge that can come from pursuing

management and forecasting careers. In the cases they prepare, students often see some

problems they are glad they do not have to face in real life, and others that they recognize

from first-hand experience. They should come to recognize that being a manager or

working as a forecaster for a manager can be a great challenge intellectually, politically,

and socially.

There is no single form of case preparation that works best for everyone. However, some

general guidelines can be offered that might well be adapted to the way each individual

student does his or her work. These guidelines would include the following.

1. Go through the case almost as fast as you can turn the pages, asking what the case

is about and what types of information are being provided for analysis.

2. Read the case very carefully, underlining key facts and perhaps writing those and

key issues in the margin. The students should try to put themselves in the position

of the manager or forecaster in the case and develop a sense of involvement in that

persons problem.

40 Part C. Additional materials for teaching forecasting

3. After a thorough reading of the case its often useful to prepare a list of the key prob-

lems. The case can then be gone through once more, picking up those considerations

and data that are relevant for resolving each of those problem areas.

4. Perform the analysis that will enable the student to conclude what recommendations

are most consistent with the situation and the facts provided.

5. Develop a set of recommendations that can be supported by case data and the

students analysis, and indicate how those recommendations should be implemented

in this situation.

These five steps can best be applied by the student working individually. The next step

which can aid in case preparation is to have students meet in discussion groups to present

their arguments to others as well as to listen to the arguments of their peers. This testing

of the students analysis and recommendations is an important preparatory step for class

discussion. The purpose of the group discussion is not to develop a consensus or a group

plan of action, but rather to help each member refine, adjust, and amplify their own

thinking. It is not necessary or even desirable that the discussion group members agree.

In class the instructor will usually let the class direct the discussion toward those topics

where most of the individuals have concentrated their attention. However, the faculty

member is also likely to prod the class to explore fully those avenues that are most relevant

based on the faulty members experience and based on the purposes for which the case was

included in that course. Often the faculty member will summarize the discussion and draw

out the useful lessons and observations toward the end of the class discussion. However,

this might also be done by asking a student to provide that type of summary. It should be

emphasized that learning through the case method results from rigorous discussion and

controversy. Each member of the class and the instructor must assume responsibility for

preparing a case and for contributing ideas to the class discussion.

Just as there are many types of cases, there are many purposes for which they may be

used in course design. Perhaps the simplest is to select exercise cases that can be used

to present data for the students to use with various forecasting techniques. Many of the

cases that will be suggested later in this chapter are of this type and, thus, can be used

simply as exercises or work problems for students.

A second way in which cases are often used by instructors who do not teach largely by

the case method is as a way to get at implementation and management decision-making

issues. In such a use of cases, an instructor might choose to cover a topic such as regression

C/1 Using cases in teaching forecasting 41

analysis using a more traditional lecture method with exercises and problems and then

conclude that section of the course with one or two classes built around management

process-oriented cases. These can be viewed as a way for students not only to apply the

techniques related to regression that they have learned but to look at the implications

of those applications for managers and to consider how they might be effectively sold

to management and implemented. If this were to be the only use of cases the instructor

might simply choose to end each of four or five major sections of the course with one or

two cases, resulting in a course that is approximately 80% exercises and lecture/problems,

and 20% case applications.

A third way in which cases often have been used effectively in teaching forecasting and

planning is to teach the basic techniques and their applications using exercises and prob-

lems and then at the every end of the course to have a major section on implementation.

In that section, cases requiring the use of different techniques and illustrating the range

of management issues related to implementation could be addressed. If this approach is

followed a class of thirty sessions might have only the last four or five built around case

studies of the management process and implementation type.

Still a fourth approach to utilizing cases in this subject area would be to build the entire

course around cases. While this is certainly feasible given the amount of material readily

available, this can often be a most challenging task if the students are not used to case

courses from other parts of their curriculum. The authors experience would suggest that

it is best to use one of the foregoing forms of case use initially, before building an entire

course around cases.

When students are not particularly well versed in the case method, it has often been

found effective to assign study groups to meet in preparation for the classes in which cases

will be used and then to have two or three individuals briefly (5 minutes each) present

their recommendations and analysis at the start of class in order to get the discussion

going. That provides a complete set of thoughts and ideas on the case situation that the

rest of the class discussion can build on. It also ensures that students are well prepared for

that class since they know they might be called on to make such a starting presentation

Prepared cases with teaching notes are available from Harvard Business School Publishing,

60 Harvard Way, Box 230-5A, Boston, MA 02163. They can also be obtained through the

internet at

www.hbsp.harvard.edu

These can be reprinted by the instructor and used to complement the exercises in the text

itself.

42 Part C. Additional materials for teaching forecasting

C/2 Assignments

The Phrygian Thread Factory was founded in 1947 by Ikos Matzakis shortly after

emigrating from Greece. The enterprise had begun on a small scale, supplying thread for

the local garment industry. In these days, Matzakis would buy cotton fiber from relatives

in Greece, import it to the United States, and dye it and spin it produce a rather wide

range of end products.

Since those humble beginnings Phrygian had grown to become a not inconsequential

thread supplier for the Northeastern United States. In addition to supplying the garment

industry and various distributors and retailers of sewing threads, Phrygian was now sup-

plying large industrial users. Major customers included the auto industry (for upholstery

and seat belts) and the telecommunications industry (for wrapping and insulating cables)

Similarly, Phrygian no longer restricted itself to cotton thread. The bulk of its output

was now nylon, although significant amounts of rayon, cotton, and silk were produced

as well. Phrygians product line was virtually unlimited, for it was standard operating

procedure to do custom dye jobs to match customer color specifications. However, color

notwithstanding, there still remained nearly a hundred distinct items in the product line.

One of Phrygians most important products was NC-216. This was a bonded nylon

thread customarily used by the auto industry in sewing seams in upholstery. To make it

Phrygian began with raw nylon fiber of weight 210 denier. Two strands were spun together

with a right-hand twist to form a thread. Then three threads were twisted together, again

with a right-hand twist, to form the final thread. Once this was ready, it was loosely

wound into large spools and sent to the dyehouse.

The dyehouse staff would dye the thread into batches of up to ninety pounds. From the

dye vats, the thread would go directly to large walk-in ovens to accelerate drying. After

2448 hours in the ovens the thread was moved to large drying rooms to finish drying.

After one to five days in the drying room, the thread was ready to be sent upstairs for

bonding.

In the bonding room the thread was passed slowly through a hot liquid plastic solution

and then through heaters and on to winders. Once this process was completed, random

samples were taken and tested, primarily for breaking strength. The thread was finally

sent down to the spooling room to be put on customer-specified spools (usually one pound

spools). Once finished, the completed order was sent down to shipping to be packaged

and shipped

C/2 Assignments 43

Other products followed the same general flow, although some required additional proce-

dures (such as skeining before dying) and others required fewer (for example, no bonding).

In the office suite life was characterized by a constant effort to track down and expedite

orders. Orders were phoned in by salesmen in the field. In the case of custom color

requirements, color samples followed by mail. For most orders, salesmen, wanted to know

a projected delivery date. The production manager, Roy, and his assistant, Fred would

characteristically supply delivery dates off the top of their heads. In this process they

relied on their intelligence guided by experience. For all products they were aware of the

normal production time. They also knew that these lead times were quite flexible. With

constant monitoring, a product could be shipped in a much shorter time than its expected

processing time. However, with no monitoring a product often took much longer than the

normal lead time.

For very important orders, Roy and Fred would promise an early date and then ride

the department managers closely to make sure the date was met.

Other aspects of production control were done in the same sort of ad hoc manner.

Workforce levels, overtime, and extra shifts were decided on pretty much a day-to-day

basis. Bernie, the new plant manager, had decided things had to change. This decision

had been made in response to the latest catastrophe. Phrygian had just received a large

rush order from Non-Specific Motors for three thousand pounds of NC-216, a thousand

pounds in each of three colors. Bernie was at first jubilant when he heard of it. But his

jubilation was short-lived when he learned that there was not enough 210 denier nylon to

meet such a large order. There was already additional nylon backordered but it was not

scheduled to arrive in time to be of use for the Non-Specific order.

While Roy and Fred scrambled orders, robbing Peter to pay Paul, and combed the coun-

tryside for additional supplies of 210 nylon, Bernie sat in his office and plotted strategy.

Bernie decided the first requirement was a good forecasting system so they would not be

caught off guard like this again. By going over orders for the last three years, he noticed

that each year Non-Specific had placed a large rush order for NC-216 at about this same

time. He felt sure such information could be used in planning operations. He decided to

call Roy and Fred in and have them set up a forecasting system.

The next day Bernie made his pitch. Look, you guys, things have been going pretty

well. Phrygians profits are up, even our market share is up. But now were getting bigger

and I think weve just about hit our breaking point. Last year you hired Fred, Roy. That

took some of the load off you, but already its getting ahead of you again. We are getting

more customer complaints about orders being shipped late. And now weve gotten caught

short on 210 for the Non-Specific order.

Look, Bernie, were getting burdened on this I know. But its the first time this has

44 Part C. Additional materials for teaching forecasting

January 1340 3690 4110 4500 2600 5330 5140 6820

February 1500 3520 3870 4290 5830 5290 4900 6540

March 1570 3330 3550 4010 5400 4960 4400 6030

April 1360 3120 3420 3830 4210 4730 4090 5770

May 1350 2880 3250 3570 3900 4370 4600 c 5510

June 1400 2670 2910 3250 3640 4020 4540 5000

July 1610 2790 3080 3520 4010 4020b 4930 5430

August 2280 3540 3890 4280 4830 4830 5920 6520

September 2730 3920 4310 4830 5270 4880 6480 7180

October 260 3210 4310 4860 5310 5960 5540 7170

November 550 3350 4200 4660 5180 5830 5430 7080

December 930 3620 4070 4520 5030 5510 5210 6930

held for NC-216.

b Phrygian dropped NC-336, from their regular product line, retaining it

as an option at a price premium.

c Phrygian introduced NC-236, identical to NC-216, except with a right

hand spin and a left-hand twist.

Monthly order for NC in poundsa

Roy, I know its the first time and Im not really blaming you. But I want to make

sure it doesnt happen a second time. I want you two to develop a forecasting system so

youll have a better idea of what to expect.

You mean some sort of automated technique for the new computer you got?

Thats right, Fred. Since sales orders are now being entered into a data file for the

billing system anyway, there must be some way of accessing that information and using

it.

After the meeting Roy and Fred sat in Roys office kicking around ideas. Look, Roy,

heres the orders for NC-216 since late 1972. Thats when we introduced NC-216, to meet

the demand caused by federally mandated seat belts (see Table 3-1). Suppose we just use

this one item for discussion purposes. Now what do you propose to do?

Well, Fred, Im not sure. In the past I always used to talk with the salesmen periodi-

cally to get a feel for what they thought was coming. Then Id use that with my intuition

C/2 Assignments 45

January 5983 6796

February 6767 7481

March 6355 6910

April 6380 6305

May 5326 6116

June 4960 5698

July 5565 6105

August 6646 7350

September 7018 7980

October 7353 8006

November 7009 7736

December 6795 7578

For January 1981 through September 1981, inflation factor was calcu-

lated using data from October 1980.

Freds forecasts for October 1979 through September 1981

to make decisions on ordering raw materials, setting up vacation and maintenance sched-

ules.

Well, your intuition didnt intuit the big Non-Specific order.

Actually, Id thought of it. Thats why we have had that big order already in on 210.

Its just I though we wouldnt need it for almost another month.

Hey, Roy, maybe we could just use last years figure for a months demand, plus an

inflation factor to get a forecast for the same month this year.

That might make sense, Fred. But look here with the NC-216. Notice in January of

1977 the low figure. Thats probably because of our wildcat strike that entire month. If

wed used that figure to predict January 1978 we would have really been caught short.

Well, perhaps this forecasting system should have a manual component as well. A

place for our intuition to pick up on facts like that.

Were busy enough already, Fred. With almost a hundred products, not considering

color differences, wed be buried under the reams of data and the damned computer output.

Of course we should be able to take advantage of the fact that most of our products fall

into one of three or four demand patterns.

Hows this then, Roy? Suppose we are trying to forecast January 1981. We need

46 Part C. Additional materials for teaching forecasting

the forecast about three months in advance; that means the first week of October, 1980.

Suppose we take the average January demand for the five years preceding and inflate it by

a certain percentage. Since we will know October 1980s real demand by then, we will let

the inflation percentage be the percentage difference between October 1980s real demand

and the five preceding years average of October demands. Following that procedure, we

should be able to develop forecasts for even the next twelve months. (See Table 3-2.)

That sounds really good, Fred. But lets test it by going back in the data and fore-

casting our last twelve months demand. Then we can compare it with reality (see Table

3-2). In any case, just to protect ourselves, I think we should hire a consultant and see

what ideas he has.

Sounds good to me. Boy, this could cut Phrygian Threads Gordian Knot.

Assignment

You are to act as a consultant to Phrygian Thread. Prepare a careful analysis of the

demand pattern for NC-216. Then present a forecasting system for Fred and Roy to

consider for their product line. Evaluate your model on whatever criteria you feel to be

appropriate. Make explicit any assumptions you make.

C/2 Assignments 47

M.A. Verage is a retired stockbroker who on occasion still tries her hand at picking

stocks. Last summer she invested her entire lifes savings in Nike, a company that sells

not only running shoes, but also a wide range of athletic apparel. Having made a bundle

on her original investment, M.A. is now concerned that the athletic fad will fade. (She

privately fears that this is already happening.) She is also concerned that the current

speculative bubble of optimism on Wall Street will burst, and this too will send Nike stock

prices plummeting.

Her dilemma is this: since last August, the value of her stock has ranged from one and

a half to two times what she paid for it. If she sells now, she can realize a sure return on

her original investment. On the other hand, Nikes stock price has historically been quite

strongeven in recessionsand it may return to its 1983 high of $24.00. Not wanting to

make a foolish decision either way, she decides to call in an expert who can predict what

will happen to the price of Nike stock over the next 15 trading days.

For a rather modest fee, you have a suddenly become an expert on stock price behavior.

Having not had a course in finance and not having access to any information about Nikes

prospects in the athletic apparel market, you must resort to forecasting Nikes stock price

using only information on past stock prices.

Assignment

1. In a four-page (or less) document, you or your group must present your forecast of

how Nikes stock price will behave. If you present more than one set of forecasts,

you must state which one is your most preferred (and why).

2. You may assume that she has taken this subject and is familiar with the material

covered in the text.

3. You should spend at least three-quarters of your discussion describing the data

(e.g., plots, transformations, statistics . . .) and why you have elected to use your

forecasting techniques. Running every possible method on the data would be terribly

time consuming and burdensome. Please limit your efforts by first looking at the

data and then deciding what to do. You will be graded not so much on how you

do at forecasting the price and the sophistication of your techniques, as you will

on the thoughtfulness and clarity of your discussion. Remember, the idea here is

not so much to give you a once-and-for all grade, as it is for the experiment with

techniques used in class. If you get to the point where you are disgusted with moving

48 Part C. Additional materials for teaching forecasting

averages, exponential smoothing, etc, you have probably tried too many techniques

and written too much.

4. Your paper is limited to four double-spaced pages of text. You can attach as many

plots or copies of output as you want.

5. The data on Nikes stock price are below. Please use only these data. The data are

arrayed by date and price. February data are given first, followed by March.

Date Price Date Price Date Price Date Price Date Price

01 19.875 02 19.375 03 19.500 04 20.125 07 20.000

09 20.125 10 20.250 11 21.500 14 22.750 15 23.000

16 22.625 17 23.125 18 23.125 21 23.125 22 22.625

23 22.625 24 19.125 25 17.000 28 17.875 01 17.625

02 17.125 03 16.625 04 16.125 07 16.000 08 15.625

09 16.000 10 16.375 11 16.250 14 16.125 15 16.125

16 16.125 17 16.000 18 15.625 21 15.875 22 16.000

23 16.250 24 15.875

C/2 Assignments 49

Paper requirements

All papers must be typed, double-spaced and have neat corrections. Your paper can be

anywhere from seven to sixteen pages in length. I suggest you aim for seven to ten pages,

but if you require more pages to make your arguments clear, you may use it. Footnotes,

references and exhibits are not counted in the above page limits.

In writing your papers, remember that, although content is paramount, style and clear

prose are also important. Any sources that contribute significant or little known facts

must be referenced with footnotes. You are also not permitted to turn in any paper

submitted for another course (or any modified course paper). All data sources must be

clearly detailed.

Topics

You pretty much have freedom to choose your own topic as long as it is related to a

forecasting problem. This forecasting problem can be an event change study, a time series

study, or a characteristics-based study. You must use real data. This data can be gathered

from public or private sources. You must reserve 5 to 10 per cent of your data for out of

sample predictions. You may not use this data when fitting your model. Once you have

settled on perhaps several models, you are then to forecast the out of sample data, as

well as periods (or phenomena) beyond your data. You are then to write up your results,

evaluating how you did in forecasting your reserved data and how you expect to do with

your future forecasts. You must not alter your models once you have simulated them over

the reserved data. I will not penalize you for bad out of sample predictions as long as you

intelligently evaluate why they occurred and you indicate how you might have gone about

fixing your model once these new data were revealed (you can actually revise your model

if you find it a compelling exercisebut you must report your initial models and results

first).

As far as content, you should in the beginning of the paper lay out the issue or topic

you are discussing. This should include a statement of the forecasting problem or topic,

any analytical or numerical frameworks you wish to use, and a brief statement of the

forecasting methods you have chosen to explore. You then should present your analysis

and facts back to your analysis . Finally, you should give a brief summary of your models,

50 Part C. Additional materials for teaching forecasting

and their out of sample predictions, and an evaluation of the reliability of your model.

Suggestions for actions based on your forecasts should also be made at this point.

Sample topics might include:

quality, advertising, etc.).

6. Judgmental forecasting

C/2 Assignments 51

This assignment is based on the article Box-Jenkins vs multiple regression: some

adventures in forecasting the demand for blood tests by Everette S. Gardner, Interfaces,

Vol. 9, August 1979.

This paper is a report on consulting activities performed for the Clinical Coagulation

Laboratory at the North Carolina Memorial Hospital. You are to put yourself in the

position of the Laboratory Director. Consider the papers as the consultants final report

to you on their study of the problem of forecasting the Laboratorys demand for blood

tests. While you are responsible for this project, your boss has taken a keen interest

in it as well and she is eager to have the Laboratory start using an analytically based

forecasting method. She has also received a copy of the consultants report and you can

assume she has read it. She is now awaiting your report on the project and how you will

proceed. Your presentation should include a brief analysis of the consultants work and

your own recommendation for implementing the proposed forecasting procedure. What is

your presentation?

52 Part C. Additional materials for teaching forecasting

As the operations manager for Winning Wines, Sandra McDougal has become quite con-

cerned about managing her product release. The firm produces a large range of wines

primarily for local consumption. Lately Winning Wines has found a valuable new ex-

port market in South East Asia. Currently, of the 39 distinct products Winning Wines

produces, 12 are for this new market.

Ms McDougal is worried about managing the release of her products gradually so that

customer demand is satisfied without affecting the price by oversupply. Consequently she

has decided to introduce a formal forecasting procedure at Winning Wines. In order to

begin her analysis of the demand pattern for the companys products, Ms McDougal has

selected a product for which the company has a detailed historya popular sparkling

white wine.

From marketing she has obtained extensive monthly sales data for this wine. Graphs

of the data with ACF and PACF plots are attached. What forecasting techniques should

Ms McDougal be considering at this point? Defend your choice. What advice can you

give to help her in developing a forecasting system for demand for Winning Wines entire

product line?

C/2 Assignments 53

o

o

600

o

o o

500

o o

400

o

liters

o

o

o o

o

300

o o

o

o

o o o

o

200

o o

o o o o o o

o o

o o o o o o o o

o o o o o o o o

o o o o o o o

o o o o

o o o o

o

o o

o

100

1 2 3 4 5 6

0.8

0.6

0.4

0.4

PACF

ACF

0.2

0.0

0.0

-0.2

-0.4

0 10 20 30 0 10 20 30

54 Part C. Additional materials for teaching forecasting

Select one time series of real data. The series can be selected from among those available

in the Time Series Data Library (www.maths.monash.edu.au/~hyndman/tseries/) or can be

published data or data you have collected. The data series must be seasonal and comprise

at least 30 observations.

1. Make a time plot of your data and describe the main features of the series.

2. Transform your series if necessary. Explain which transformation was used and why.

If no transformation was used, explain why not.

plot and a seasonal sub-series plot for the decomposition.

4. Forecast the next two years of your series using HoltWinters additive method. Give

the parameters of the method and report the MSE, MAPE and MAD of the one-step

forecasts from your method. If you transformed the series, give the forecasts on the

original scale.

5. Find a prediction interval for the next observation using the MSE. Check the as-

sumption of normality.

6. Add your forecasts and prediction interval to the graph of the data.

7. Explain why the MSE cannot be used to obtain prediction intervals for longer-term

forecasts.

C/2 Assignments 55

Select one time series of real data. The series can be selected from among those available

in the Time Series Data Library (www.maths.monash.edu.au/~hyndman/tseries/) or can be

published data or data you have collected. The data series must comprise at least 30

observations.

You should produce forecasts of the series using an ARIMA model. Write a brief report

(about 4 pages) of your analysis including

transformations

model selection

estimation

diagnostics

Explain carefully what you have done and why you have done it. You should also compare

your results with those obtained using an exponential smoothing method. Which method

do you think gives the better forecasts?

You should write as if your report is to a client who is interested in forecasts of your

data. You may assume that your client is familiar with the material covered in the text.

You will be graded not so much on the sophistication of your techniques, as you will on

the thoughtfulness and clarity of your discussion and the communication of your results.

You will also be required to give a presentation of your analysis in class.

56 Part C. Additional materials for teaching forecasting

C/3 Exams

C/3/1 Sample exam questions for a time series and forecasting subject

Question 1 The graphs in Figure 1 concern the production of sulphuric acid in Australia

between March 1956 and March 1992.

1.1 Describe the series in a few sentences. Does transforming or differencing seem

appropriate?

1.2 Given that Australian economic policy was radically different between 1972

and 1975 due to the Whitlam Labor Government and that there was a severe

recession in 19911992, explain in a couple of sentences some of the unusual

features of the time series plot. What other information might be helpful in

modelling this series?

1.3 Your client has asked you to provide forecasts of this series for the next two

years. She has no specific idea of the expected behavior of the forecasts and

does not require forecast intervals. Consider each of the methods listed below.

Say, in a few words each, if and why you think each of the methods listed might

be appropriate or not for this situation. If you find more than one method

that might be appropriate, discuss in about two sentences the relative merits

of the appropriate methods. Assume methods a)e) will be applied to the

data as given, without any preceding actions taken.

a) Single exponential forecasting

b) Holts method

c) HoltWinters method

d) AR(1) with 1 < < 1

e) ARIMA(0,1,1) with 1 < < 1 with a constant

f) ARIMA(0, 1, 4) applied to the data differenced at lag 4.

C/3 Exams 57

o

o

o

600 o

o

o o

o

o

o

o o

o o o

o o oo

o o

o o

500

o o o o

o o o

o

o o o o

o o o o

o o o o

o o

o o o o

o

o o

o o o o o

o o o o

oo o o o o o o

o o

o o

o o o

o o o o

400

o

o o

o

o o

o o o

o o o

o o o

o o o

o o

o o o o

o

o o

o o

300

o o

o o

o o

o

o o o

o

o o o

o o o o

o o o o

o o o o

o

o o

o o

200

o

o

oo

Time

1.0

0.8

0.8

0.6

0.4

0.6

Partial ACF

ACF

0.2

0.4

0.0

0.2

-0.2

0.0

-0.4

-0.2

0 5 10 15 20 5 10 15 20

Lag Lag

58 Part C. Additional materials for teaching forecasting

300

200

100

0

Year

1.0

0.8

0.8

0.6

0.6

Partial ACF

0.4

ACF

0.4

0.2

0.2

0.0

0.0

-0.2

0 5 10 15 20 25 5 10 15 20 25

Lag Lag

C/3 Exams 59

Question 2 The graphs in Figure 2 concern the Beveridge wheat price index from 1500

1869.

2.1 Describe the series in a few sentences. Explain why taking logarithms of the

series is appropriate. Does differencing also seem appropriate? Explain why

or why not.

2.2 Suppose you wish to forecast the series for the next 10 years. Consider each of

the methods listed below. Say, in a few words each, if and why you think each

of the methods listed might be appropriate or not for this situation. If you

find more than one method that might be appropriate, discuss in about two

sentences the relative merits of the appropriate methods. Assume methods

(a)(f) will be applied to the logged data without any other actions taken.

(a) Single exponential forecasting

(b) Holts method

(c) HoltWinters method

(d) AR(1) with 1 < < 1

(e) ARIMA(p,1,q) with no constant

(f) ARIMA(p,1,q) with a constant

(g) ARMA(p,q) applied to the logged data differenced at lag 12.

Question 3 The graph in Figure 3 is of the number of housing starts in the US each

month for nine years.

3.1 Consider forecasting the time series using the various methods listed in the

previous question. Say, in a few words each, if and why you think each of the

methods might be appropriate or not for the client in this situation. If you

find more than one method that might be appropriate, discuss in about two

sentences the relative merits of the appropriate methods. Assume methods

(a)(f) will be applied to the data as given without any preceding actions

taken.

3.2 Describe what the ACF would probably look like for this series and describe

any actions you would take before trying to fit a stationary ARMA model.

3.3 Discuss in about 45 sentences (but without giving any equations) what ac-

tions you would take after you have obtained the parameter estimates from

your ARMA model but before you produce any forecasts.

60 Part C. Additional materials for teaching forecasting

200

thousands

150

100

50

1966 1967 1968 1969 1970 1971 1972 1973 1974 1975

Question 4 The graphs in Figure 4 concern the total building and construction activity

in Australia each quarter. The units represent the value of work done in millions of

dollars at 1984/1985 prices. Data are available from July 1976 to September 1994.

However, the graphs are based on a restricted set of data. The first quarter on the

graph is JulySeptember 1976; the last quarter on the graph is JulySeptember

1991. Let Yt denote the raw series shown in the time plot and let X t denote the

series after differencing at lags 1 and 4. The ACF and PACF graphs are for the X t

series.

4.1 Describe the series in a few sentences. Does transforming seem appropriate?

If so, what transformation would you try? What features of the series suggest

differencing is appropriate?

4.2 You are developing a forecasting model for the Housing Industry Association

and you wish to test the model by forecasting the data from December 1991 to

September 1994. Consider each of the methods listed below. Comment, in a

few words each, on whether the methods listed might be appropriate for these

data. If more than one method might be appropriate, discuss in about two

sentences the relative merits of the appropriate methods. Assume methods

a)e) will be applied to the data, Yt , without any preceding actions taken.

C/3 Exams 61

o

o

o

5000

o

o

4500

o o

o

o

o

o

o

4000

o o o o

o

o o o

o o

o

o o o o

o

3500

o

o o

o

o o

o o o

o o

o o o o

o o

3000

o o

o o o o

o

o o

o

o o

o

2500

1.0

0.2

0.8

0.6

0.0

Partial ACF

0.4

ACF

0.2

-0.2

0.0

-0.2

-0.4

-0.4

0 5 10 15 20 5 10 15 20

Lag Lag

Figure 4: Graphs relating to quarterly totals of building and construction activity in Australia,

First quarter: JulSep 1976; last quarter: JulSep 1991.

62 Part C. Additional materials for teaching forecasting

b) Holts method

c) HoltWinters method

d) AR(1) with 1 < < 1

e) ARIMA(0, 4, 1)

f) AR(8) applied to Xt

g) ARIMA(1,0,0) applied to Xt

4.3 Explain why it is better when evaluating forecast performance to fit the model

using the data up to September 1991 rather than using the complete data set

up to September 1994.

C/3 Exams 63

C/3/2 Two hour exam for a time series and forecasting subject

25

20

15

10

5

Figure 1: Time plot of monthly retail turnover ($ million) of recreational goods in Tasmania

between April 1982 and March 1996.

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

1995 13.7 14.7 14.8 13.0 14.0 13.4 13.6 14.9 13.5 14.7 15.7 21.9

1996 16.9 16.3 14.7

Question 1

1.1 Describe the series plotted above in a few sentences. Comment on trend,

seasonality, cycles and changes in variance and discuss the causes for these.

1.2 Explain why it is easier to analyze the logarithms of the data rather than the

raw data.

1.3 Your client has asked you to provide forecasts of this series for the next two

years. Consider each of the methods listed below. Assume the methods will

be applied to the logged data. Say, in a few words each, if and why you think

each of the methods listed might be appropriate or not for this situation. If

64 Part C. Additional materials for teaching forecasting

you find more than one method that might be appropriate, discuss in about

two sentences the relative merits of the appropriate methods.

a) Single exponential forecasting

b) Holts method

c) HoltWinters method

d) AR(1) with 1 < < 1

e) ARIMA(0,1,1) with 1 < < 1 and with the mean removed after differ-

encing at lag 1

f) ARMA(p,q) model fitted to the series after differencing at lag 12.

g) Seasonal means method.

Question 2 Figure 2 shows the results of a STL decomposition applied to the logarithm

of the data shown in Figure 1. The seasonal component is assumed to be constant

from year to year. Figure 3 shows the seasonal pattern.

2.1 Say which quantities are plotted in each graph of Figures 2 and 3.

2.2 Explain how seasonally adjusted data can be obtained using the quantities

plotted in Figure 2.

2.3 If you were using a classical decomposition, what sort of moving average

smoother would be appropriate for estimating the trend of the series? Ex-

press the smoother as a weighted moving average smoother and explain how

the weights ensure there is no seasonal contamination of the trend estimate.

2.4 Explain why there is a problem with computing a moving average smoother

near the ends of the series. Explain why a loess smoother does not have this

problem.

2.5 What sort of decomposition would have been necessary if we had used the

raw data instead of the logged data?

C/3 Exams 65

3.0

2.5

2.0

1.5

2.8

2.4

2.0

1.6

0.3

0.1

-0.1

0.10

0.0

-0.10

time

66 Part C. Additional materials for teaching forecasting

Seasonal pattern

0.3

0.2

0.1

0.0

-0.1

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Figure 3: Seasonal pattern (or indices) based on the STL decomposition in Figure 2.

3.0

Forecasts

2.5

2.0

1.5

Figure 4: Forecasts of the logarithms of the data shown in Figure 1, computed using Holt-

Winters method with parameters a = 0.47, b = 0.41 and c = 0.0.

C/3 Exams 67

Question 3 Holt-Winters method was used to forecast the logged data. The forecasts

are shown in Figure 4. The MSE for the one-step forecasts is 0.0045. The first few

forecasts are:

2.69 2.74 2.70 2.74 2.79

3.1 Give the forecasts for AprilAugust 1996 on the original scale.

3.2 Compute a 95% prediction interval for the first forecast (on the original scale).

3.3 The smoothing parameters (, and ) were chosen to minimize the one-step

MSE. Using the Holt-Winters equations, explain what = 0 implies about

the data. Discuss how this feature of the data is also seen in the seasonal

decomposition in Figure 2.

Question 4 Let the series plotted in Figure 1 be denoted by {X t }, let Yt = log(Xt ) and

let Wt = Yt Yt12 . Then the following model was fitted:

Wt = 0.52Wt1 + 0.38Wt2 + Zt

4.1 What sort of ARIMA model is Wt (i.e. what are p, d and q)?

4.2 Is the model for Wt stationary?

4.3 Write down the model for Yt . Is the model for Yt stationary?

4.4 Compute the forecasts for Xt for April 1996 and May 1996.

4.5 Compare the forecast performance of this model with the method used in

Question 3, referring to the one-step MSE for both models.

4.6 The Ljung-Box statistic for h = 24 is 50.3. Complete the Ljung-Box test and

comment on the adequacy of the model.

End of Exam

68 Part C. Additional materials for teaching forecasting

Part I: True or False (Worth 100 points. Suggested time allotment: 1 hour)

The following statements are either always true or they are false. Answer each

statement True or False and give a one or two sentence justification that supports

your conclusion. You will receive zero for an incorrect answer, one for the right

answer, and 1 to 3 more points for your justification.

1. A forecast error cost function based upon absolute deviations (MAD) penalizes

you more for large errors than does mean squared error (MSE).

2. The larger R2 , the better the model.

3. The smaller the standard error of a multiple regression model, the more accurate

are the predictions of the future.

4. The first four randomness tests we discussed in class will (with an extremely

high probability) perform equally well in detecting any trend or seasonality in

a time series.

5. Given that we now know how to use Box-Jenkins procedures, there is absolutely

no reason to ever use a simple moving average of past values of the time series.

6. The mean of a time series (that has any variance) can never be an optimal

forecasting rule.

t = (Xt1 + Xt2 )/2) is always sta-

7. A single moving average (for example, X

tionary.

8. Single moving averages as a smoothing and forecasting tool work well on data

that have a strong linear trend.

9. Single Exponential Smoothing (SES) puts most of its averaging weight on values

that are mid-way between the start and end of the observations being averaged.

10. Holts method is always better than single exponential smoothing because it

removes more trend.

11. In SES, if the value of that minimizes MSE is equal to one, this implies X t

t = Xt1 .

is best represented as a moving average of length one (i.e. X

12. If you definitely have seasonality, trend and randomness in your time series

data, Winters method should be used before any other smoothing method.

13. In trying to pick the best exponential smoothing method, it is always best to

compare MSE. The method with the smallest MSE is always preferred.

C/3 Exams 69

15. A high F statistic for your regression indicates that you have a good forecasting

model.

16. A multiplicative decomposition method will usually be preferred to an additive

decomposition method when the variability of the data increases through time.

17. The adjusted R2 is a much better measure of goodness-of-fit and forecasting

accuracy because it penalizes you for data mining.

18. If the residuals from a regression show evidence of first-order serial correlation,

then we can obtain better forecasts of the dependent variable by incorporating

that serial dependence in our forecasting rule.

19. The F sub-block test in regression analysis is not useful if the linear regression

slope coefficients change through time.

20. Econometric system models are always better forecasting tools than linear re-

gression models because simultaneous equations models use more information

about the process generating the data.

21. Correlation always implies causality.

22. Autocorrelation coefficients are not very useful for detecting the presence of an

autoregressive process in time series data.

23. Differencing techniques can always make a time series stationary. Besides, we

can never over-difference a time series.

24. Box-Jenkins forecasting methods do not work well in the presence of quadratic

time trends in the data.

25. If there are patterns in your out-of-sample forecast errors, but your within-

sample forecast errors appear random, it was just by chance that you obtained

this nonrandomness out of sample.

3

Part II: Thought Questions (Worth 70 points. Suggested time allotment: 4 hour.)

Please answer only one of the following questions. Long answers are not required,

so please pay attention to your time. The situations described are hypothetical and

the names have been changed to protect the offenders.

1. Recently you received a forecasting report that used a multiple regression model

to construct the forecasts. The author of the report chose his model because

of the several hundred tried, it had by far the highest R-squared, adjusted

R-squared and F statistic. It also had the lowest sum of squared residuals and

a Durbin-Watson of 1.40 for n = 100 and k = 5. The author also claims that

on a reserved sample of the data the model has very little bias and the MSE

is only slightly worse than within sample. Given only this information,

70 Part C. Additional materials for teaching forecasting

(b) Suggest a number of diagnostics or additional steps you would like to see

the author take.

2. Recently the Bank of the GSB issued its forecasts of macroeconomic aggregate

variables (e.g. GNP, the money supply, interest rates, etc.) for 1998. In their

report they say that they have a ten equation model of the U.S. economy.

Discuss:

(a) What types of data (or models) they would have to have in order to predict

the 1998 values.

(b) What sorts of questions you as a consumer of these forecasts would ask in

order to be more confident that their forecast of a 20 percent inflation rate

was reasonable.

Part III: Forecasting analysis (Worth 130 points. Suggested time allotment: 1 14

hours).

(a) Look at the following retail sales data (see the attached pages) and write a

short description of the data.

(b) A colleague has proposed that

Box-Jenkins, time trend and seasonal regression, and decomposition

appear to work best as forecasting tools.

Looking at the attached output that contains data and diagnostics on each of

these best methods, evaluate the above statement. Comment on the strengths

and weakness of the three approaches.

C/3 Exams 71

o o

1700

o

o

o

o

o

1600

o

o

1500

o o

o o

o

o o

o o o

o o o

1400

o

o o o o

o

o

o o o

o o

1300

o o

o o

1 2 3 4 5 6 7 8 9 10

0.6

0.6

PACF

ACF

0.2

0.2

-0.2

-0.2

2 4 6 8 10 12 14 2 4 6 8 10 12 14

72 Part C. Additional materials for teaching forecasting

1700

Data

1500

1300

1500

Trend-cycle

1400

150

Seasonal

50

-50 0

20

10

Remainder

0

-20

2 4 6 8 10

time

Seasonal indices

-51.80256 -51.52647 189.64777 -86.31869

C/3 Exams 73

Regression Coefficients:

(Intercept) 1271.2339 9.1093 139.5535 0.0000

t 5.0954 0.2961 17.2103 0.0000

Q1 36.8263 9.5463 3.8577 0.0005

Q2 35.6839 9.5417 3.7398 0.0007

Q3 275.2755 9.5463 28.8360 0.0000

Multiple R-Squared: 0.9771

F statistic: 363.2 on 4 and 34 degrees of freedom, the p-value is 0

Period(s) of Differencing = 1,4.

Number of observations = 34

NOTE: The first 5 observations were eliminated by differencing.

Approx.

Parameter Estimate Std Error T Ratio Lag

MA1,1 0.74282 0.20763 3.58 4

Std Error Estimate = 193.026599

AIC = 458.540303

SBC = 460.066663

Number of Residuals= 34

To Chi Autocorrelations

Lag Square DF Prob

6 2.05 5 0.842 -0.007 -0.005 -0.164 0.106 -0.083 -0.074

12 5.92 11 0.878 -0.100 0.055 0.191 -0.161 -0.004 -0.054

18 13.49 17 0.703 -0.118 -0.067 0.112 0.233 0.165 0.015

24 20.71 23 0.599 -0.118 -0.031 0.091 -0.225 0.019 -0.014

No mean term in this model.

Period(s) of Differencing = 1,4.

Factor 1: 1 - 0.74282 B**(4)

74 Part C. Additional materials for teaching forecasting

o

40

o

o

o o o

o o

20

o

o

o o

o

o o

o o

o o

0 o

o o

o o

o o

o o o

o

o o

-20 o

o

o

o

o

o

-40 o

1 2 3 4 5 6 7 8 9 10

0.6 0.6

0.4 0.4

0.2

PACF

0.2

ACF

0.0

0.0

-0.2

-0.2

-0.4

2 4 6 8 10 12 14 2 4 6 8 10 12 14

C/3 Exams 75

o

40

o

o

o o

o

o

20 o

o

o

o o

o o

o

o

o

0 o

o

o o

o

o

o

o o

o o

o

o o

-20 o

-40

1 2 3 4 5 6 7 8 9 10

0.2 0.2

PACF

ACF

0.0 0.0

-0.2 -0.2

2 4 6 8 10 12 14 2 4 6 8 10 12 14

D/Solutions to exercises

1.1 Look for pragmatic applications in the real world. Note that there are no fixed

answers in this problem.

(a) Dow theory: There is an element of belief that past patterns will continue

into the future. So first, look for the patterns (support and resistance levels)

and then project them ahead for the market and individual stocks. This is a

quantitative time series method.

(b) Random walk theory: This is quantitative, and involves a time series rather

than an explanatory approach. However, the forecasts are very simple because

of the lack of any meaningful information. The best prediction of tomorrows

closing price is todays closing price. In other words, if we look at first differences

of closing prices (i.e., todays closing price minus yesterdays closing price) there

will be no pattern to discover.

(c) Prices and earnings: Here instead of dealing with only one time series (i.e., the

stock price series) we look at the relation between stock price and earnings per

share to see if there is a relationshipmaybe with a lag, maybe not. There-

fore this is an explanatory approach to forecasting and would typically involve

regression analysis.

1.2 Step 1: Problem definition This would involve understanding the nature of the indi-

vidual product lines to be forecast. For example, are they high-demand prod-

ucts or specialty biscuits produced for individual clients? It is also important

to learn who requires the forecasts and how they will be used. Are the forecasts

to be used in scheduling production, or in inventory management, or for bud-

getary planning? Will the forecasts be studied by senior management, or by

the production manager, or someone else? Have there been stock shortages so

that demand has gone unsatisfied in the recent past? If so, would it be better

to try to forecast demand rather than sales so that we can try to prevent this

76

Chapter 1: The forecasting perspective 77

happening again in the future? The forecaster will also need to learn whether

the company requires one-off forecasts or whether the company is planning on

introducing a new forecasting system. If the latter, are they intending it to

be managed by their own employees and, if so, what software facilities do they

have available and what forecasting expertise do they have in-house?

Step 2: Gathering information It will be necessary to collect historical data on each

of the product lines we wish to forecast. The company may be interested in

forecasting each of the product lines for individual selling points. If so, it is

important to check that there are sufficient data to allow reasonable forecasts

to be obtained. For each variable the company wishes to forecast, at least a

few years of data will be needed.

There may be other variables which impact the biscuit sales, such as economic

fluctuations, advertising campaigns, introduction of new product lines by a

competitor, advertising campaigns of competitors, production difficulties. This

information is best obtained by key personnel within the company. It will be

necessary to conduct a range of discussions with relevant people to try to build

an understanding of the market forces.

If there are any relevant explanatory variables, these will need to be collected.

Step 3: Preliminary (exploratory) analysis Each series of interest should be graphed

and its features studied. Try to identify consistent patterns such as trend

and seasonality. Check for outliers. Can they be explained? Do any of the

explanatory variables appear to be strongly related to biscuit sales?

Step 4: Choosing and fitting models A range of models will be fitted. These models

will be chosen on the basis of the analysis in Step 3.

Step 5: Using and evaluating a forecasting model Forecasts of each product line will

be made using the best forecasting model identified in Step 4. These forecasts

will be compared with expert in-house opinion and monitored over the period

for which forecasts have been made.

There will be work to be done in explaining how the forecasting models work

to company personnel. There may even be substantial resistance to the in-

troduction of a mathematical approach to forecasting. Some people may feel

threatened. A period of education will probably be necessary.

A review of the forecasting models should be planned.

78 Part D. Solutions to exercises

2.1 (a) One simple answer: choose the mean temperature in June 1994 as the forecast

for June 1995. That is, 17.2 C.

(b) The time plot below shows clear seasonality with average temperature higher

in summer.

20

18

16

Celsius

14

12

10

8

6

Jan Feb May Jul Sep Nov Jan Mar May

Month

Exercise 2.1(b): Time plot of average monthly temperature in Paris (January 1994May

1995).

(b) Seasonal pattern of period 24 (low when asleep); occasional peaks due to stren-

uous activity.

(c) Seasonal pattern of period 7 with peaks at weekends; possibly also peaks during

holiday periods such as Easter or Christmas.

(d) Strong seasonality with a weekly pattern (low on weekends) and a yearly pat-

tern. Peaks in either summer (air-conditioning) or winter (heating) or both

depending on climate. Probably increasing trend with variation increasing with

trend.

2.3 (a) Smooth series with several large jumps or direction changes; very large range

of values; logs help stabilize variance.

(b) Downward trend (or early level shift); cycles of about 15 days; outlier at day

8; no transformation necessary.

(c) Cycles of about 910 years; large range and little variation at low points indi-

cating transformation will help; logs help stabilize variance.

Chapter 2: Basic forecasting tools 79

(d) No clear trend; seasonality of period 12; high in July; no transformation neces-

sary.

(e) Initial trend; level shift end of 1982; seasonal period 4 (high in Q2 and Q3, low

in Q1); no transformation necessary.

2.4 1-B, 2-A, 3-D, 4-C. The easiest approach to this question is to first identify D.

Because it has a peak at lag 12, the time series must have a pattern of period 12.

Therefore it is likely to be monthly. The slow decay in plot D shows the series has

trend. The only series with both trend and seasonality of period 12 is Series 3. Next

consider plot C which has a peak at lag 10. Obviously this cannot reflect a seasonal

pattern since the only series remaining which is seasonal is series 2 and that has

period 12. Series 4 is strongly cyclic with period approximately 10 and series 1 has

no seasonal or strong cyclic patterns. Therefore C must correspond to series 4. Plot

A shows a peak at lag 12 indicating seasonality of period 12. Therefore, it must

correspond with series 2. That leaves plot B aligned with series 1.

2.5 (a)

X Y

Mean 52.99 43.70

Median 52.60 44.42

MAD 3.11 2.47

MSE 15.94 8.02

St.dev. 4.14 2.94

(b) Mean and median give a measure of center; MAD, MSE and St.dev. are mea-

sures of spread.

(c) r = 0.660. See plot on next page.

(d) It is inappropriate to compute autocorrelations since there is no time component

to these data. The data are from 14 different runners. (Autocorrelation would

be appropriate if they were data from the same runner at 14 different times.)

(b) and (c)

Notation: Error 1 = (actual demand) (method 1 forecast)

Error 2 = (actual demand) (method 2 forecast)

80 Part D. Solutions to exercises

48

Y: maximal aerobic capacity

46

44

42

40

48 50 52 54 56 58 60

X: running times

240

220

200

Demand

180

Actual

160

Forecast Method 1

Forecast Method 2

140

5 10 15 20

Month

Chapter 2: Basic forecasting tools 81

1 139 157 18 170 31

2 137 145 8 162 25

3 174 140 34 157 17

4 142 162 20 173 31

5 141 149 8 164 23

6 162 144 18 158 4

7 180 156 24 166 14

8 164 172 8 179 15

9 171 167 4 177 6

10 206 169 37 180 26

11 193 193 0 199 6

12 207 193 14 202 5

13 218 202 16 211 7

14 229 213 16 221 8

15 225 223 2 232 7

16 204 224 20 235 31

17 227 211 16 225 2

18 223 221 2 232 9

19 242 222 20 233 9

20 239 235 4 243 4

(periods 120) MAE 14.45 14.29

MSE 307.25 294.00

MPE 2.55 3.61

MAPE 7.87 8.24

Theils U 0.94 0.85

is better than Method 2. Note that this is different from the conclusion drawn

in Section 4/2/3 where these two methods are compared. The difference is that

we have used a different time period over which to compare the results. Holts

method (Method 2) performs quite poorly at the start of the series. In Chapter

4, this period is excluded from the analysis of errors.

2.7 (a) Changes: 0.25, 0.26, 0.13, . . . , 0.09, 0.77. There are 78 observations in

the DOWJONES.DAT file. Therefore there are 77 changes.

(b) Average change: 0.1336. So the next 20 changes are each forecast to be 0.1336.

(c) The last value of the series is 121.23. So the next 20 are forecast to be:

79 = 121.23 + 0.1336 = 121.36

X

80 = 121.36 + 0.1336 = 121.50

X

81 = 121.50 + 0.1336 = 121.63

X etc.

82 Part D. Solutions to exercises

In general, X

(d) See the plot below.

120

Dow Jones index

115

110

0 20 40 60 80 100

day

1 Pn n+h =

(e) The average change is c = n1 t=2 (Xt Xt1 ) and the forecasts are X

Xn + hc. Therefore,

n

n+h = Xn + h 1 X

X (Xt Xt1 )

n 1 t=2

h

= Xn + (Xn X1 ).

n1

This is a straight line with slope equal to (X n X1 )/(n 1). When h = 0,

n+h = Xn and when h = (n 1), X

X n+h = X1 . Therefore, the line is drawn

between the first and last observations.

2.8 (a) See the plot on the next page. The variation when the production is low is

much less than the variation in the series when the production is high. This

indicates a transformation is required.

(b) See the plot on the next page.

(c) See the table on page 84.

Chapter 2: Basic forecasting tools 83

10 12

Forecast

Vehicles (thousands)

8

6

4

2

0

Logarithms of vehicles

8

6

4

Exercise 2.8 (a) and (b): Time plots of Japanese automobile production and the logarithms

of Japanese automobile production.

84 Part D. Solutions to exercises

1947 11 2.40

1948 20 3.00 2.40 0.598 0.357 0.1996

1949 29 3.37 3.00 0.372 0.138 0.1103

1950 32 3.47 3.37 0.098 0.010 0.0284

1951 38 3.64 3.47 0.172 0.030 0.0472

1952 39 3.66 3.64 0.026 0.001 0.0071

1953 50 3.91 3.66 0.249 0.062 0.0635

1954 70 4.25 3.91 0.337 0.113 0.0792

1955 69 4.23 4.25 0.014 0.000 0.0034

1956 111 4.71 4.23 0.475 0.226 0.1009

1957 182 5.20 4.71 0.495 0.245 0.0950

1958 188 5.24 5.20 0.032 0.001 0.0062

1959 263 5.57 5.24 0.336 0.113 0.0602

1960 482 6.18 5.57 0.606 0.367 0.0981

1961 814 6.70 6.18 0.524 0.275 0.0782

1962 991 6.90 6.70 0.197 0.039 0.0285

1963 1284 7.16 6.90 0.259 0.067 0.0362

1964 1702 7.44 7.16 0.282 0.079 0.0379

1965 1876 7.54 7.44 0.097 0.009 0.0129

1966 2286 7.73 7.54 0.198 0.039 0.0256

1967 3146 8.05 7.73 0.319 0.102 0.0396

1968 4086 8.32 8.05 0.261 0.068 0.0314

1969 4675 8.45 8.32 0.135 0.018 0.0159

1970 5289 8.57 8.45 0.123 0.015 0.0144

1971 5811 8.67 8.57 0.094 0.009 0.0109

1972 6294 8.75 8.67 0.080 0.006 0.0091

1973 7083 8.87 8.75 0.118 0.014 0.0133

1974 6552 8.79 8.87 0.078 0.006 0.0089

1975 6942 8.85 8.79 0.058 0.003 0.0065

1976 7842 8.97 8.85 0.122 0.015 0.0136

1977 8514 9.05 8.97 0.082 0.007 0.0091

1978 9269 9.13 9.05 0.085 0.007 0.0093

1979 9636 9.17 9.13 0.039 0.002 0.0042

1980 11043 9.31 9.17 0.136 0.019 0.0146

1981 11180 9.32 9.31 0.012 0.000 0.0013

1982 10732 9.28 9.32 0.041 0.002 0.0044

1983 11112 9.32 9.28 0.035 0.001 0.0037

1984 11465 9.35 9.32 0.031 0.001 0.0033

1985 12271 9.41 9.35 0.068 0.005 0.0072

1986 12260 9.41 9.41 0.001 0.000 0.0001

1987 12249 9.41 9.41 0.001 0.000 0.0001

1988 12700 9.45 9.41 0.036 0.001 0.0038

1989 13026 9.47 9.45 0.025 0.001 0.0027

1990 9.47

Chapter 2: Basic forecasting tools 85

MAPE=3.21% (average of values in last column multiplied by 100).

(e) See graph. Forecast is e9.47 = 13026.

(f ) There are a large number of possible methods. One method, which is discussed

in Chapter 5, is to consider only data after 1970 and use a straight line fitted

through the original data (i.e. without taking logarithms).

(g) The data for 1974 is lower than would be expected. If this information could be

included in the forecasts, the MSE and MAPE would both be smaller because

the forecast error in 1974 would be smaller.

86 Part D. Solutions to exercises

3.1

Y 3-MA 5-MA 7-MA 3 3-MA 5 5-MA

42 55.50 70.33 81.50 62.92 81.14

69 70.33 81.50 91.60 73.50 88.71

100 94.67 91.60 99.83 93.56 96.65

115 115.67 111.40 107.57 113.22 111.10

132 129.33 128.40 126.00 129.11 125.92

141 142.33 142.60 141.86 142.33 141.60

154 155.33 155.60 156.71 155.33 156.80

171 168.33 170.00 172.86 169.56 172.32

180 185.00 187.40 189.29 185.78 189.80

204 204.00 206.00 210.71 205.11 210.96

228 226.33 230.00 236.86 228.56 236.72

247 255.33 261.40 268.29 257.78 262.54

291 291.67 298.80 283.00 295.56 289.27

337 339.67 316.50 298.80 331.78 304.09

391 364.00 339.67 316.50 351.83 318.32

400

3-MA

3x3 MA

5-MA

5x5 MA

7-MA

300

200

100

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

Chapter 3: Time series decomposition 87

The graph on the previous page shows the five smoothers. Because moving average

smoothers are flat at the ends, the best smoother in this case is the one with the

smallest number of terms, namely the 3-MA.

3.2 1

Tt = 1

3 5 (Yt3 + Yt2 + Yt1 + Yt + Yt+1 )

+ 15 (Yt2 + Yt1 + Yt + Yt+1 + Yt+2 )

+ 15 (Yt1 + Yt + Yt+1 + Yt+2 + Yt+3 )

1 2 1 1 1 2 1

= 15 Yt3 + 15 Yt2 + 5 Yt1 + 5 Yt + 5 Yt+1 + 15 Yt+2 + 15 Yt+3 .

3.3 (a) The 4 MA is designed to eliminate seasonal variation because each quarter

receives equal weight. The 2 MA is designed to center the estimated trend at

the data points. The combination 2 4 MA also gives equal weight to each

quarter.

(b) Tt = 1 Yt2 + 1 Yt1 + 1 Yt + 1 Yt+1 + 1 Yt+2 .

8 4 4 4 8

3.4 (a) Use 24 MA to get trend. If the end-points are ignored, we obtain the following

results.

Data: Trend:

Y1 Y2 Y3 Y4 Y1 Y2 Y3 Y4

Q1 99 120 139 160 Q1 110.250 129.250 150.125

Q2 88 108 127 148 Q2 114.875 134.500 154.750

Q3 93 111 131 150 Q3 100.375 119.635 138.875

Q4 111 130 152 170 Q4 105.500 124.375 145.125

Data trend:

Y1 Y2 Y3 Y4 Ave

Q1 9.750 9.750 9.875 9.792

Q2 6.875 7.500 6.500 7.042

Q3 7.375 8.625 8.875 8.292

Q4 5.500 5.625 6.875 6.000

(b) Hence, the seasonal indices are:

S1 = 9.8, S2 = 7.0, S3 = 8.3 and S4 = 6.0.

The seasonal component consists of replications of these indices.

(c) End points ignored. Other approaches are possible.

3.5 (a) See the top plot on the next page. There is clear trend which appears close to

linear, and strong seasonality with a peak in AugustOctober and a trough in

JanuaryMarch.

(b) Calculations are given at the bottom of the next page. The decomposition plot

is shown at the top of the next page.

88 Part D. Solutions to exercises

1600

1200

Plastic sales

data

800

trend-cycle

1200

1000

110

seasonal

70 80 90

104

remainder

100

96

2 3 4 5

Year J F M A M J J A S O N D

Data

1 742 697 776 898 1030 1107 1165 1216 1208 1131 971 783

2 741 700 774 932 1099 1223 1290 1349 1341 1296 1066 901

3 896 793 885 1055 1204 1326 1303 1436 1473 1453 1170 1023

4 951 861 938 1109 1274 1422 1486 1555 1604 1600 1403 1209

5 1030 1032 1126 1285 1468 1637 1611 1608 1528 1420 1119 1013

212 MA Trend

1 977.0 977.0 977.1 978.4 982.7 990.4

2 1000.5 1011.2 1022.3 1034.7 1045.5 1054.4 1065.8 1076.1 1084.6 1094.4 1103.9 1112.5

3 1117.4 1121.5 1130.7 1142.7 1153.6 1163.0 1170.4 1175.5 1180.5 1185.0 1190.2 1197.1

4 1208.7 1221.3 1231.7 1243.3 1259.1 1276.6 1287.6 1298.0 1313.0 1328.2 1343.6 1360.6

5 1374.8 1382.2 1381.2 1370.6 1351.2 1331.2

Ratios

1 119.2 124.5 123.6 115.6 98.8 79.1

2 74.1 69.2 75.7 90.1 105.1 116.0 121.0 125.4 123.6 118.4 96.6 81.0

3 80.2 70.7 78.3 92.3 104.4 114.0 111.3 122.2 124.8 122.6 98.3 85.5

4 78.7 70.5 76.2 89.2 101.2 111.4 115.4 119.8 122.2 120.5 104.4 88.9

5 74.9 74.7 81.5 93.8 108.6 123.0

Seasonal indices

Ave 77.0 71.3 77.9 91.3 104.8 116.1 116.8 122.9 123.6 119.3 99.5 83.6

Exercise 3.5(a) and (b): Multiplicative classical decomposition of plastic sales data.

Chapter 3: Time series decomposition 89

(c) The trend does appear almost linear except for a slight drop at the end. The

seasonal pattern is as expected. Note that it does not make much difference

whether these data are analyzed using a multiplicative decomposition or an

additive decomposition.

3.6 Period Trend Seasonal Forecast

t Tt St Yt = Tt St /100

61 1433.96 76.96 1103.6

62 1442.81 71.27 1028.3

63 1451.66 77.91 1131.0

64 1460.51 91.34 1334.0

65 1469.36 104.83 1540.3

66 1478.21 116.09 1716.1

67 1487.06 116.76 1736.3

68 1495.91 122.94 1839.1

69 1504.76 123.55 1859.1

70 1513.61 119.28 1805.4

71 1522.46 99.53 1515.3

72 1531.31 83.59 1280.0

3.7 (a) See the top of the figure on the previous page.

(b) The calculations are given below.

Year Q1 Q2 Q3 Q4

Data

1 362 385 432 341

2 382 409 498 387

3 473 513 582 474

4 544 582 681 557

5 628 707 773 592

6 627 725 854 661

42 MA

1 382.5 388.0

2 399.3 413.3 430.4 454.8

3 478.3 499.6 519.4 536.9

4 557.9 580.6 601.5 627.6

5 654.8 670.6 674.9 677.0

6 689.4 708.1

Ratios

1 112.9 87.9

2 95.7 99.0 115.7 85.1

3 98.9 102.7 112.1 88.3

4 97.5 100.2 113.2 88.7

5 95.9 105.4 114.5 87.4

6 91.0 102.4

Seasonal indices

Ave 95.8 101.9 113.7 87.5

90 Part D. Solutions to exercises

600

data

400

700

trend-cycle

600

500

400

seasonal

105

90 95

remainder

100

98

96

2 3 4 5 6

increasing with the level of the series. The most interesting feature of the

decomposition is that the trend has levelled off in the last year or so. Any

forecast method should take this change in the trend into account.

3.8 (a) The top plot shows the original data followed by trend-cycle, seasonal and

irregular components. The bottom plot shows the seasonal sub-series.

(b) The trend-cycle is almost linear and the small seasonal component is very small

compared to the trend-cycle. The seasonal pattern is difficult to see in time

plot of original data. Values are high in March, September and December and

low in January and August. For the last six years, the December peak and

March peak have been almost constant. Before that, the December peak was

growing and the March peak was dropping. There are several possible outliers

in 1991.

Chapter 3: Time series decomposition 91

(c) The recession is seen by several negative outliers in the irregular component.

This is also apparent in the data time plot. Note: the recession could be made

part of the trend-cycle component by reducing the span of the loess smoother.

3.9 (a) and (b) Calculations are given below. Note that the seasonal indices are

computed by averaging the de-trended values within each half-year.

Data 22 MA Detrended Seasonal Seasonal

Trend Data Component Adjusted Data

1.09 0.017 1.073

1.07 1.0825 -0.0125 -0.014 1.084

1.10 1.0825 0.0175 0.017 1.083

1.06 1.0750 -0.0150 -0.014 1.074

1.08 1.0625 0.0175 0.017 1.063

1.03 1.0450 -0.0150 -0.014 1.044

1.04 1.0300 0.0100 0.017 1.023

1.01 1.0225 -0.0125 -0.014 1.024

1.03 1.0075 0.0225 0.017 1.013

0.96 -0.014 0.974

(c) With more data, we could take moving averages of the detrended values for

each half-year rather than a simple average. This would result in a seasonal

component which changed over time.

92 Part D. Solutions to exercises

4.1

Period Data MA(3) SES( = 0.7)

t Yt

Yt Et Yt Et

1974 1 1 5.4

2 2 5.3 5.40 -0.10

3 3 5.3 5.33 -0.03

4 4 5.6 5.33 0.27 5.31 0.29

1975 1 5 6.9 5.40 1.50 5.51 1.39

2 6 7.2 5.93 1.27 6.48 0.72

3 7 7.2 6.57 0.63 6.99 0.21

4 8 7.10 7.14

Accuracy statistics from period 4 through 7

ME 0.92 0.65

MAE 0.92 0.65

MAPE 13.22 9.56

MSE 1.08 0.64

Theils U 1.40 1.14

Theils U statistic suggests that the nave (or last value) method is better than

either of these. If SES is used with an optimal value of chosen, then = 1 is

selected. This is equivalent to the nave method. Note different packages may give

slightly different results for SES depending on how they initialize the method. Some

packages will also allow > 1.

Method MA(3) MA(5) MA(7) MA(9) MA(11)

Forecast 24.0 48.6 55.6 51.7 53.1

MSE 1484.3 1031.2 757.5 860.8 1313.8

(b) Forecasts for May 1992

Method = 0.1 = 0.3 = 0.5 = 0.7 = 0.9

Forecast 45.5 41.9 33.1 29.1 28.7

MSE 1421.35 1211.80 1193.98 1225.40 1298.49

(c) Of these forecasting methods, the best MA(k) method has k = 7 and the best

SES method has = 0.5. However, it should be noted that the MSE values for

the MA methods are taken over different periods. For example, the MSE for the

MA(7) method is computed only over 9 observations because it is not possible

to compute an MA(7) forecast for the first seven observations. So the MSE

Chapter 4: Exponential smoothing methods 93

values are not strictly comparable for the MA forecasts. It would be better to

use a holdout sample but there are too few data.

4.3 Optimizing for SES over the period 3 through 10:

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

MAPE 65.60 53.46 44.43 37.60 32.32 28.16 24.82 22.08 19.80 17.86

MSE 79.34 47.24 29.95 20.10 14.17 10.41 7.91 6.17 4.92 4.00

With Holts method, any combination of and will give MAPE=0. This is so

because the differences between successive values of (4.13) are always going to be

zero with this errorless series. Using = 1 for SES and = 0.5 and = 0.5 for

Holts method gives the following results.

Yt

Yt Et

Yt Et

2

4 2 2 4 0

6 4 2 6 0

8 6 2 8 0

10 8 2 10 0

12 10 2 12 0

14 12 2 14 0

16 14 2 16 0

18 16 2 18 0

20 18 2 20 0

(a) Clearly Holts method is better as it allows for the trend in the data.

(b) For SES, = 1. Because of the trend, the forecasts will always lag behind the

actual values so that the forecast errors will always be at least 2. Choosing

= 1 makes the forecast errors as small as possible for SES.

(c) See above.

4.4 (a) (b) and (c) See the table on the following page.

(d) Theres not much to choose between these methods. They are both bad! Look

at Theils U values for instance. The last value method over the same period

(1328) gives MSE=6.0, MAPE=2.05 and Theils U=1.0.

94 Part D. Solutions to exercises

t Yt MA(12) Et MA(6) Et

1 108

2 108

3 110

4 106

5 108

6 108

7 105 108.00 -3.00

8 100 107.50 -7.50

9 97 106.17 -9.17

10 95 104.00 -9.00

11 95 102.17 -7.17

12 92 100.00 -8.00

13 95 102.67 -7.67 97.33 -2.33

14 95 101.58 -6.58 95.67 -0.67

15 98 100.50 -2.50 94.83 3.17

16 97 99.50 -2.50 95.00 2.00

17 101 98.75 2.25 95.33 5.67

18 104 98.17 5.83 96.33 7.67

19 101 97.83 3.17 98.33 2.67

20 99 97.50 1.50 99.33 -0.33

21 95 97.42 -2.42 100.00 -5.00

22 95 97.25 -2.25 99.50 -4.50

23 96 97.25 -1.25 99.17 -3.17

24 96 97.33 -1.33 98.33 -2.33

25 97 97.67 -0.67 97.00 0.00

26 98 97.83 0.17 96.33 1.67

27 94 98.08 -4.08 96.17 -2.17

28 92 97.75 -5.75 96.00 -4.00

29 97.33 95.50

Accuracy criteria: periods 1328

ME -1.51 -0.10

MAE 3.12 2.96

MSE 14.40 12.64

MAPE 3.23 3.03

Theils U 1.58 1.45

Chapter 4: Exponential smoothing methods 95

Paperbacks Hardcovers

SES Holt SES Holt

Smoothing parameters = 0.213 = 0.335 = 0.347 = 0.437

= 0.453 = 0.157

Forecast Day 31 210.15 224.24 240.38 250.73

Forecast Day 32 210.15 231.79 240.38 254.63

Forecast Day 33 210.15 239.33 240.38 258.53

Forecast Day 34 210.15 246.88 240.38 262.43

MAE 29.6 33.9 27.3 28.6

MSE 1252.2 1701.7 1060.6 1273.0

MAPE 17.1 18.4 13.5 14.3

Theils U 0.68 0.92 0.81 0.92

For both series, SES forecasting is performing better than Holts method.

(d) SES forecasts are flat and Holts forecasts show a linear trend. Both series

show an upward linear trend and we would expect the forecasts to reflect that

trend. Perhaps an out-of-sample analysis would give a better indication of the

merits of the two methods.

(e) The autocorrelation functions of the forecast errors in each case are plotted on

the next page. In each case, there is no noticeable pattern. Only a few spikes

are just outside the critical bounds which is expected.

4.6 Here is a complete run for one set of values ( = 0.1 and 1 = 0.1). Note that in this

program we have chosen to make the first three values of be equal to the starting

value. This is not crucial, but it does make a difference.

t Yt Ft Et At Mt t

1 200.0 200.00 0.00 0.00 0.00 0.100

2 135.0 200.00 -65.00 -6.50 6.50 0.100

3 195.0 193.50 1.50 -5.70 6.00 0.100

4 197.5 193.65 3.85 -4.74 5.79 0.950

5 310.0 197.31 112.69 7.00 16.48 0.820

6 175.0 289.74 -114.74 -5.18 26.30 0.425

7 155.0 241.00 -86.00 -13.26 32.27 0.197

8 130.0 224.08 -94.08 -21.34 38.45 0.411

9 220.0 185.43 34.57 -15.75 38.06 0.555

10 277.5 204.62 72.88 -6.89 41.55 0.414

11 235.0 234.77 0.23 -6.17 37.41 0.166

12 234.81 0.165

96 Part D. Solutions to exercises

0.4

0.2

0.2

ACF

ACF

0.0

0.0

-0.2

-0.2

-0.4

-0.4

2 4 6 8 10 12 14 2 4 6 8 10 12 14

Lag Lag

0.4

0.4

0.2

0.2

ACF

ACF

0.0

0.0

-0.2

-0.2

-0.4

-0.4

2 4 6 8 10 12 14 2 4 6 8 10 12 14

Lag Lag

For other combinations of values for and starting values for , here is what the

final value is:

= 0.1 0.165 0.327 0.732 0.143

= 0.2 0.058 0.454 0.783 0.133

= 0.3 0.140 0.618 0.797 0.133

The time series is not very long and therefore the results are somewhat fickle. In

any event, it is clear that the value and the starting values for have a profound

effect on the final value of .

4.7 Holt-Winters method is best because the data are seasonal. The variation increases

with the level, so we use Holt-Winters multiplicative method. The optimal smooth-

ing parameters (giving smallest MSE) are a = 0.479, b = 0.00 and c = 1.00. These

give the following forecasts (read left to right):

Chapter 4: Exponential smoothing methods 97

327.5 330.5 333.6 336.7 339.7 342.8

345.9 348.9 352.0 355.0 358.1 361.2

364.2 367.3 370.4 373.4 376.5 379.6

4.8 First choose any values for the three parameters. Here we have used = = = 0.1.

Different values will choose different initial values. Our program uses the method

described in the textbook and gave the following results:

-240.5 240.5 62469.9 37.5 0.70 2.6

Now compare with the optimal values: = 0.917, = 0.234 and = 0.000. Using

the same initialization, we obtain the results in Table 4-11, namely

-9.46 24.00 824.75 3.75 0.17 0.29

98 Part D. Solutions to exercises

(b) False. The correlation is negative. So below-average values of one are associated

with above-average values of the other variable.

(c) Wages have been increasing over time due to inflation. At the same time,

population has been increasing and consequently, new houses need to be built.

So, because they are both increasing with time, they are positively correlated.

(d) There are many factors affecting unemployment and it is simplistic to draw a

causal connection with inflation on the basis of correlation. As in the previous

question, both vary with time and the correlation could be induced by their

time trends. Or they could both be related to some third variable such as

business confidence or government spending.

(e) The older people in the survey had much less opportunity for education than the

younger people. This negative correlation is caused by the increase in education

levels over time.

2 = 20, P(Xi X)(Y

i Y ) = 78. So b = 78/20 = 3.9

and a = 25 3.9(5) = 5.5. Hence, the regression line is Y = 5.5 + 3.9X.

P P p

(b) (Yi Y )2 = 304.20, (Yi Yi )2 = 25.80, e = 25.80/(5 2) = 2.933. So

304.20/(2 1)

F = = 35.4.

25.80/(5 2)

This has (21) = 1 df for the numerator and (52) = 3 df for the denominator.

From Table C in Appendix III, the P -value is slightly smaller than 0.010. (Using

a computer, it is 0.0095.) Standard errors:

q

s.e.(a) = (2.93) 15 + 25 20 = 3.53

q

1

s.e.(b) = (2.93) 20 = 0.656.

On 3 df, t = 3.18 for a 95% confidence interval. Hence 95% intervals are

: 3.900 3.18(0.656) = [1.8, 6.0]

Chapter 5: Simple regression 99

Regression Analysis

Y = 5.50 + 3.90X

Constant 5.500 3.531 1.56 0.217

X 3.9000 0.6557 5.95 0.010

Analysis of Variance

Source DF SS MS F P

Regression 1 304.20 304.20 35.37 0.010

Error 3 25.80 8.60

Total 4 330.00

(d) The line through the middle of the graph is the line of best fit. The 95%

prediction interval shown is the interval which would contain the Y value with

probability 0.95 if the X value was 17. The 80% prediction interval shown is

the interval which would contain the Y value with probability 0.80 if the X

value was 26. The dotted line at the boundary of the light shaded region gives

the ends of all the 95% prediction intervals. The dotted line at the boundary

of the dark shaded region gives the ends of all the 80% prediction intervals.

5.3 (a) See the plot on the next page and the Minitab output on page 101. The straight

line is Y = 0.46 + 0.22X.

(b) See the plot on the next page. The residuals may show a slight curvature (

shaped). However, the curvature is not strong and the fitted model appears

reasonable.

(c) R2 = 90.2%. Therefore, 90.2% of the variation in melanoma rates is explained

by the linear regression.

(d) From the Minitab output:

Prediction: 9.286. Prediction interval: (6.749, 11.823)

100 Part D. Solutions to exercises

10

8

melanoma

6

4

2

10 20 30 40

ozone

1.5

1.0

Residuals

0.5

0.0

-0.5

-1.0

10 20 30 40

ozone

Chapter 5: Simple regression 101

SUBC> predict 40.

Melanoma = 0.460 + 0.221 Ozone

Constant 0.4598 0.6258 0.73 0.481

Ozone 0.22065 0.02426 9.09 0.000

Analysis of Variance

Source DF SS MS F P

Regression 1 81.822 81.822 82.70 0.000

Error 9 8.905 0.989

Total 10 90.727

9.286 0.517 ( 8.116, 10.456) ( 6.749, 11.823)

Note that it is the prediction interval (PI) we want here. Minitab also gives

the confidence interval (CI) for the line at this point, something we have not

covered in the book.

(e) This analysis has assumed that the susceptibility to melanoma among people

living in the various locations is constant. This is unlikely to be true due to

the diversity of racial mix and climate over the locations. Apart from ozone

depletion, melanoma will be affected by skin type, climate, culture (e.g. is

sun-baking encouraged?), diet, etc.

5.4 (a) See plot on the next page and computer output on page 103.

(b) Coefficients: a = 4.184, b = 0.9431. Only b is significant, showing the relation-

ship is significant. (We could refit the model without the intercept term.)

(c) If X = 80, Y = 4.184 + 0.9431(80) = 79.63. Standard error of forecast is 1.88

(from computer output).

102 Part D. Solutions to exercises

90

80

70

Production rating

60

50

40

30

20 40 60 80

Manual dexterity

Exercise 5.4(a): Scatterplot of production rating against manual dexterity test scores.

100

80

Production rating

60

40

20

20 40 60 80 100

Manual dexterity

Chapter 5: Simple regression 103

SUBC> Predict newX.

Y = 4.18 + 0.943 X

Constant 4.184 3.476 1.20 0.244

X 0.94306 0.05961 15.82 0.000

Analysis of Variance

Source DF SS MS F P

Regression 1 6576.8 6576.8 250.29 0.000

Error 18 473.0 26.3

Total 19 7049.8

23.05 2.38 ( 18.04, 28.05) ( 11.17, 34.92)

41.91 1.46 ( 38.85, 44.97) ( 30.71, 53.10)

60.77 1.18 ( 58.28, 63.26) ( 49.71, 71.82)

79.63 1.88 ( 75.68, 83.58) ( 68.16, 91.10)

(d) For confidence and prediction intervals, use Table B with 18 df. 95% CI for

is 0.94306 2.10(0.05961) = [0.82, 1.07].

(e) See output. Again it is the prediction interval (PI) we want here, not the

confidence interval (CI). The prediction intervals are shown in the plot on the

previous page.

5.5 (a) See the plot on the following page. The straight line regression model is Y =

20.20.145X where Y = electricity consumption and X = temperature. There

is a negative relationship because heating is used for lower temperatures, but

there is no need to use heating for the higher temperatures. The temperatures

are not sufficiently high to warrant the use of air conditioning. Hence, the

electricity consumption is higher when the temperature is lower.

104 Part D. Solutions to exercises

19

Electricity consumption (Mwh)

18

17

16

10 15 20 25 30

Temperature

Exercise 5.5(a): Electricity consumption (Mwh) plotted against temperature (degrees Cel-

sius).

Possible outlier

2

1

Residuals

0

-1

10 15 20 25 30

Temperature

Exercise 5.5(c): Residual plot for the straight line regression of electricity consumption

against temperature.

Chapter 5: Simple regression 105

(b) r = 0.791

(c) See the plot on the previous page. Apart from the possible outlier, the model

appears to be adequate. There are no highly influential observations.

(d) If X = 10, Y = 20.2 0.145(10) = 18.75. If X = 35, Y = 20.2 0.145(35) =

15.12. The first of these predictions seems reasonable. The second is unlikely.

Note that X = 35 is outside the range of the data making prediction danger-

ous. For temperatures above about 20 C, it is unlikely electricity consumption

would continue to fall because no heating would be used. Instead, at high

temperatures (such as X = 35 C), electricity consumption is likely to increase

again due to the use of air-conditioning.

(b) When H = 40 and W = 150, r = 0.001.

(c) The following table shows the influence of outliers at various positions.

H W r

129 0 -0.393

128 22 0.032

122 44 0.527

112 64 0.773

99 83 0.846

83 99 0.810

65 112 0.627

44 122 0.151

22 128 -0.365

0 129 -0.624

The point about all this is that an outlier (and skewness in general) can seriously

affect the correlation coefficient. It is a good idea to look at the scatterplot

before computing any correlation.

5.7 (a) See the plot on the next page. The winning time has been decreasing with year.

There is an outlier in 1896.

(b) The fitted line is Y = 1960.0768X where X denotes the year of the Olympics.

Therefore the winning time has been decreasing an average 0.0768 seconds per

year.

(c) The residuals are plotted on the next page. The residuals show random scatter

about 0 with only one usual point (the outlier in 1896). But note that the

last five residuals are positive. This suggests that the straight line is levelling

outthe winning time is decreasing at a slower rate now than it was earlier.

106 Part D. Solutions to exercises

54

52

50

winning.time

48

46

44

year

3

2

fit$resid

1

0

-1

year

Exercise 5.7(c): Residual plot for linear regression model of winning times.

Chapter 5: Simple regression 107

This would smash the world record. But given the previous five results (with

positive residuals), it would seem more likely that the actual winning time

would be higher. A prediction interval is

5.8 (a) There is strong seasonality with peaks in November and December and a trough

in January. The surfing festival shows as a smaller peak in March from 1988.

The variation in the series is increasing with the level and there is a strong

positive trend due to sales growth.

(b) Logarithms are necessary to stabilize the variance so it does not increase with

the level of the series.

(c) See the plot on the next page and the computer output on page 109. The fitted

line is Y = 526.57 + 0.2706X where X is the year and Y is the logged annual

sales.

(d)

X = 1994 : Y = 526.57 + 0.2706(1994) = 12.98

X = 1995 : Y = 526.57 + 0.2706(1995) = 13.25

X = 1996 : Y = 526.57 + 0.2706(1996) = 13.52

X = 1995 : [12.80, 13.71]

X = 1996 : [13.03, 14.02]

(e) We transform the forecasts and intervals with the exponential function:

Total annual sales for 1995 exp(13.25) = $569, 439

Total annual sales for 1996 exp(13.52) = $746, 383

Prediction intervals:

X = 1995 : [e12.80 , e13.71 ] = [361994, 895764]

X = 1996 : [e13.03 , e14.02 ] = [455060, 1224208]

108 Part D. Solutions to exercises

80000

60000

Sales

40000

20000

0

12.5

Log Total annual sales

12.0

11.5

Chapter 5: Simple regression 109

MTB > regress Log Sales 1 Year;

SUBC> predict new years;

Log Sales = - 527 + 0.271 Year

Constant -526.57 46.44 -11.34 0.000

Year 0.27059 0.02334 11.60 0.000

Analysis of Variance

Source DF SS MS F P

Regression 1 2.0501 2.0501 134.45 0.000

Error 5 0.0762 0.0152

Total 6 2.1263

12.9818 0.1044 ( 12.7135, 13.2502) ( 12.5661, 13.3975)

13.2524 0.1257 ( 12.9293, 13.5755) ( 12.7994, 13.7054) X

13.5230 0.1476 ( 13.1435, 13.9025) ( 13.0282, 14.0178) X

X denotes a row with X values away from the center

These prediction intervals are very wide because we are only using annual totals

in making these predictions. A more accurate method would be to fit a model

to the monthly data allowing for the seasonal patterns. This is discussed in

Chapter 7.

(f ) One way would be to calculate the proportion of sales for each month compared

to the total sales for that year. Averaging these proportions will give a rough

guide as to how to split the annual totals into 12 monthly totals.

110 Part D. Solutions to exercises

14

12

Percentage mortality

10

8

6

4

0 20 40 60 80 100

Y = 4.38 + 0.0154X

(b) From the computer output:

Predictor Coef StDev T P

Constant 4.3817 0.6848 6.40 0.000

% Type A 0.015432 0.007672 2.01 0.046

So the t-test is significant (since P < 0.05). A 95% confidence interval for the

slope is

This suggests that the Type A birds have a higher mortality than the Type B

birds, the opposite to what the farmers claim.

(c) For a farmer using all Type A birds, X = 100. So Y = 4.38 + 0.0154(100) =

5.92%. For a farmer using all Type B birds, X = 0. So Y = 4.38%. Prediction

intervals for these are [2.363, 9.487] and [0.587, 8.177] respectively.

(d) R2 = 2.6. So only 2.6% of the variation in mortality is due to bird type.

Chapter 5: Simple regression 111

140

Model 1

Model 2

120

100

consumption

80

60

40

40 60 80 100

price

(e) This information suggests that heat may be a lurking variable. If Type A birds

are being used more in summer and the mortality is higher in summer, than the

increased mortality of Type A birds may be due to the summer rather than the

bird type. A proper randomized experiment would need to be done to properly

assess whether bird type is having an effect here.

5.10 (a) Cross sectional data. There is no time component.

(b) See the plot above.

(c) When the price is higher, the consumption may be lower due to the pressure of

increased cost. Therefore, we would expect b 1 < b2 < 0.

(d) Model 1: First take logarithms of Y i , then use simple linear regression to obtain

a = 5.10, b = 0.0153, e2 = 0.0735.

Model 2: Split data into two groups. Fit each group separately using simple

linear regression to obtain

a1 = 221, b1 = 2.91 and a2 = 84.8, b2 = 0.447.

Using the equation given in the question, we obtain

e2 = 2913.7/16 = 182.06.

The fitted lines are shown on the graph above.

112 Part D. Solutions to exercises

1

Model 2: R2 = rY2 Y = 0.859. The second model is better with higher R 2 value.

2

The residual plots are given on the following page. Again, the second model is

much better showing random scatter about zero. The first model show pattern

in the residuals.

(f ) The graph on page 114 shows a local linear regression through the data. The

fitted curve resembles the fitted lines for model 2. This suggests that model 2

is a reasonable model for the data. However, our approach has also meant the

two lines do not join at X = 60. A better model would force them to join. This

means the parameters must be restricted which makes the estimation much

harder.

(g) and (h) Using model 2, forecasts are obtained by

220.9 2.906X when X 60

Y =

84.8 0.447X when X > 60.

r

1 (X 63)2

s.e.(Y ) = 182.06 1 + + .

20 10672.11

The 95% PI are obtained using Y t (s.e.) where t = 2.12 (from Table B with

16 df). Hence, we obtain the following values.

X Y s.e. [ 95% PI ]

40 104.67 14.15 [74.7 , 134.7]

60 46.55 13.83 [17.2 , 75.9]

80 49.03 14.00 [19.3 , 78.7]

100 40.09 14.65 [ 9.0 , 71.1]

120 31.15 15.70 [ -2.1 , 64.4]

For example, at a price of 80c, the gas consumption will lie between 19.3 and

78.7 for 95% of towns.

Chapter 5: Simple regression 113

40

20

Residuals model 1

0

-20

40 60 80 100

Price

20

10

Residuals model 2

0

-10

-20

40 60 80 100

Price

114 Part D. Solutions to exercises

140

120

100

Consumption

80

60

40

40 60 80 100

Price

Exercise 5.10(f ): Local linear regression through the gas consumption data. The fitted line

suggests that model 2 is more appropriate.

140

120

100

Consumption

80

60

40

20

0

40 60 80 100 120

Price

Chapter 6: Multiple regression 115

observations and k = number of explanatory variables. Here, k = 16 so that

n 16 1 = 30. Hence, n = 30 + 16 + 1 = 47.

(b)

2 = 1 (1 R2 ) n1 = 1 (1 0.943) 47 1

R nk1 = 0.913.

48 16 1

(c) F = 31.04 on (17,30) df. From Table C in Appendix III, the P -value is much

smaller than 0.01. So the regression is highly significant.

(d) The coefficients should be compared with a t 30 distribution. From Table B in

Appendix III, any value greater than 2.04 in absolute value will be significant

at the 5% level. So the constant and variables 4, 8, 12, 13, 14, 15 and 17 are

significant in the presence of other explanatory variables. Note that the signif-

icance level of 5% is arbitrary. There is no reason why some other significance

level (e.g. 2%) could not be used.

(e) The next stage would be to reduce the number of variables in the model by

removing some of the least significant variables and re-fitting the model.

6.2 (a) The fitted model is C = 273.935.68P +0.034P 2 . For this model, R2 = 0.8315.

[Recall: in exercise 5.6, model 1 had R 2 = 0.721 and model 2 had R2 = 0.859.]

So the R 2 values for each model are:

R nk1 45

= 1 (1 0.859)

2 n1 46

Model 2 R nk1 = 1 (1 0.859) 43 = 0.849.

R nk1 44

These values show that model 2 is the best model, followed by model 3. The t

values for the coefficients are:

Model 1 : t = 10.22 : t = 5.47

Model 2 1 : t = 10.33 1 : t = 6.61 2 : t = 4.11 2 : t = 1.99

Model 3 0 : t = 8.83 1 : t = 5.62 2 : t = 4.57

Of these, only 2 from model 2 is not significantly different from zero. This

suggests that a better model would be to allow the second part of model 2 to

be a constant rather than a linear function.

(b) From the computer output the following 95% prediction intervals are obtained.

116 Part D. Solutions to exercises

MTB > regress C 2 P Psq;

SUBC> predict newP newPsq.

C = 274 - 5.68 P + 0.0339 Psq

Constant 273.93 31.03 8.83 0.000

P -5.676 1.009 -5.62 0.000

Psq 0.033904 0.007412 4.57 0.000

Analysis of Variance

Source DF SS MS F P

Regression 2 17327.0 8663.5 41.95 0.000

Error 17 3511.0 206.5

Total 19 20838.0

Source DF Seq SS

P 1 13005.7

Psq 1 4321.3

173.97 14.29 ( 143.82, 204.13) ( 131.21, 216.74) XX

101.14 4.77 ( 91.08, 111.21) ( 69.19, 133.10)

55.43 4.91 ( 45.07, 65.80) ( 23.38, 87.48)

36.85 4.95 ( 26.40, 47.29) ( 4.77, 68.92)

45.38 7.14 ( 30.31, 60.46) ( 11.52, 79.25)

81.04 18.49 ( 42.02, 120.06) ( 31.62, 130.46) XX

X denotes a row with X values away from the center

XX denotes a row with very extreme X values

Chapter 6: Multiple regression 117

200

150

consumption

100

50

0

20 40 60 80 100 120

price

Exercise 6.2: Quadratic regression of gas consumption against price. 95% prediction inter-

vals shown.

P C [ 95% PI ]

20 173.97 [ 131.21 , 216.74 ]

40 101.14 [ 69.19 , 133.10 ]

60 55.43 [ 23.38 , 87.48 ]

80 36.85 [ 4.77 , 68.92 ]

100 45.38 [ 11.52 , 79.25 ]

120 81.04 [ 31.62 , 130.46 ]

It is clear from the plot that it is dangerous predicting outside the observed

price range. In this case, the predictions at P = 20 and P = 120 are almost cer-

tainly wrong. Predicting outside the range of the explanatory variable is always

dangerous, but much more so when a quadratic (or higher-order polynomial) is

used.

(c) rP P 2 = 0.990. If we were to use P , P 2 and P 3 , the correlations among these

explanatory variables would be very high and we would have a serious multi-

collinearity problem on our hands. The coefficients estimates would be unstable

(i.e. have large standard errors). Multicollinearity will often be a problem with

polynomial regression.

118 Part D. Solutions to exercises

Period Actual Forecast

54 4.646 1.863

55 1.060 1.221

56 -0.758 0.114

57 4.702 2.779

58 1.878 1.959

59 6.620 5.789

Analysis of errors: periods 54 through 59.

ME MAE MSE MPE MAPE ACF1 Theils U

0.74 1.11 2.15 34.82 41.32 -0.35 0.34

Strictly speaking, we should not compute relative measures when the data cross

the zero line (i.e., when there are positive and negative values) because relative

measures will blow up if divided by zero.

(b) and (c) Optimizing the coefficients for Holts method will give better forecasts.

Another approach is to use a simple MA forecast. An MA(2) forecast actually

works better than Holts method for both series. Other approaches are also

possible.

Calculate accuracy statistics for your forecasts and compare them with the

forecasts in Table 6-14.

95% confidence intervals for the parameters are calculated using a t 6 distribu-

tion. So the multiplier is 2.45:

1.52 2.45(0.1295) = [1.20, 1.84]

0.38 2.45(0.1941) = [0.09, 0.85]

0.27 2.45(0.1841) = [0.72, 0.18]

(b) F = 123.3 on (3,6) df. P = 0.000. This means that the probability of results

like this, if the three explanatory variables were not relevant, is very small.

(c) The residual plots on page 120 show the model is satisfactory. There is no

pattern in any of the residual plots.

(d) R2 = 0.984. Therefore 98.4% of the variation in Y is explained by the regression

relationship.

Chapter 6: Multiple regression 119

SUBC> Predict 10 40 30;

SUBC> Confidence 90.

Y = 73.4 + 1.52 X1 + 0.381 X2 - 0.268 X3

Constant 73.40 14.69 5.00 0.002

X1 1.5162 0.1295 11.71 0.000

X2 0.3815 0.1941 1.97 0.097

X3 -0.2685 0.1841 -1.46 0.195

Analysis of Variance

Source DF SS MS F P

Regression 3 2001.54 667.18 123.32 0.000

Error 6 32.46 5.41

Total 9 2034.00

Source DF Seq SS

X1 1 1118.36

X2 1 871.67

X3 1 11.51

95.762 1.632 ( 92.590, 98.934) ( 90.239, 101.285)

120 Part D. Solutions to exercises

4

3

3

2

2

residuals

residuals

1

1

0

0

-1

-1

-2

-2

5 10 15 20 30 40 50 60 70

X1 X2

4

3

2

residuals

1

0

-1

-2

10 20 30 40 50 60

X3

(e) The signs of the coefficients indicate the direction of the effect of each variable.

X1 increases heat and has the greatest effect (the largest coefficient). The other

variables are not significant, so they may not have any effect. If they do, the

coefficients suggest that X2 might increase heat and X3 might decrease heat.

(f ) For X1 = 10, X2 = 40 and X3 = 30, Y = 73.40+1.52(10)+0.38(40)0.27(30) =

95.76. 90% Prediction interval: [90.24,101.29]

6.5 The data for this exercise were taken from McGee and Carleton (1970) Piecewise re-

gression, Journal of the American Statistical Association, 65, 11091124. It might

be worthwhile to get this paper to compare what conventional regression can ac-

complish when there are special features in the data. In this case, the relationship

Chapter 6: Multiple regression 121

between the Boston dollar volume and the NYSE-AME dollar volume underwent a

series of changes over the time period of interest. In this paper, the solution was as

follows:

from Nov 67 through Jul 68 Y = 20.905 + 0.0114X

from Aug 68 through Nov 68 Y = 79.043 + 0.0205X

from Dec 68 through Nov 69 Y = 11.075 + 0.0067X

Notice the slope coefficients in these four equations. They are small (because

Bostons dollar volume is small relative to the big board volumes) but they get

increasingly stronger (from6 1 to 114 to 205) in successive periods of commission

splitting. Then in Dec 68, the SEC said no more commission splitting and it

hurt the Boston dollar volume. The slope went back to 67, which is almost where it

started.

(a) The fitted equation is Y = 66.2 + 0.014X. The following output was obtained

from a computer package.

(Intercept) -66.2193 39.6809 -1.6688 0.1046

X 0.0138 0.0029 4.7856 0.0000

F statistic: 22.9 on 1 and 33 degrees of freedom

the p-value is 0.00003465

R-sq = 0.4097 Rbar-sq = 0.3918 D-W = 0.694

(b) Output from computer package:

(Intercept) -67.2116 40.2550 -1.6696 0.1047

X 0.0135 0.0030 4.5025 0.0001

time 0.2737 0.6518 0.4199 0.6773

F statistic: 11.25 on 2 and 32 degrees of freedom

the p-value is 0.0001992

R-sq = 0.4129 Rbar-sq = 0.3762 D-W = 0.6814

Here, the regression is significant, but time is not significant. In fact, comparing

these two models shows that adding time to the regression equation is actually

worse than not adding it. See the R 2 values. And for both analyses, the D-W

122 Part D. Solutions to exercises

250

200

Y

150

100

50

Exercise 6.5(c): Connected scatterplot for the Boston and American stock exchanges.

statistic shows that there is a lot of pattern left in the residuals. A piecewise

regression approach does far better with this data set.

(c) See the plot above.

6.6 (a) and (b) Here are the seasonality indices based on the regression equations

(6.10) and (6.12). They represent the intercept term in the regression for each

of the 12 first differences.

Using (6.10) Using (6.12)

Mar-Feb -2.6 -6.2

Apr-Mar -6.7 -10.6

May-Apr -3.5 -7.4

Jun-May -5.3 -9.2

Jul-Jun -3.6 -7.4

Aug-Jul -5.2 -9.2

Sep-Aug -5.9 -9.7

Oct-Sep -6.9 -10.7

Nov-Oct -4.1 -7.9

Dec-Nov -4.7 -8.5

Jan-Dec -0.8 -4.6

Feb-Jan -2.2 -6.2

These two sets of seasonal indices are not quite the same. In the first equa-

Chapter 6: Multiple regression 123

tion (6.10), all eleven dummy variables for seasonality were allowed to be in

the regression. In the second equation (6.12), the best subsets regression pro-

cedure did not allow the first seasonal dummy into the final equation. The

absolute values are not so important because, in the presence of different sets

of explanatory variables, we expect the intercept terms to be different.

(c) The seasonal indices should be the same regardless of which month is used as

a base.

(b) DW = 0.57. dL = 1.04 at 1% level. Therefore there is significant positive

autocorrelation.

124 Part D. Solutions to exercises

7.1 (a) In general, the approximate standard error of the sample autocorrelations is

1/ n. So the larger the value of n, the smaller the standard error. Therefore,

the ACF has more variation for small values of n than for large values of n. All

three series show the autocorrelations mostly falling with the 95% bands. The

few that lie just outside the bands are not of concern since we would expect

about 5% of spikes to cross the bands. There is no reason to think these series

are anything but white noise.

(b) The lines shown are 95% critical values. These are calculated as 1.96/ n. So

they are closer to zero when n is larger. The autocorrelations vary randomly,

but they mostly stay within the bounds.

7.2 The time plot shows the series as a non-stationary level. It wanders up and down

over time in a similar way to a random walk. The ACF decays very slowly which

also indicates non-stationarity in the level. Finally, the PACF has a very large value

at lag 1, indicating the data should be differenced.

AR(1) Yt = 0.6Yt1 + et .

MA(1) Yt = et + 0.6et1 .

ARMA(1,1) Yt = 0.6Yt1 + et + 0.6et1 .

AR(2) Yt = 0.8Yt1 + 0.3Yt2 + et .

MA(2) Yt = et + 0.8et1 0.3et2 .

In each case, we assume Yt = 0 and et = 0 for t 0. The generated data are shown

on the following two pages. There is a lot of similarity in the shapes of the series

because they are based on exactly the same errors.

7.4 (a) The ACF is slow to die out and the time plot shows the series wandering in a

non-stationary way. So we take first differences. The ACF of the first differences

show one significant spike at lag 1 indicating an MA(1) is appropriate. So the

model for the raw data is ARIMA(0,1,1).

(b) There is not consistent trend in the raw data and the differenced data have

mean close to zero. Therefore, there is no need to include a constant term.

(c) (1 B)Yt = (1 1 B)et .

(d) See the output on page 127. There may be slight differences with different

software packages and even different versions of the same package. The Ljung-

Box statistics are not significant and the ACF and PACF of residuals show no

significant differences from white noise.

Chapter 7: The Box-Jenkins methodology for ARIMA models 125

1 0.010 0.010 0.010 0.010 0.010

2 1.386 1.386 1.392 1.372 1.388

3 1.362 1.358 2.193 -0.565 1.631

4 2.397 1.898 3.214 2.443 1.590

5 2.758 2.268 4.196 -0.804 2.425

6 2.695 1.832 4.350 2.416 1.622

7 1.947 0.954 3.564 -1.844 0.766

8 0.968 -0.002 2.136 2.000 -0.248

9 2.481 1.780 3.062 -0.253 1.641

10 2.209 1.860 3.697 1.523 2.300

11 1.055 0.162 2.380 -1.564 -0.264

12 -0.797 -1.592 -0.164 0.278 -1.862

13 -1.628 -2.008 -2.106 -1.842 -2.213

14 -1.047 -0.760 -2.024 1.487 -0.561

15 1.062 1.648 0.434 -0.052 1.979

16 0.917 1.294 1.554 0.768 1.653

17 0.560 0.178 1.111 -0.620 -0.273

18 1.276 0.946 1.612 1.666 0.864

19 -1.334 -1.536 -0.569 -3.619 -1.351

20 -0.711 -1.170 -1.511 3.485 -1.872

21 0.484 0.964 0.057 -2.964 1.612

22 2.050 2.306 2.340 5.176 2.461

23 2.070 1.896 3.300 -4.190 1.975

24 0.112 -0.626 1.354 3.775 -0.986

25 0.987 0.242 1.054 -3.357 -0.236

26 2.262 2.222 2.855 5.488 2.745

27 0.327 -0.028 1.685 -6.428 0.030

28 -1.514 -2.328 -1.317 5.079 -3.035

29 0.272 0.154 -0.636 -4.811 0.121

30 -0.427 0.118 -0.264 4.783 0.867

126 Part D. Solutions to exercises

AR(1) MA(1)

2

2

1

1

0

0

-1

-1

-2

0 10 20 30 0 10 20 30

ARMA(1,1) AR(2)

4

4

3

2

2

0

1

-2

0

-4

-1

-6

-2

0 10 20 30 0 10 20 30

MA(2)

2

1

0

-1

-2

-3

0 10 20 30

Chapter 7: The Box-Jenkins methodology for ARIMA models 127

MTB > ARIMA 0 1 1 Strikes;

SUBC> NoConstant;

SUBC> Forecast 3.

Type Coef StDev T

MA 1 0.3174 0.1886 1.68

Number of observations: Original series 30, after differencing 29

Residuals: SS = 9256634 (backforecasts excluded)

MS = 330594 DF = 28

Lag 12 24

Chi-Square 8.1(DF=11) 34.1(DF=23)

95 Percent Limits

Period Forecast Lower Upper

31 4164.87 3037.70 5292.04

32 4164.87 2800.11 5529.63

33 4164.87 2598.14 5731.60

(e) The last observation is yt = 3885; the last residual in series is e t = 881.87

(obtained from the computer package). Now

Yt = Yt1 + et 0.3174et1 .

So Y31 = Y30 + e31 0.3174

e30

= 3885 + 0 0.3174(881.87) = 4164.9

Y32 = Y31 + 0 0.3174(0) = 4164.9

Y33 = Y32 + 0 0.3174(0) = 4164.9

7.5 (a) The monthly data show strong seasonality and the seasonal pattern is reason-

ably stable. There is no trend in the data (this is a mature product).

(b) The pattern in the ACF plot shows the dominance of the seasonality. The

autocorrelations at lags 6, 18 and 30 are negative (because we are correlating

128 Part D. Solutions to exercises

6000

5000

4000

3000

Exercise 7.4(f ): Predicted number of strikes in USA. 95% prediction intervals shown.

the high periods with the low periods) and at lags 12, 14 and 36 they are

positive (because we are correlating high periods with high periods).

(c) The pattern in the PACF plot is not particularly revealing. However, there

is little need to try to interpret this plot when the analysis clearly shows the

dominance of the seasonality. The best approach would be to difference the

series to reduce the effect of the seasonality and then see what is left over.

(d) These graphs suggest a seasonal MA(1) because of the spike at lag 12 in the

ACF and the decreasing spikes at lags 12 and 24 in the PACF. Overall, the

suggested model is ARIMA(0,1,0)(0,1,1) 12 .

(e) Using the backshift operator: (1 B)(1 B 12 )Yt = (1 B 12 )et . Rewriting

gives

Yt Yt12 Yt1 + Yt13 = et et12 .

(b) For the differenced data, the PACF has a significant spikes at lags 1, 2 and

3 and a spike at lag 17 which is marginally significant. The spike at lag 17

is probably due to chance. Therefore an AR(3) is an appropriate model for

the differenced data. Consequently, an ARIMA(3,1,0) model is suitable for the

original data.

Chapter 7: The Box-Jenkins methodology for ARIMA models 129

(c) Now

Y1941 = 1.42(1778.1) 0.62(1797) 0.10(1791) + 0.30(1627) = 1719.8

Y1942 = 1.42(1719.8) 0.62(1778.1) 0.10(1797) + 0.30(1791) = 1697.3

(b) The model was chosen because the last significant spike in the PACF was at

lag 4. Note that the spikes at lags 2 and 3 were not significant. This makes

no difference. It is the last significant spike which determines the order of the

model.

(c) The model is

So

Y1970 = 146.1 + 0.891(528.7) 0.257(545) + 0.392(552) 0.333(534) = 515.7

Y1971 = 146.1 + 0.891(515.7) 0.257(528.7) + 0.392(545) 0.333(552) = 499.5

7.8 (a) The centered 12-MA smooth is shown in the plot on the next page. The trend

is generally linear and increasing with a flat period between 1990 and 1993.

(b) The variation does not change much with the level, so transforming will not

make much difference.

(c) The data are not stationary. There is a trend and seasonality in the data.

Differencing at lag 12 gives the data shown in the plot on page 131. These

appear stationary although it is possible another difference at lag 1 is needed.

(d) From the plots on page 131 it is clear there is a seasonal MA component of order

1. In addition there is a significant spike at lag 1 in both the ACF and PACF.

Hence plausible models are ARIMA(1,0,0)(0,1,1) 12 and ARIMA(0,0,1)(0,1,1)12 .

Comparing the two models we have the following results

ARIMA(1,0,0)(0,1,1)12 AIC=900.2

ARIMA(0,0,1)(0,1,1)12 AIC=926.9

130 Part D. Solutions to exercises

US electricity generation

300

280

260

240

220

200

Year

Hence the better model is the first one. Note that different packages will give

different values for the AIC depending on how it is calculated. Therefore the

same package should be used for all calculations.

(e) The residuals from the ARIMA(1,0,0)(0,1,1) 12 are shown in the plots on page

132. Because there are significant spikes in the ACF and PACF, the model is

not adequately describing the series. These plots suggest we need to add an

MA(1) term to the model. So we fit the revised model ARIMA(1,0,1)(0,1,1) 12 .

This time, the residual plots (not shown here) look like white noise. The AIC

is 876.7. Part of the computer output for fitting the revised model is shown

below.

Approx.

Parameter Estimate Std Error T Ratio Lag

MA1,1 0.74427 0.05887 12.64 1

MA2,1 0.77650 0.09047 8.58 12

AR1,1 0.99566 0.0070613 141.00 1

So the fitted model is

Chapter 7: The Box-Jenkins methodology for ARIMA models 131

30

o

o

o o

20

o

o o o

o o oo o

o o

o

o o o o o

oo o

o o

10

o o o

o

o

o o oo oo o o

o

o o

o o oo o o o

o o o o

o o o o o o o

o o o o o o

o o o o

o o o oo o o

oo o o o o o oo o o

o o o o

o

0

o o o o o o

o o

o o o

o o o o

o o o oo

o o o

o o

o

o

o

-10

o

o

o

o

o o

-20

1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997

0.2

0.2

PACF

ACF

0.0

0.0

-0.2

-0.2

0 10 20 30 40 0 10 20 30 40

132 Part D. Solutions to exercises

3

o

o

o

o

o

2

o o

o o o o

o

o o o o

o o

o

o o o

o

1

o o o o

o o o

o o o o oo

oo o o o o o o o

o o o o o o o o

o o o o o o

o o o o o o o o

o o o o o o

o o o

o ooo

0

o

oo o oo o o o o o

o o o o o o

o oo

o o o o o o

o o o o

o

o o

o

o o o

-1

o

o

o

o o

o

o o

o

-2

o

-3

1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997

0.2

0.1

0.1

0.0

PACF

ACF

-0.1

-0.1

-0.3

-0.3

0 10 20 30 40 0 10 20 30 40

Exercise 7.8(e): Residuals from ARIMA(1,0,0)(0,1,1)12 model fitted to the electricity data.

Chapter 7: The Box-Jenkins methodology for ARIMA models 133

Approx.

Parameter Estimate Std Error T Ratio Lag

MA1,1 0.86486 0.06044 14.31 1

MA2,1 0.80875 0.09544 8.47 12

AR1,1 0.27744 0.10751 2.58 1

Std Error Estimate = 6.46914574

AIC = 864.616345

SBC = 873.195782

Number of Residuals= 129

Lag Square DF Prob

6 1.60 3 0.659 0.028 -0.036 -0.020 -0.010 -0.014 0.095

12 7.67 9 0.568 0.004 -0.082 0.073 0.128 -0.095 0.072

18 15.32 15 0.429 0.126 0.016 -0.091 0.125 -0.105 -0.020

24 18.67 21 0.607 0.065 -0.048 0.051 0.005 0.069 -0.085

Note that the first term on the left is almost the same as differencing (1 B).

This suggests that we probably should have taking a first difference as well as a

seasonal difference. We repeated the above analysis and arrived at the following

model: ARIMA(1,1,1)(0,1,1)12 which has AIC=864.6.

The computer output for the final model is shown above. The figures under

the heading Chi Square concern the Ljung-Box test. Clearly the model passes

the test (see Table E in Appendix III).

(f ) Forecasts for the next 24 months are given on the following page.

134 Part D. Solutions to exercises

Nov 96 143 240.1614 6.4691 227.4821 252.8407

Dec 96 144 262.5516 6.9981 248.8356 276.2677

Jan 97 145 270.2423 7.1820 256.1659 284.3187

Feb 97 146 244.0064 7.3027 229.6934 258.3194

Mar 97 147 249.8899 7.4074 235.3718 264.4081

Apr 97 148 232.7683 7.5069 218.0550 247.4816

May 97 149 249.3720 7.6042 234.4680 264.2759

Jun 97 150 270.7257 7.6999 255.6341 285.8173

Jul 97 151 295.5439 7.7944 280.2671 310.8207

Aug 97 152 295.6598 7.8878 280.2000 311.1196

Sep 97 153 257.1358 7.9800 241.4952 272.7764

Oct 97 154 246.4526 8.0712 230.6332 262.2719

Nov 97 155 245.0224 8.4340 228.4920 261.5528

Dec 97 156 267.1930 8.6077 250.3222 284.0638

Jan 98 157 274.8228 8.7406 257.6914 291.9541

Feb 98 158 248.5699 8.8622 231.2003 265.9395

Mar 98 159 254.4488 8.9796 236.8491 272.0484

Apr 98 160 237.3258 9.0948 219.5004 255.1513

May 98 161 253.9291 9.2083 235.8811 271.9772

Jun 98 162 275.2828 9.3205 257.0150 293.5506

Jul 98 163 300.1009 9.4313 281.6160 318.5858

Aug 98 164 300.2168 9.5407 281.5173 318.9163

Sep 98 165 261.6929 9.6490 242.7812 280.6045

Oct 98 166 251.0096 9.7560 231.8881 270.1311

7.9 (a) See the plot on the following page. Note that there is strong seasonality and

a pronounced trend-cycle. One way to study the consistency of the seasonal

pattern is to compute the seasonal sub-series and see how stable each month

is. The results are given below.

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

1955: 94.7 94.0 96.5 101.3 102.4 103.7 104.5 104.3 104.1 101.2 98.3 95.4

1956: 94.1 93.5 96.8 103.1 104.1 102.8 103.7 103.6 103.6 101.7 98.6 96.8

1957: 95.9 96.8 99.0 97.7 99.5 101.1 102.0 103.3 105.1 103.2 99.8 96.7

1958: 95.0 94.8 96.1 100.4 101.7 102.1 103.6 104.9 104.4 101.1 96.7 95.3

1959: 93.9 94.5 96.4 100.9 102.1 103.3 104.7 106.0 104.4 100.9 98.7 96.1

1960: 95.7 95.7 95.1 98.9 100.8 102.5 104.6 106.0 104.0 100.1 98.0 96.7

1961: 95.2 94.8 96.5 101.3 101.7 103.7 105.2 105.3 104.3 101.2 97.7 96.5

1962: 93.5 93.7 95.6 100.7 102.3 102.5 104.4 106.4 103.5 101.0 97.6 96.9

1963: 94.6 93.4 95.5 99.1 100.8 104.1 106.1 107.4 104.1 100.7 97.9 97.4

1964: 93.6 93.2 94.6 98.6 100.2 103.5 106.6 107.5 103.6 101.7 97.9 96.9

1965: 95.6 92.7 94.0 96.7 99.4 103.7 108.2 108.0 104.7 100.5 98.4 99.6

1966: 97.0 93.7 95.2 97.0 98.4 104.1 105.9 107.2 104.2 99.7 97.1 96.8

1967: 93.9 93.6 94.1 99.0 102.5 105.7 109.2 109.9 104.9 99.8 98.3 93.8

1968: 91.2 91.7 94.5 99.0 101.9 103.1 105.7 106.0 103.5 100.2 100.7 99.1

1969: 96.0 94.3 94.1 96.8 100.7 104.5 106.3 107.2 103.7 102.5 100.2 99.4

1970: 95.8 93.0 92.0 96.0 100.2 103.7 106.0 105.8 102.7 98.9 97.1 96.5

These detrended data are relatively consistent from year to year with only minor

Chapter 7: The Box-Jenkins methodology for ARIMA models 135

240

o

o oo

o

o o

o o

o o oo

o o

o

220

o o

o o

oo o o

o oo

o o

o oo o o

o oo o o

o o o

o

oo o o

o o o

o oo o o o

o o o

o o o

oo o o o

200

o oo o o oo

o o

o o o o o o

o o

o o o o

o oo o o o

oo o o

o o o

o o o o o oo

o o o

o o oo o o o o

o o o o o

o o o

o o o o o

o o o

o oo oo o o o o oo

o o

180

o o o o o o o o

o

o

o o o o oo

o o o o

o o o o oo

o o

o o oo o

o o

oo o

o o

o

o oo

1.0

0.8

0.5

PACF

ACF

0.4

0.0

0.0

-0.5

0 10 20 30 40 0 10 20 30 40

136 Part D. Solutions to exercises

110

105

ratios

100

95

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

variations occurring here and there. For example, December 1967 and January

and February 1968 were noticeably lower than surrounding years.

Another way to look at seasonal patterns is via autocorrelation functions. Note

that for the raw data, the ACF shows strong seasonality over several seasonal

lags. This is further evidence of the consistency of the seasonal pattern. The

plot on the previous page shows the detrended data. Again, the seasonal pat-

tern is very consistent although the amplitude of the pattern each year varies.

Unusual results in early 1968 and early 1970 are seen.

(b) For the first 96 months, we identified an ARIMA(0,1,0)(0,1,1) 12 . For the second

96 months, we identified an ARIMA(0,1,0)(1,1,0) 12 : In practice, there is little

difference between these models. This means that once the trend has been

eliminated (by differencing), the seasonal patterns are very similar.

(c) Using the above ARIMA(0,1,0)(0,1,1) 12 model, we obtained the following fore-

casts.

Chapter 7: The Box-Jenkins methodology for ARIMA models 137

110

o

o

oo

o o

o o

o o

o

o o o o o o

o o o

105

o o oo o

oo o o o o

o o o o

o o o o o o o

o ooo o o o o o o o o

oo o o

o o

o o

o o oo o o

o o o

o o o o o

o o o o o

o o o o

o o o o o o

100

o o

o o o o o

o o o o

o o o

o o o o o o

o o

o o o o

o o

o o o

o o o o

o o o o o o o

o oo o o o o o o

o o o o

o o

o oo o o

o o o

o o

95

o o o

oo o

o o o o o o oo

o o o o oo

o o

o oo o o o

o

o

o

o

o

o

0.5

0.5

PACF

ACF

0.0

0.0

-0.5

-0.5

0 10 20 30 40 0 10 20 30 40

138 Part D. Solutions to exercises

Jan 1963 167.2 167.3286 172.4228 162.2345 -0.1286

Feb 1963 165.0 166.7030 171.7902 161.6159 -1.7030

Mar 1963 168.8 167.7141 172.8012 162.6269 1.0859

Apr 1963 175.0 176.5972 181.6844 171.5100 -1.5972

May 1963 177.9 176.9498 182.0369 171.8626 0.9502

Jun 1963 183.7 179.2212 184.3084 174.1340 4.4788

Jul 1963 187.2 186.1708 191.2579 181.0836 1.0292

Aug 1963 189.3 188.7598 193.8470 183.6727 0.5402

Sep 1963 183.4 186.1678 191.2550 181.0806 -2.7678

Oct 1963 177.3 177.2392 182.3264 172.1520 0.0608

Nov 1963 172.3 170.7711 175.8583 165.6839 1.5289

Dec 1963 171.4 168.8708 173.9580 163.7837 2.5292

Jan 1964 164.9 167.5935 172.6807 162.5063 -2.6935

Feb 1964 164.4 163.9412 169.0247 158.8577 0.4588

Mar 1964 166.9 167.4086 172.4920 162.3251 -0.5086

Apr 1964 174.2 174.2641 179.3475 169.1806 -0.0641

May 1964 177.5 176.4075 181.4909 171.3240 1.0925

Jun 1964 183.6 180.0358 185.1193 174.9523 3.5642

Jul 1964 189.5 186.3499 191.4333 181.2664 3.1501

Aug 1964 191.6 191.2063 196.2898 186.1228 0.3937

Sep 1964 185.1 187.7172 192.8007 182.6338 -2.6172

Oct 1964 181.9 178.9557 184.0392 173.8722 2.9443

Nov 1964 175.4 175.7857 180.8692 170.7022 -0.3857

Dec 1964 174.2 172.6567 177.7402 167.5732 1.5433

(d) For the second half of the data we used the ARIMA(0,1,0)(1,1,0) 12 to obtain

the forecasts at the top of the following page. The actual 19711972 figures

are also shown. The source is Employment and Earnings, US 19091978,

published by the Department of Labor, 1979.

A good exercise would be to take these forecasts and check the MAPE for 1971

and 1972 separately. The MAPE for the first forecast year should be smaller

than the MAPE for the second year.

(e) If the objective is to forecast the next 12 months then the latest data is obviously

the most relevant but to get seasonal indices we have to go back several years

and to anticipate what the next move the large cycle is going to be, we really

need to look at as much data as possible. So a good strategy would be

i. study the trend-cycle by looking at the 12-month moving average;

ii. remove the trend-cycle and study the consistency of the seasonality;

iii. decide how much of the data series to retain for the ARIMA modeling;

iv. forecast the next 12 months and use some judgment as to how to modify

the ARIMA forecasts on the basis of anticipated trend-cycle movements.

Chapter 7: The Box-Jenkins methodology for ARIMA models 139

Jan 1971 194.5 196.0141 201.4418 190.5864 -1.5141

Feb 1971 187.9 191.2939 198.9698 183.6180 -3.3939

Mar 1971 187.7 189.9446 199.3456 180.5436 -2.2446

Apr 1971 198.3 197.3595 208.2149 186.5042 0.9405

May 1971 202.7 205.7424 217.8790 193.6057 -3.0424

Jun 1971 204.2 213.1446 226.4396 199.8495 -8.9446

Jul 1971 211.7 217.8789 232.2392 203.5186 -6.1789

Aug 1971 213.4 218.8543 234.2061 203.5025 -5.4543

Sep 1971 212.0 213.2939 229.5769 197.0109 -1.2939

Oct 1971 203.4 208.3371 225.5009 191.1733 -4.9371

Nov 1971 199.5 204.7595 222.7611 186.7580 -5.2595

Dec 1971 199.3 203.4670 222.2690 184.6650 -4.1670

Jan 1972 191.3 196.2469 217.0365 175.4573 -4.9469

Feb 1972 192.1 191.0587 213.6617 168.4557 1.0413

Mar 1972 193.3 189.3618 213.6432 165.0804 3.9382

Apr 1972 203.4 196.9907 222.8417 171.1396 6.4093

May 1972 205.5 205.3066 232.6373 177.9760 0.1934

Jun 1972 218.2 212.5618 241.2960 183.8276 5.6382

Jul 1972 220.3 217.4298 247.5022 187.3575 2.8702

Aug 1972 219.9 218.2314 249.5848 186.8780 1.6686

Sep 1972 211.9 213.0587 245.6428 180.4746 -1.1587

Oct 1972 204.5 207.6473 241.4174 173.8773 -3.1473

Nov 1972 198.5 204.3907 239.3064 169.4750 -5.8907

Dec 1972 200.5 203.2051 239.2300 167.1802 -2.7051

7.10 (a) There is strong seasonality as can be seen from the time plot and the seasonal

peaks in the ACF.

(b) The trend in the series is small compared to the seasonal variation. However,

there is a period of downward trend in the first four years, followed by an

upward trend for four years. At the end the trend seems to have levelled off.

(c) The one large spike in the PACF of Figure 7-34 suggests the series needs dif-

ferencing at lag 1. This is also apparent from the slow decay in the ACF and

the non-stationary mean in the time plot.

(d) You would need to difference again at lag 1 and plot the ACF and PACF of the

new series (differenced at lags 12 and 1). It is not possible to identify a model

from Figures 7-33 and 7-34.

140 Part D. Solutions to exercises

8.1 (a) The fitted model in Exercise 6.7 (using OLS) was

Yt = 78.7 + 0.534xt + Nt .

The computer output below shows the results for fitting the straight line re-

gression with AR(1) errors. Hence the new model is

In this case, the error model makes very little difference to the parameters.

Approx.

Parameter Estimate Std Error T Ratio Lag Variable Shift

MU 79.27236 0.76093 104.18 0 SALES 0

AR1,1 0.72469 0.14647 4.95 1 SALES 0

NUM1 0.50801 0.02318 21.91 0 ADVERT 0

Variance Estimate = 1.11639088

Std Error Estimate = 1.056594

AIC = 74.2915405

SBC = 77.825702

Number of Residuals= 24

To Chi Autocorrelations

Lag Square DF Prob

6 3.46 5 0.630 0.027 0.099 -0.037 0.111 -0.060 -0.274

12 9.31 11 0.593 0.055 0.126 0.229 -0.227 0.060 -0.095

18 16.39 17 0.497 -0.117 -0.238 -0.080 0.054 -0.108 0.101

(b) The ACF and PACF of the errors is plotted on the following page. An AR(1)

model for the errors is appropriate since there is a single significant spike at

lag 1 in the PACF and geometric decay in the ACF. This is confirmed by the

Ljung-Box test in the computer output above. The Q values are given under

the column Chi Square. None are significant showing the residuals from the

full model are white noise.

Chapter 8: Advanced forecasting models 141

o

2

o

o

o

1

o

o

o

o

0

o o

o

o

o

-1

o o o

o

o o

-2

o

-3

10 20

0.4

0.4

PACF

ACF

0.0

0.0

-0.4

-0.4

2 4 6 8 10 12 2 4 6 8 10 12

Exercise 8.1: Errors from regression model with AR(1) error term.

142 Part D. Solutions to exercises

Approx.

Parameter Estimate Std Error T Ratio Lag Variable Shift

MU 9.56328 0.40537 23.59 0 HURON 0

AR1,1 0.78346 0.06559 11.94 1 HURON 0

NUM1 -0.02038 0.01066 -1.91 0 YEAR 0

Std Error Estimate = 0.71568001

AIC = 216.450147

SBC = 224.205049

Number of Residuals= 98

To Chi Autocorrelations

Lag Square DF Prob

6 8.35 5 0.138 0.222 -0.100 -0.133 -0.056 -0.007 -0.042

12 15.01 11 0.182 -0.051 0.009 0.175 0.017 -0.121 -0.107

18 16.36 17 0.499 -0.053 0.014 0.019 0.058 0.006 -0.067

24 25.47 23 0.326 -0.071 -0.166 -0.043 0.051 0.160 0.092

8.2 (a) To reduce numerical error, we subtracted 1900 from the year to create an ex-

planatory variable. Hence the year ranged from -25 (1875) to 72 (1972). The

computer output above shows the fitted model to be

Yt = 9.56 0.02xt + Nt where Nt = 0.78Nt1 + et

where xt is the year 1900.

(b) The errors are shown in the plot on the following page. This demonstrates

that a better model would have an AR(2) error term since the PACF has two

significant spikes at lags 1 and 2. The spike at lag 10 is probably due to chance.

The ACF shows geometric decay which is possible with an AR(2) model. So

the full regression model is

Yt = 0 + 1 xt + N t where Nt = 1 Nt1 + 2 Nt2 + et .

Fitting this model gives the output shown on page 144. So the fitted model is

Yt = 9.53 0.02xt + Nt where Nt = Nt1 0.29Nt2 + et .

Chapter 8: Advanced forecasting models 143

o

2

o

o o o

o

o o o

oo o

o

o o o

1

o o oo o

o o o

o

o oo

o o

oo o

o

o ooo o

o o o

o o

o o o

0

o o o

o o

o o o o

o

o o o

o o

oo oo o

oo o o

o o

o

-1

o o

o o o oo

o o o

o o o

o o

o

oo oo

-2

oo o

0.8

0.8

0.6

0.6

0.4

0.4

PACF

ACF

0.2

0.2

0.0

-0.2

-0.2

5 10 15 5 10 15

Exercise 8.2(b): Errors from regression model with AR(1) error term.

144 Part D. Solutions to exercises

Approx.

Parameter Estimate Std Error T Ratio Lag Variable Shift

MU 9.53078 0.30653 31.09 0 HURON 0

AR1,1 1.00479 0.09839 10.21 1 HURON 0

AR1,2 -0.29128 0.10030 -2.90 2 HURON 0

NUM1 -0.02157 0.0082537 -2.61 0 YEAR 0

Std Error Estimate = 0.68996319

AIC = 210.396534

SBC = 220.736404

Number of Residuals= 98

To Chi Autocorrelations

Lag Square DF Prob

6 0.60 4 0.964 0.018 -0.028 -0.003 0.040 0.054 -0.007

12 5.35 10 0.867 -0.032 -0.037 0.167 -0.007 -0.098 -0.055

18 6.21 16 0.986 -0.036 0.005 -0.025 0.035 -0.006 -0.063

24 10.49 22 0.981 -0.003 -0.141 -0.007 0.006 0.116 0.008

8.3 (a) ARIMA(0,1,1)(2,1,0)12 . This model would have been chosen by first identifying

that differences at lags 12 and 1 are necessary to make the data stationary. Then

looking at the ACF and PACF of the differenced data would have shown two

significant spikes in the PACF at lags 12 and 24. There would have also been

a significant spike in the ACF at lag 1 and geometric decay in the early lags of

the PACF.

(b) Since both parameter estimates are positive (and significantly different from

zero), we can conclude that electricity consumption increases with both heating

degrees and cooling degrees. Because b 2 is larger, we know that there is a greater

increase in electricity usage for each heating degree than for each cooling degree.

(c) To use this model for forecasting, we would first need forecasts of both X 1,t

and X2,t into the future. These could be obtained by taking averages of these

variables over the equivalent months of the previous few decades. Then the

model can be used to forecast electricity demand over the next 12 months by

Chapter 8: Advanced forecasting models 145

forecasting the Nt series using the method discussed in chapter 7 and plugging

the forecasts of X1,t , X2,t and Nt into the formula for Yt .

(d) If the model was fitted using a standard regression package (thus modeling N t

as white noise), then the seasonality and autocorrelation in the data would have

been ignored. This would result in less efficient parameter estimates and invalid

estimates of their standard errors. In particular, tests for significance would be

incorrect, as would prediction intervals. Also, when producing forecasts of Y t ,

the forecasts of Nt would be all be zero. Hence, the model would not adequately

allow for the seasonality or autocorrelation in the data.

8.4 (a) b = 3, r = 1, s = 2.

(b) ARIMA(2,0,0)

(c) 0 = 0.53, 1 = 0.37, 2 = 0.51, 1 = 0.57, 2 = 0, 1 = 2 = 0, 1 = 1.53,

2 = 0.63.

(d) 27 seconds.

8.6 (a) The three series are shown on page 147. For Set 1, four X t values are needed

(since v1 , v2 , v3 and v4 are all non-zero). Therefore 27 Yt values can be produced.

Similarly 26 Yt values for Set 2 and 24 Yt values for Set 3 can be calculated.

(b) The first model is

2.0B

Yt = Xt + N t .

1 0.7B

The simplest way to generate data for this transfer function is to rewrite it as

follows

Thus Yt values can only be generated for times t = 3, 4, . . . since we need at

least two previous Xt values. However, for t = 3, we also need Y 2 . To start the

process going, we have assumed here that Y 2 = 0. Other values could also have

been used. The effect of this initialization is negligible after a few time periods.

The second model is easier to generate as we can write it

146 Part D. Solutions to exercises

(a) (b)

2.0

3.0

1.5

2.5

1.0

2.0

weight

weight

0.5

1.5

0.0

1.0

0.5

-1.0

0.0

0 2 4 6 8 10 0 2 4 6 8 10

lag lag

(c) (d)

0.8

1.0

0.6

0.8

0.4

0.6

weight

weight

0.2

0.4

0.0

0.2

-0.2

-0.4

0 2 4 6 8 10 0 2 4 6 8 10

lag lag

Exercise 8.5: Impulse response weights for the four different transfer functions.

Chapter 8: Advanced forecasting models 147

t Nt Xt Yt Yt Yt

1 -0.8003 50

2 0.8357 90 0.0

3 1.4631 50 110.9 201.5

4 0.7332 30 58.7 51.3 64.7

5 0.3260 80 52.3 58.3 25.7 116.3

6 -0.7442 80 61.3 51.3 135.0 231.3

7 0.7362 30 65.7 62.7 74.7 143.8 132.7

8 1.1931 70 59.2 66.2 88.2 49.3 81.2

9 -1.4681 60 55.5 56.5 87.5 130.2 186.5

10 -0.5285 10 49.5 56.5 79.5 113.7 75.5

11 0.4314 40 37.4 50.4 83.4 16.4 20.4

12 -1.6341 20 27.4 35.4 72.4 75.5 94.4

13 0.8198 40 29.8 29.8 54.8 38.8 56.8

14 0.4183 20 30.4 29.4 47.4 79.0 88.4

15 -0.4065 10 23.6 29.6 41.6 38.6 19.6

16 -0.0615 30 19.9 23.9 40.9 19.3 39.9

17 0.1432 60 29.1 20.1 36.1 59.7 124.1

18 -1.0747 70 46.9 27.9 27.9 118.6 178.9

19 -0.5355 40 56.5 47.5 38.5 139.2 139.5

20 -0.1454 70 56.9 56.9 59.9 79.7 107.9

21 0.2088 10 49.2 57.2 74.2 140.1 120.2

22 -0.6854 30 34.3 48.3 76.3 19.2 -0.7

23 0.1182 30 28.1 35.1 73.1 60.1 88.1

24 0.6971 40 30.7 28.7 54.7 60.7 84.7

25 0.3698 30 34.4 30.4 43.4 80.3 92.4

26 -0.0802 100 46.9 33.9 43.9 59.9 147.9

27 -0.9202 60 64.1 46.1 44.1 199.1 247.1

28 1.1483 90 76.1 66.1 61.1 121.1 149.1

29 -0.1663 60 75.8 74.8 84.8 179.8 203.8

30 -0.5461 100 76.5 75.5 97.5 119.4 167.5

148 Part D. Solutions to exercises

(b) There are a number of ways this could be done. The simplest is to define the

monthly CPI to be the same as that of the quarter. For example, January,

February and March of 1980 would each have a CPI of 45.2; April, May and

June 1980 would each have a CPI of 46.6; and so on. Other methods might

involve fitting a smooth curve through the quarterly figures and using the curve

to predict the CPI at other points along the time axis.

(c) See figure below.

100

80

60

40

Yt = a + (0 + 1 B + + 6 B 6 )Xt + Nt

where Yt denotes the average room rate, Xt denotes the CPI and Nt is an AR(1)

process. The estimated errors from this model are shown in the figure on the

previous page. They are clearly non-stationary and have some seasonality.

So we difference both Yt and Xt and refit the model with Nt specified as an

ARIMA(1,0,0)(1,0,0)12 . The parameter estimates are shown below (as given

by SAS).

Chapter 8: Advanced forecasting models 149

o o o

o o o

o o o

o o o

o o o

o o o o

oo

o

0.05

o o oo o

o o

o o o

o oo o o o

o o

o o

oo o

o o o

o o

o o o o

o o o o

o o o

o o

o o oo o

o o o

o o

o oo o o o

0.0

o o o

o o o o o

o oo o o o

oo

o o o o o o

oo o o o o

o o o o o

o o o o o o

o o o o

o o o oo oo oo

o o o

o o o o o o o o o

o o o

o o o o

o o o o

-0.05

o o o

o o

o o o

o o

o o o o

o o o o

o o

o

o o

o o

0.6

0.6

0.4

0.4

PACF

ACF

0.2

0.2

0.0

0.0

-0.4

5 10 15 20 5 10 15 20

Exercise 8.7(d): Errors from regression model with AR(1) error term.

150 Part D. Solutions to exercises

a 0.20200 0.2848 0.4791

0 0.20730 0.2602 0.4267

1 -0.41687 0.2634 0.1154

2 0.23165 0.2655 0.3842

3 0.32048 0.2716 0.2396

4 -0.72093 0.2665 0.0075

5 0.74707 0.2633 0.0051

6 -0.36272 0.2656 0.1739

Thus the intercept and first four coefficients are not significant and can be

omitted. Hence we select b = 4. We shall retain the last three coefficients for

the moment. Since they show no clear pattern, we select r = 0 and s = 3 giving

the model

Yt = (0 + 1 B + 2 B 2 )B 4 Xt + Nt .

Looking at the ACF and PACF of the error series (not shown) and trying a

number of alternative models led us to the model ARIMA(2,1,0)(2,0,0) 12 for

Nt . That is

Parameter 0 1 2 1 2 1 2

Estimate 0.52 0.61 -0.47 -0.49 -0.33 0.37 0.41

The model suggests that there is a lag of four months between changes in the

CPI and changes in the price of travel accommodation. The seasonality inherent

in the model may be due to seasonal price variation or due to the way CPI was

estimated from quarterly data.

(e) Forecasts of CPI were obtained using Holts method. These are only needed

from November 1995 because of the time lag of 4 months. Actual data beyond

June 1995 are given in the second column for comparison.

Chapter 8: Advanced forecasting models 151

Yt Yt Xt4 Xt5 Xt6

Jul 1995 90.4 94.0 115.0 115.0 115.0

Aug 1995 91.8 96.7 116.2 115.0 115.0

Sep 1995 92.0 94.8 116.2 116.2 115.0

Oct 1995 91.6 89.6 116.2 116.2 116.2

Nov 1995 93.4 95.8 116.9 116.2 116.2

Dec 1995 90.3 91.5 117.3 116.9 116.2

Jan 1996 90.4 92.0 117.7 117.3 116.9

Feb 1996 92.6 95.5 118.1 117.7 117.3

Mar 1996 94.7 100.6 118.5 118.1 117.7

Apr 1996 90.7 94.1 118.9 118.5 118.1

May 1996 91.5 97.2 119.3 118.9 118.5

Jun 1996 93.0 102.9 119.8 119.3 118.9

80

Perpetual speed score

60

40

0 20 40 60 80 100 120

Day

Exercise 8.8: Time plot of daily perceptual speed scores for a schizophrenic patient. The

drug intervention is shown at day 61.

(b) The step intervention model with an ARIMA(0,1,1) error was used:

152 Part D. Solutions to exercises

where Yt denotes the perceptual speed score and X t denotes the step dummy

variable. The estimated coefficients were

Parameter

Estimate -22.1 0.76

(c) The drug has lowered the perceptual speed score by about 22.

(d) The new model is

Yt = Xt + N t where (1 B)Nt = et1 + et

1 B

(An ARIMA(0,1,1) error was found to be the best model again.) Here the

estimated coefficients were

Parameter

Estimate -13.21 0.54 0.76

The following accuracy measures show that the delayed effect model fits the

data better.

Model Step Delayed step

MAPE 15.1 15.0

MSE 92.5 91.1

AIC 542.8 538.4

The forecasts for the two models are very similar. This is because the effect of

the step in the delayed step model is almost complete at the end of the series,

60 days after the drug intervention.

(e) The best ARIMA model we found was an ARIMA(0,1,1) with = 0.69. This

gave MAPE=15.4, MSE=100.8 and AIC=550.9.

This model gives a flat forecast function (since we did not include a constant

term). The forecast values are 33.9. Because the step effect is almost complete

in the delayed step model, it also gives a virtually flat forecast function with

forecast values of 34.1. Hence there is virtually no difference. If forecasts had

been made earlier (for example, at day 80), there would have been a difference

because the step effect would still be in progress and so the delayed step model

would have showed a continuing decline in perceptual speed. The real advantage

of the intervention model over the ARIMA model is that the intervention model

provides a way of measuring and evaluating the effect of an intervention.

(f ) If the drug varied from day to day and the reaction times depended on dose,

then a better model would be a dynamic regression model with the the quantity

of drug as an explanatory variable.

Chapter 8: Advanced forecasting models 153

8.9 (a)

Yt Yt1 Yt1 Yt2 Yt2 Yt3

= 1 + 2

Xt Xt1 Xt1 Xt2 Xt2 Xt3

Yt12 Yt13

+ + 12 + Zt.

Xt12 Xt13

(b)

Yt = Yt1 0.38(Yt1 Yt2 ) + 0.15(Xt1 Xt2 )

0.37(Yt2 Yt3 ) + 0.13(Xt2 Xt3 ) +

= 0.62Yt1 + 0.01Yt2 + 0.15Xt1 0.02Xt2 +

(c)

Multivariate model assumes feedback. That is, X t depends on past values

of Yt . But regression does not allow this.

Regression model does not assume Xt is random.

Regression model allows Yt to depend on Xt as well as past values

Xt1 , Xt2 , . . .. Multivariate AR only allows dependence on past values

of {Xt }.

For these data, it is unlikely room rates will substantially affect Y t although

it is possible. Small values in lower left of coefficient matrices suggest that

Xt is not affecting Yt . Yt should depend on Xt . So regression is probably

better.

8.10 (a) An AR(3) model can be written using the same procedure as the AR(2) model

described in Section 8/5/1. Thus we define X 1,t = Yt , X2,t = Yt1 and X3,t =

Yt2 . Then write

1 2 3 at

X t = 1 0 0 X t1 + 0

0 1 0 0

and Yt = [1 0 0]X t .

This is now in state space form with

1 2 3 1 0 0 at

F = 1 0 0 ,G = 0 1 0 , H = [1 0 0], et = 0 and zt = 0.

0 1 0 0 0 1 0

write this in state space form by letting F = 0, G = , e t = at1 , zt = at and

H = 1. Thus

Yt = X t + a t and Xt = at1 .

154 Part D. Solutions to exercises

Lt = Yt + (1 )(Lt1 + bt1 ),

bt = (Lt Lt1 ) + (1 )bt1 ,

et = Yt Lt1 bt1 . The first row can be written

= et + Lt1 + bt1

= bt1 + (Lt Lt1 bt1 )

= bt1 + et

Now let Xt,1 = Lt and Xt,2 = bt . Then the state space form of the model is

1 1

Xt = X t1 + et

0 1

Yt = [1 1] X t1 + et .

it allows missing values to be handled easily;

it is easy to generalize to allow the parameters to change over time;

the Kalman recursion equations can be used to calculate the forecasts and

likelihood.

Chapter 9: Forecasting the long term 155

9.1 There is little doubt that the trends in computer power and memory show a very

clear exponential growth while that of price is declining exponentially. It is therefore

a question of time until computers that cost only a few hundred dollars will exist

that can perform an incredible array of tasks which until now have been the sole

prerogative of humans, for example playing chess (a high-power judgmental and

creative process). It is therefore up to our imaginations to come up with future

scenarios when such computers will be used as extensively as electrical appliances

are used today. The trick is to free our thinking process so that we can come up with

scenarios that are not constrained by our perception of the present when computers

are being used mostly to make calculations.

9.2 As the cost of computers (including all of the peripherals such as printers and scan-

ners) is being reduced drastically, and at the same time we will be getting soon to

devices that will perform a great number of functions now done by separate ma-

chines, it will become more practical and economical to work at home. Furthermore,

the size of these all-purpose machines is being continuously reduced. In the next

five to ten years we will be able to have everything that is provided to us now in

an office at home with two machines: one a powerful all-inclusive computer and the

other a printer-scanner-photocopier-fax machine. Moreover these two machines will

be connected to any network we wish via modems so that we can communicate and

get information from anywhere.

9.3 As it was also mentioned in Exercise 9.1, there is no doubt that the trend in computer

and equipment prices are declining exponentially at a fast rate. This would make it

possible for everyone to be able to afford them and be able to have an office not only

at home but at any other place he or she wishes, including ones car, a hotel room,

a summer vacation residence, or a sail boat.

9.4 Statements like those referred to in Exercise 9.4 abound and demonstrate the short-

sightedness of peoples ability to predict the future. As a matter of fact as late

as the beginning of our century people did not predict all four major inventions of

the Industrial Revolution (cars, telephones, electrical appliances and television) that

have dramatically changed our lives. Moreover, they did not predict the huge impact

of computers even as late as the beginning of the 1950s. This is why we must break

from our present mode of thinking and see things in a different, new light. This is

where scenarios and analogies can be extremely useful.

156 Part D. Solutions to exercises

10.1 Phillips problems have to do with the management bias of overoptimism, that is

believing that all changes will be successful and that they can overcome peoples

resistance to change. This is not true, but we tend to believe that most organisational

changes are successful because we hear and we read about the successful ones while

there is very little mention of those that fail. Introducing changes must be considered,

therefore, in an objective manner and our ability to succeed estimated correctly.

10.2 The quote by Glassman illustrates the extent to which professional, expert invest-

ment managers underperform the average of the market. Business Week, Fortune

and other business journals regularly publish summary statistics of the performance

of mutual funds and other professionally-managed funds, benchmarking them with

the Standard & Poor or other appropriate indexes. The instructor can therefore get

some more recent comparisons than those shown in Chapter 10 and show them in

class.

10.3 Assuming that the length of cycles varies considerably we have little way to say how

long it will take until the expansion started in May 1991 will be interrupted. Un-

fortunately the length (and depth) of cycles varies a great deal making it extremely

difficult to say how long an expansion will last. It will all depend on the specific

situation involved that will require judgmental inputs, structured in such a way as

to avoid biases and other problems.

10.4 There are twenty 8s that one will encounter when counting from 0100. When given

this exercise most people say nine or ten because they are not counting the eights

coming from 81 to 87, and 89 (they usually count the 8s in 88 often one time).

Chapter 11: The use of forecasting methods in practice 157

11.1 The results of Table 1 are very similar to those of the previous M-Competitions. As

a matter of fact the resemblance is phenomenal given the fact that the series used

and the time horizon they refer to are completely different.

11.2 In our view the combined method will do extremely well. More specifically its accu-

racy will be higher than the individual methods being combined while its variance

of forecasting error will be smaller than that of the methods involved.

11.3 It seems that proponents of new forecasting methods usually exaggerate their ben-

efits. This has been the case with methods under the banner of neural networks,

machine learning and expert systems. These methods did not do well in the M3-IJF

Competition. In addition only few experts participated in the competition using

such methods, even though more than a hundred were contacted (and invited to

participate) and more than fifty expressed an interest in the M3-IJF Competition,

indicating that they would possibly participate . In the final analysis it seems that

it is not so simple to run a great number of series by such methods resulting in not

too many participants from such methods.

158 Part D. Solutions to exercises

limitations

The exercises for Chapter 12 are general and can be answered by referring to the text

of Chapter 12 which covers each one of the topics. Each instructor can therefore form

his/her way of answering these exercises.

- Noise Resilient Periodicity Mining In Time Series Data basesUploaded byseventhsensegroup
- 06rr410308 Production Planning and ControlUploaded bymurugan1987
- Forecasting SkUploaded byNirmay Mufc Shah
- Student Slides Chapter 3Uploaded byNiaz Ahmad
- Forecating Questionnaire-Customer (Process)Uploaded bymshubhadeep
- OM Assignment 2Uploaded byUdit_Dave_251
- 210094081-Speciality-Packaging-Case-study.pdfUploaded byরাকিবইসলাম
- 2inventory ManagementUploaded byrichard
- Forecasting LectureUploaded byAlice Star
- Rao, Ambar G. and Gary Lilien - A System of Promotional ModelsUploaded byDana Constantin
- UNIT II DFUploaded byRupika2316
- F-GD02 STLFUploaded byEmar Orea
- AnalyticUploaded bysandy_skvv
- Forecasting 1Uploaded byKizaGaki Nagato
- IAME 14Uploaded byDejan Dragan
- PEC Productivity (Equipment)Uploaded bytruongngv
- End Aut 2011Uploaded byMridul Pant
- Sales Objectives and QuotasUploaded bysubbugss
- Class Notes RM2Uploaded byShailja Rajinder Raina
- OSCM SummaryUploaded byCaterpillar
- Mcqs Time Series 2Uploaded byBassamSheryan
- M and W PatternsUploaded byIconoclast
- SC - 3.pptxUploaded byDweep Chokshi
- Fundamentals of Applied Econometrics.pdfUploaded byImam Awaluddin
- F9FM-MT2A-As_d08Uploaded byErclan
- Chapter 5 - Cost EstimationUploaded byalleyezonmii
- Statistics for ManagementUploaded bySammir Malhotra
- 1.Homework Week4 Session 2Uploaded byTony Lâm
- Transforming Time SeriesUploaded byZenzen Miclat
- Business AnalysisUploaded bySantosh Ravindra Nadagouda

- 2001-principlesforecastingUploaded byFrederico Sande Viana
- notes_5.pdfUploaded byPragyan Nanda
- MarkdownUploaded byPragyan Nanda
- 17GeometricSearch.pdfUploaded byAnil
- Lecture 12Uploaded byPragyan Nanda
- Practical 9Uploaded byPragyan Nanda
- Model Selection for Primal SVMUploaded byPragyan Nanda
- PHI-LEARNING-Computer-Science-IT-Engineering-Electrical-Electronics-Mechanical-Civil-Chemical-Metallurgy-and-Agricultural-Catalogue-2016.pdfUploaded byPragyan Nanda
- Chap 7Uploaded bystruggler17
- Data_Clustering_Theory_Algorithms_and_Ap.pdfUploaded byPragyan Nanda
- Time Series Forecasting Using Holt-Winters Exponential SmoothingUploaded byphu_2000_99
- ZhangEtAl1998_ForecastingwithArtificialNeuralNetworks_IntJForecastV14p35-52_10.1.1.138Uploaded byalexandru_bratu_6
- Notes on Nonseasonal ARIMA Models--Robert NauUploaded byPragyan Nanda
- Projects OverviewUploaded byPragyan Nanda
- Matlab and Its ApplicationUploaded byPragyan Nanda
- manual.pdfUploaded byPragyan Nanda
- A Study on Using Business Intelligence fUploaded byPragyan Nanda
- The 7 Basic PlotsUploaded byPragyan Nanda
- NN ExamplesUploaded byMinh Map Vu
- NN ExamplesUploaded byMinh Map Vu
- test.xlsxUploaded byPragyan Nanda
- index cheat.pdfUploaded byPragyan Nanda
- IBF04 Neural ForecastingUploaded byPragyan Nanda
- SumitabahdasUploaded byShweta Verma
- Indexing CheatsheetUploaded byPragyan Nanda
- time_series Machine LearningUploaded byakragnarock
- indexing_cheatsheet.pdfUploaded byPragyan Nanda
- KIITEE - 2014Uploaded byPragyan Nanda
- 10.1.1.466.8377.pdfUploaded byPragyan Nanda

- vol50no2rad8Uploaded bySteva Milosev
- BIBDUploaded bydini5793
- Segunda Asignación de Estadística Aplicada a La IngenieríaUploaded byBrian Williams
- Phase Rule and Phase DiagramsUploaded byMazhar Ali
- Tutorial 9 Solutions S1 2015Uploaded bycynthia
- Explain 1Uploaded bySudiv Gulla
- Chap 008Uploaded byChajar Matari Fath Mala
- SPSS_lab2RegressionUploaded bysangeetavjoice
- 06850072 (1)Uploaded byPradeep Kundu
- A Phenology-Based Classificationof Time-Series MODIS Data for Rice Crop Monitoring in Mekong Delta, VietnamUploaded byTâm Trần
- mseschsyllUploaded bySAMEERA
- fferUploaded bylhz
- Tutorials in Statistics-chapter 12 NewUploaded byApam Benjamin
- 7 - 4 - Lecture 6-4 - Applying Generalizations (17-52)Uploaded bymalatesta4
- MLR IntroductionUploaded byAndrew Watson
- 293383817-Predicting-Football-Matches-Using-Neural-Networks-in-MATLAB.pdfUploaded byAlexgri
- Shumway Hazard ModelUploaded byRoshan Raman
- CLB20703 Chemical Engineering Thermodynamics Experiment 4: Centrifugal compressorUploaded bySiti Hajar Mohamed
- Measure of Central TendencyUploaded bydemondida
- Enfa Con GLMUploaded bySantiago Barros
- Lyman, R. L. 1984 - Bone Density and Differential Survivorship of Fossil ClassesUploaded byJosé Camuñas
- 5Other TestsUploaded byProfesor Jorge
- Effects of Lead Exposure, Flock Behavior, And Management of Californian Condor.Uploaded byRodrigo Carral
- holtwintersv1chavcnge_tcm77-238155Uploaded byNeil John Apps
- ASTM D E202Uploaded byphamthuyha
- intro to b-networksUploaded byGeorge190
- Evaluation and Comparison of Sealing Ability of Three Different Obturation Techniques — Lateral Condensation, Obtura II, And GuttaFlow: An in Vitro StudyUploaded byLucia Arguello
- Dastek.pptUploaded byNur Hidayanti
- Effect of High Pressure Processing on Avocado SlicesUploaded byDoraGarcia
- Structural Equation ModellingUploaded byMuhammad Imran Alam