Forecast Pro

Statistical Reference Manual

Eric A. Stellwagen and Robert L. Goodrich

Business Forecast Systems, Inc.

Copyright © 1992-2017 Business Forecast Systems, Inc.

All Rights Reserved Worldwide. No part of this document may be reproduced without the
express written permission of Business Forecast Systems, Inc.

Manual Last Revised: December 5, 2017

Business Forecast Systems, Inc.

68 Leonard Street, Belmont, MA 02478 USA
Phone: 617-484-5050  Fax: 617-484-9219
Email: [email protected]  Web: www.forecastpro.com
Copyright Notice

Copyright © 1992-2017 by Business Forecast Systems, Inc. All Rights Reserved

Worldwide. No part of this manual may be reproduced, transmitted, transcribed, stored in
a retrieval system, or translated into any human or computer language, in any form or by
any means, electronic, mechanical, magnetic, optical, chemical, manual or otherwise,
without express written permission of Business Forecast Systems, Inc., 68 Leonard Street,
Belmont, MA 02478 USA.

Disclaimer

Business Forecast Systems, Inc. makes no representations or warranties with respect to

the contents hereof and specifically disclaims any implied warranties of merchantability
or fitness for any particular purpose. Further, Business Forecast Systems, Inc. reserves the
right to revise this publication and to make changes from time to time in the contents
hereof without obligation of Business Forecast Systems, Inc. to notify any person or
organization of such revision or change.

Trademarks

Forecast Pro is a registered trademark of Business Forecast Systems, Inc.

Other product names mentioned in this manual are trademarks or registered trademarks of
their respective companies and are hereby acknowledged.
Contents

Statistical Reference
Expert Selection ................................................................................................................ 2

Simple Methods ................................................................................................................. 2

Exponential Smoothing .................................................................................................... 3

Conceptual Overview.......................................................................................................... 4
Models of the Exponential Smoothing Family ................................................................... 5
Implementation of Exponential Smoothing in Forecast Pro ............................................... 7
Statistical Description of Exponential Smoothing .............................................................. 9

Discrete Distributions ..................................................................................................... 16

Poisson Distribution .......................................................................................................... 16
Negative Binomial Distribution ........................................................................................ 17

Croston’s Intermittent Demand Model......................................................................... 18

Curve Fitting ................................................................................................................... 19

Box-Jenkins Statistical Models ...................................................................................... 20

Implementation of Box-Jenkins in Forecast Pro ............................................................... 20
Box-Jenkins Background .................................................................................................. 21
Description of the ARIMA Model .................................................................................... 23
Seasonal Models ............................................................................................................... 26
Selecting Model Orders .................................................................................................... 27

Dynamic Regression........................................................................................................ 28
Description of Dynamic Regression Model ...................................................................... 29
Dynamic Regression Diagnostics ..................................................................................... 30

iii
Bass Diffusion Model ...................................................................................................... 32

Forecasting By Analogy .................................................................................................. 33

Model Statistics ............................................................................................................... 34

Box-Cox Power Transforms........................................................................................... 36

Safety Stocks .................................................................................................................... 37

Outlier Detection and Correction .................................................................................. 39

Trading Day Effects ........................................................................................................ 39

Methodology of Automatic Forecasting ...................................... 41

Introduction ....................................................................................................................... 41
Classification of Time Series ............................................................................................ 42
Multiple-level Forecasting ................................................................................................ 43
Incorporation of Additional Information .......................................................................... 44
Selection of Forecasting Method ...................................................................................... 44
Model Selection via Out-of-Sample Testing .................................................................... 46

Glossary......................................................................................... 49

Bibliography .................................................................................. 53

Index............................................................................................... 59

iv
Statistical Reference

This manual describes the statistical techniques, statistics, and strategies that are
implemented in Forecast Pro. It is not necessary that you fully understand, or even read,
this manual in order to produce accurate forecasts with the product.

Those who would like a more thorough coverage of this topic should consult the book
Applied Statistical Forecasting or any of the other texts found in the bibliography.
Applied Statistical Forecasting was written by Dr. Robert L. Goodrich, the author of
Forecast Pro, and is available from Business Forecast Systems.

This chapter begins by presenting each of the forecasting models and concludes with a
discussion of the model statistics presented by the program. The topics are:

Expert selection

Simple methods

Exponential smoothing

Discrete distributions

Croston’s intermittent demand model

Curve fitting

Box-Jenkins

Dynamic regression

Model statistics

1
Expert Selection
Expert selection allows Forecast Pro to select an appropriate univariate forecasting
technique automatically. Expert selection operates as follows.

If the data set is very short, Forecast Pro defaults to simple moving average.

Otherwise Forecast Pro examines the data for the applicability of the intermittent or
discrete forecast models. Although the forecasts produced from such models are just
straight horizontal lines, they often provide forecasts superior to those from exponential
smoothing for low-volume, messy data.

If neither of these models are applicable to the data, the choice is now narrowed down to
different forms of exponential smoothing and Box-Jenkins models. Forecast Pro next
runs a series of tests on the data and applies a rule-based logic that may lead to a model
selections based on data characteristics.

If the rule-based logic does not lead to a definitive answer, Forecast Pro performs an out-
of-sample test to choose between an exponential smoothing model and a Box-Jenkins
model.

There is also a question of selecting the form of the exponential smoothing and Box-
Jenkins models. This procedure is documented in the Implmentation of Exponential
Smoothing in Forecast Pro and Implmentation of Box-Jenkins in Forecast Pro sections of
this manual.

Simple Methods
Forecast Pro supports three variants of the n-term simple moving average, which we
symbolize as SMA(n). The essence SMA(n) is to estimate the current level St of the series
as the average of the last n observations. The level of the series is defined as the value
that the observation would take if it were not obscured by noise.

1 n−1
St = ∑Y
n s= 0 t − s

The forecast for time t+m from the forecast base t is simply a horizontal line at the level
St.

Yt (m) = S t

Confidence limits for SMA(n) are determined by assuming that the true underlying
process is a random walk with observation error.

SMA(n) has one purpose—to decrease the effect of noise on the estimated true value of
the series. It cannot pick up the effects of seasonality or trending. Thus its capabilities are

2 Expert Selection
very similar to those of simple exponential smoothing, except that the model has no
parameters that need to be fitted to the data.

SMA(n) should be used only when the historical data record is so short and so noisy that
it is meaningless to try to extract patterns from the data or even to estimate a smoothing
weight. In any other circumstance, one of the exponential smoothing models will
outperform SMA(n).

Forecast Pro offers three versions of SMA(n)—Automatic, Moving average and Random
walk. Automatic determines the number of terms n in the moving average by determining
the n that minimizes error over the historic sample. Moving average lets the user set n.
Random walk sets n to 1, so that the forecast consists of the last observed value.

Exponential Smoothing
Exponential smoothing is the most widely applicable of the univariate time series
methods to business data. In the absence of information to the contrary, it is probably the
best choice for the typical user.

Although exponential smoothing was first developed over thirty years ago, it is still very
much a hot topic in research circles. If anything, its reputation as a robust, easy to
understand methodology has increased in recent years, often at the expense of
Box-Jenkins.

The main reason for this is that Box-Jenkins models are built upon the abstract statistical
concept of autocorrelation, while exponential smoothing models are built upon clear-cut
features like level, trend, and seasonality. Exponential smoothing models are therefore
less likely to be influenced by purely statistical quirks in the data.

Harvey [1984, 1990] has extended the exponential smoothing approach in his
development of so-called structural models. Structural model forecasts are generated
from a Kalman filter built upon a formal statistical model involving the same features as
exponential smoothinglevel, trend and seasonality. We now recognize exponential
smoothing for what it really isapproximate Kalman filters fitted directly to the data.

This establishes a framework for extending the basic exponential smoothing

methodology. You will see two such extensions in the methodological descriptions
below.

Proportional error models extend exponential smoothing to the case where errors
tend to be proportional to the level of the data. The majority of business data seem
to exhibit this trait.

Event adjustment models extend exponential smoothing to include the estimation

of, and adjustment for, promotional or other nonperiodic events.

Conceptual Overview 3
Conceptual Overview
Exponential smoothing is based upon a structural model of time series data. We assume
that the time series process manifests some or all of the following structural components.

Level. The level of a time series is a smooth, slowly changing, nonseasonal

process underlying the observations. We cannot measure the level directly because
it is obscured by seasonality, promotional events and irregularity (noise). It must
be estimated from the data.

Local Trend. The local trend is the smooth, slowly changing rate of change of the
level. We call it local to emphasize the fact that at each point in time it undergoes
a small but unpredictable change. Forecasts are based on the local trend at the end
of the historic data, not the overall global trend. We cannot measure the trend
directly. It must be estimated from the data.

Seasonal Effects. Additive or multiplicative seasonal indexes represent periodic

patterns in the time series, like the annual patterns in retail sales. Like the level
and the trend, seasonal indexes must be estimated from the data. They are
assumed to undergo small changes at each point in time.

Event Effects. Promotional events influence sales in much the same way as
seasonality but they are not usually periodic. Additive or multiplicative event
indexes are estimated from the data in much the same way as seasonal indexes.
They are assumed to undergo small changes at each point in time.

Random Events. The level, local trend, seasonal and event indexes are all
stochasticthat is their values change unpredictably from point to point in time.
These changes are caused by unpredictable events like the amount by which a
company’s actual profit or loss differs from what was expected. These are often
called random shocks.

Noise. All of the features described so far are components of an ongoing historical
process. Our measurements of the process, however, are usually corrupted by
noise or measurement error. For instance, chewing gum shipments or chewing
gum orders are noisy measurements of chewing gum consumption.

Three of these features—level, random events and noise—are present in every

exponential smoothing model. The remaining three—local trend, seasonal indexes and
event effects—may be present or absent. We identify a model by determining which of
these features should be included to describe the data properly.

Originally, exponential smoothing models were built informally on these features, with
little attention paid to the underlying statistical model. Exponential smoothing equations
were merely plausible means at estimating time series features and extrapolating them.
There was no way to estimate confidence limits properly, since they depend upon the
underlying statistical model.

4 Exponential Smoothing
Some software developers responded to the need for confidence limits with little or no
theoretical justification. While the point estimates from such software have been good,
the confidence limits have been nearly unusable.

Forecast Pro takes a more modern approach to exponential smoothing. Each variant of
exponential smoothing is based upon a formal statistical model which also serves as a
basis for computation of confidence limits. The actual smoothing equations are based
upon the Kalman filter for the formal statistical model. Of course, all of this is under the
hood, and you need not know the details.

Models of the Exponential

Smoothing Family
Here we will provide an overviewwithout equationsof the models that make up the
exponential smoothing family.

Every exponential smoothing model involves at least the following three components.

Level

Random events

Noise

Simple exponential smoothing involves only these components. The data are assumed to
consist of the level, slowly and erratically changing as random events impact it, and
corrupted by noise. Simple exponential smoothing cannot capture the effects of
seasonality or trending.

The remaining components

Trend

Seasonal indexes

Event indexes

are optional. They model features that may or may not be present in the data.

The trend can enter in four waysnone, linear, damped or exponential.

The forecasts from an untrended model are flat, except perhaps for the effects of seasonal
or event indexes.

The forecasts from a linear trend model extrapolate the last estimate of the trend without
limit. The forecasts eventually become positively or negatively infinite.

Models of the Exponential Smoothing Family 5

The forecasts from a damped trend begin almost linearly but die off exponentially until
they reach a constant level This may be appropriate for data influenced by business
cycles. Damped trend models produce forecasts that remain finite.

The forecasts from an exponential trend begin almost linearly but increase as a percentage
of themselves. This explosive growth model should only be used when the data are truly
growing exponentially.

The Holt model includes a linear trend but does not accommodate seasonal or event
effects. The level of the data changes systematically because of the trend. It is also
impacted by random events. The trend varies randomly from point to point as it too is
impacted by random events. Observations are obscured by noise.

Seasonal indexes can enter in three waysnone, additive or multiplicative.

If the indexes are multiplicative, the seasonal adjustment is made by multiplying the
index into the deseasonalized series. Thus the effect is proportional to the level of the
time series. December sales are adjusted upwards by 20% if the seasonal index is 1.2.
This is the most common form of seasonality but it applies only to positive, ratio scale
data.

If the indexes are additive, the seasonal adjustment is made by adding the index onto the
deseasonalized series. Thus the effect is independent of the level of the time series.
December sales are adjusted upwards by 1000 if the seasonal index is 1000.

The multiplicative (additive) Winters exponential smoothing model extracts the level,
trend, and multiplicative (additive) seasonal indexes. The underlying nonseasonal model
is the same as Holt.

Event indexes can also enter in three different waysnone, additive or multiplicative.
The adjustments are analogous to those for seasonal indexes. The difference is that the
adjustment is made each time a certain event occurs rather than tying the adjustment to
the calendar.

Event index models extend the Holt-Winters family of exponential smoothing models,
which includes only the four trend options and three seasonality options, or twelve
models in all. The following figure portrays the forecast profiles of these twelve models.

6 Exponential Smoothing
 Additive Multiplicative
Nonseasonal Seasonal Seasonal

Constant
Level

(Simple)

Linear
Trend

(Holt) (Winters)

Damped
Trend
(0.95)

Exponential
Trend
(1.05)

Forecast Profiles of Exponential Smoothing Models (Gardner [1985])

These forecast profiles are created by extrapolating the level, trend and seasonality index
estimates from the end of the historic data. They depict the underlying patterns of the data
as these patterns exist at the end of the data. They do not and cannot include the effects of
future random events or noise, so they are much smoother than the actual future will turn
out to be.

Exponential smoothing works as its name suggests. It extracts the level, trend and
seasonal indexes by constructing smoothed estimates of these features, weighting recent
data more heavily. It adapts to changing structure, but minimizes the effects of outliers
and noise.

The degree of smoothing depends upon parameters that must be fitted to the data. The
level, trend, seasonal index and event index estimations require one parameter each. If the
trend is damped (or exponential), the damping (or growth) constant must also be
estimated. The total number of parameters that must be fitted to the data depends on the
components of the model.

Implementation of Exponential
Smoothing in Forecast Pro
This section presents some details about the Forecast Pro implementation of exponential
smoothing.

Implementation of Exponential Smoothing in Forecast Pro 7

Model selection

To select a smoothing model automatically, Forecast Pro tries all of the “standard” Holt-
Winters candidate models and chooses the one that minimizes the Bayesian information
criterion (BIC). The BIC is a goodness-of-fit criterion that penalizes more complex
models, i.e., those that require fitting more parameters to the data. Research has shown
that this leads to the model that is likely to forecast most accurately (Koehler and
Murphree [1986]).

To determine the standard candidate models, Forecast Pro applies the following rules:

1. Automatic model selection does not consider exponential trend models due to their
ability to grow explosively in the forecast period. If you wish to build exponential trend
models you must use the custom modeling option.

2. If there are less than 5 data points, then Forecast Pro does not attempt to fit a Holt-
Winters model to the data. A simple moving average model, which does not require
parameter estimation, is substituted.

3. If there is less than two years worth of data, then Forecast Pro Unlimited does not
consider seasonal models.

4. If the data contain negatives or zeroes, multiplicative index models are not considered.

If the NA-CL model is under consideration (by default it is) and/or if seasonal
simplification models are under consideration (by default they are not). Then an out-of-
sample test is used to select amongst the standard model that minimized the BIC and the
NA-CL and/or seasonally simplified models.

Parameter optimization

To estimate model parameters, the program uses an iterative search (simplex method) to
minimize the sum of squared errors over the historic data. The search begins at default
values set by the program. Theoretically, the search could yield a local, rather than the
global, minimum. In practice, the authors know of almost no instances where this has
occurred or where the algorithm has failed to converge.

Confidence limits

Forecast Pro outputs lower and upper confidence limits for exponential smoothing
forecasts. The confidence limits for nonseasonal and additive seasonal models are
computed by making the assumption that the underlying probability model is the specific
Box-Jenkins model for which the exponential smoothing model is known to be optimal
(see Yar and Chatfield [1990]).

The confidence limits for multiplicative seasonal models are computed as described by
Chatfield and Yar [1991]. The error standard deviation is assumed to be proportional
either (1) to the corresponding seasonal index or (2) to the corresponding seasonal index
and the current estimate of the level.

8 Exponential Smoothing
For the nonseasonal models, the error standard deviation is assumed either (1) constant or
(2) proportional to the current estimate of the level. For the additive seasonal models, it is
assumed either (1) constant or (2) proportional to the current estimate of the seasonalized
level.

In each case, Forecast Pro decides which option to use by determining which fits the
historical data more closely.

These confidence limits are useful guides to expected model performance, but they are
not perfect, since the actual underlying probability model of the data is not known. Their
usefulness for multiple-step forecasts deteriorates when the historical errors appear to be
correlated.

Notice that the Chatfield-Yar confidence limits differ somewhat from those based on the
underlying Box-Jenkins models.

Statistical Description of
Exponential Smoothing
Each of the smoothing techniques uses recursive equations to obtain smoothed values for
model components. Simple uses one equation (level), Holt uses two (level and trend),
Winters uses three (level, trend and seasonal). Event index models require an additional
equation. Each equation is controlled by a smoothing parameter. When this parameter is
large (close to one), the equation heavily weights the previous values in the seriesi.e.,
the smoothing process is highly adaptive. If the parameter is small (close to zero), the
equation weights previous values decreasingly far into the pasti.e., the smoothing
process is not highly adaptive.

The following table defines the notation that will be used in the detailed discussion of
exponential smoothing. It is adapted from that of Gardner [1985].

m Forecast lead time

p Number of periods per year

Yt Observed value at time t

St Smoothed level at end of time t

Tt Smoothed trend at end of time t

It Smoothed seasonal index at end of time t

Jt Smoothed event index at end of time t

α Smoothing parameter for level of series

γ Smoothing parameter for trend

Statistical Description of Exponential Smoothing 9

 δ Smoothing parameter for seasonal indexes

λ Smoothing parameter for event indexes

ϕ Damped/exponential trend constant

Yt (m) Forecast for time t+m from base t

~
I t +m Most recent seasonal index for time t+m
~
Jt +m Most recent event index for time t+m

The Forecast Pro output calls α the level parameter, γ the trend parameter, δ the seasonal
parameter, λ the event parameter and ϕ the decay/growth constant.

General Additive Index Model

There are twelve exponential smoothing models, so it would not be practical or
interesting to discuss each individually. We will instead discuss the most fully featured
model and how it relates to simpler models.

The most complex additive index model involves the level St, the trend Tt, the seasonal
index It and the event index Jt. The trend is assumed to decay at the rate ϕ ≤ 1. The
observations Yt are assumed to be composed of these components as follows.

Yt = S t + I t + J t + e t

The components St, It and Jt in this equation are the true values for the level, seasonal and
event indexes at the time t. However, they cannot be observed directly but, rather, must
be estimated from the data. This done by using the following recursive equations, which
comprise an approximate Kalman filter for the underlying model. The italicized symbols
now refer to estimates of the true values.
~ ~
St = α (Yt − I t − J t ) + (1 − α ) ( St −1 + ϕ Tt −1 )

Tt = γ ( St − St −1 ) + (1 − γ ) ϕ Tt −1

~ ~
I t = δ (Yt − St − J t ) + (1 − δ ) I t

~ ~
J t = λ (Yt − St − I t ) + (1 − λ ) J t
~
The symbol I t refers to the most up-to-date prior estimate of the seasonal index for the
month (quarter, week) that occurs at time t. If t refers to December, 1993, then this
~
estimate will have been last updated in December, 1992. The symbol J t refers to the most
up-to-date prior estimate for an event of the type that occurs at time t. These equations

10 Exponential Smoothing
 ~ ~
update the prior estimates St-1, Tt-1, I t and J t to reflect the last observation. The posterior
estimates are the quantities on the left hand side of the equations—St, Tt, It and Jt.

All the simpler additive models are, in a sense, contained in these equations.
~
If there is no event at time t, or if event indexes are not wanted, then J t = J t = 0
and the last equation is discarded.

These equations involve a decaying trend. In this case the decay constant ϕ is
usually a little less than one. To convert the model to a linear trend model, just set
ϕ to 1.0. This is equivalent to erasing it from the equations. To convert the model
to an exponential trend model, just set ϕ to a value greater than 1.0.

If seasonal indexes are not wanted, discard the third equation and set St to 0
elsewhere.

If a trend is not wanted, discard the second equation and set Tt to 0 elsewhere.

These equations clearly show how exponential smoothing actually works. Let us look
~ ~
carefully at the first. The quantity Yt − I t − J t represents the current observation, adjusted
for seasonal and event effects by subtracting off their last available prior estimates. The
adjustment yields an estimate of the current level. The quantity St −1 + ϕTt −1 represents the
forecast of the current level St based on information available previous to the last
observation. The first term, based on the current observation, is weighted by α and the
second, based on previous information, is weighted by (1-α).

Each smoothed estimate of the level is computed as a weighted average of the current
observation and past data. The weights decrease in an exponential pattern. The rate of
decrease depends on the size of the smoothing weight α, which thus controls relative
sensitivities to newer and older historic data. The larger the value of the smoothing
parameter, the more emphasis on recent observations and the less on distant.

The parameters γ, ϕ, δ and λ are fitted to the data by finding the values that minimize the
sum of squared forecast errors for the historic data. To compute the sum of squared errors
for trial values of γ, ϕ, δ and λ, the following steps are performed.

The initial values of the four components S0, T0 ,I0 and J0 are set equal to
reasonable guesses based on the data.

The one-step forecast for the first data point t=1 is generated via the equation
~ ~
Y0 (1) = S0 + ϕT0 + I1 + J1 . The forecast error Y1 − Y1 (1) is computed and squared.

This step is repeated for t=2 to the end of the historic data t=T. The forecast
~ ~ ~ ~
formula is Yt (1) = St + ϕTt + I t +1 + J t +1 so the error is Yt − St − ϕTt − I t +1 − J t +1 . As
each point is forecasted, the forecast error is squared and accumulated.

Statistical Description of Exponential Smoothing 11

This procedure is iterated with new trial values of the parameters until the values that
minimize the sum of squared errors are found. The trial parameter values are determined
by the simplex procedure, an especially stable algorithm for nonlinear minimization.

Once the parameters have been estimated by fitting to the data, the model is used to
compute the forecasts. The equation for the forecast of YT+m from the forecast base YT
(last historic data point) is as follows.
 m i
YT (m) = S T + ∑ ϕ  TT + I T + m + J T + m .
 ~ ~
 i =1 

General Multiplicative Index Model

The general multiplicative model looks almost the same as the additive, except that
multiplication and division replace addition and subtraction. The multiplicative equations
are as follows.
Y
S t = α ~ t~ + (1 − α ) ( S t −1 + ϕ Tt −1 )
It Jt

Tt = γ ( S t − S t −1 ) + (1 − γ ) ϕ Tt −1

Yt ~
It = δ ~ + (1 − δ ) I t
St J t

Yt ~
Jt = λ ~ + (1 − λ) J t
St I t

Simpler models are obtained from these equations in much the same way that they are for
the additive case.
~
If there is no event at time t, or if event indexes are not wanted, then J t = J t = 10
.
and the last equation is discarded.

These equations involve a decaying trend. In this case the decay constant ϕ is
usually a little less than one. To convert the model to a linear trend model, set
ϕ equal to 1.0 or simply remove all references to ϕ.

If seasonal indexes are not wanted, discard the third equation and set to 1.0
elsewhere.

If a trend is not wanted, discard the second equation and set Tt to 0 elsewhere.

Now that the full additive and multiplicative smoothing equations have been presented,
we will examine some of the simpler models that they contain as special cases.

12 Exponential Smoothing
Simple Exponential Smoothing
The simple exponential smoothing model is used for data that are untrended, nonseasonal
and not driven by promotional events. We can get its equation from either the general
additive or general multiplicative model by discarding the last three equations and
eliminating the seasonal and event indexes from the first. We are left with the following.

St = α Yt + (1 − α ) St −1 (1)

Notice that when

α = 1. 0
the equation becomes

St = Yt

i.e., there is no “memory” whatsoever of previous values. The forecasts from this model
would simply be the last historic point. On the other hand, if the parameter is very small,
then a large number of data points receive nearly equal weights, i.e., the memory is long.
The other exponential smoothing models use additional smoothing parameters in
equations for smoothed values of trend and seasonality, as well as level. These have the
same interpretation. The larger the parameter, the more adaptive the model to that
particular time series component.

Equation (1) shows how the smoothed level of the series is updated when a new
observation becomes available. The m step forecast using observations up to and
including the time t is given by

Y (m) = St (2)

i.e., the current smoothed level is extended as the forecast into the indefinite future.
Clearly, simple exponential smoothing is not appropriate for data that exhibit extended
trends.

Holt Exponential Smoothing

Holt’s [1957] exponential smoothing model uses a smoothed estimate of the trend as well
as the level to produce forecasts. The forecasting equation is

Y (m) = St + mTt (3)

The current smoothed level is added to the linearly extended current smoothed trend as
the forecast into the indefinite future.

The smoothing equations are

Statistical Description of Exponential Smoothing 13

 S t = αYt + (1 − α )( S t −1 + Tt −1 ) (4)

Tt = γ ( S t − S t −1 ) + (1 − γ )Tt −1 (5)

where the symbols were defined previously. Equation (4) shows how the updated value of
the smoothed level is computed as the weighted average of new data (first term) and the
best estimate of the new level based on old data (second term). In much the same way,
equation (5) combines old and new estimates of the one period change of the smoothed
level, thus defining the current linear (local) trend.

Multiplicative Winters
In multiplicative Winters, it is assumed that each observation is the product of a
deseasonalized value and a seasonal index for that particular month or quarter. The
deseasonalized values are assumed to be described by the Holt model. The Winters model
involves three smoothing parameters to be used in the level, trend and seasonal index
smoothing equations.

The forecasting equation for the multiplicative Winters model is

Y (m) = ( S t + mTt ) It (m) (6)

i.e., the forecast is computed similarly to the Holt model, then multiplied by the seasonal
index of the current period.

The smoothing equations are obtained from the general multiplicative equations by
setting ϕ to 1 and discarding the parts that involve event indexes.

Yt (7)
St = α + (1 − α )( S t −1 + Tt −1 )
I t− p

Tt = γ ( S t − S t −1 ) + (1 − γ )Tt −1 (8)

Yt (9)
It = δ + (1 − δ ) I t − p
St

The level smoothing equation (7) is similar to equation (4) for the Holt model, except that
the latest measurement is deseasonalized by dividing by the seasonal index calculated one
year before. The trend smoothing equations of the two models are identical. The seasonal
index is estimated as the ratio of the current observation to the current smoothed level,
averaged with the previous value for that particular period.

14 Exponential Smoothing
Additive Winters
In additive Winters, it is assumed that each observation is the sum of a deseasonalized
value and a seasonal index. The deseasonalized values are assumed to be described by the
Holt model. The equations for additive Winters are nearly identical to those of
multiplicative, except that deseasonalization requires subtraction instead of division.

The forecasting equation for the additive Winters model is

Yt ( m) = S t + mTt + It ( m) (10)

The smoothing equations are obtained from the general additive equations by setting ϕ to
1 and discarding the event indexes.

St = α (Yt − I t − p ) + (1 − α )( St −1 + Tt −1 ) (11)

Tt = γ ( S t − S t −1 ) + (1 − γ )Tt −1 (12)

It = δ (Yt − St ) + (1 − δ ) I t − p (13)

NA-Constant Level Model

Forecast Pro supports a form of exponential smoothing that we refer to as an NA-
Constant Level model or a “salt” model (a word play on NaCl). The model is a variation
of the Additive Winters whereby the trend term is omitted and the smoothing weight α is
set to a very small value. By constraining α to a small value the model enforces a constant
level and the seasonal component models the departures from the constant level. The
model works particularly well for data that exhibit a “selling season” whereby the
majority of the demand occurs at specific times of the year (e.g., snow shovels, flu
vaccines, etc.).

Seasonally Simplified Exponential

Smoothing
In a standard seasonal exponential smoothing model the number of seasonal indices used
to model the seasonality equals the number of periods per year. For example there would
be 12 seasonal indices for monthly data, 52 seasonal indices for weekly data, etc. Forecast
Pro supports seasonally simplified models whereby the number of indices used to model
the seasonal component is less than the number of periods per year.

The model is particularly useful for modeling noisy weekly data sets, where using 52
seasonal indices can sometimes result in an overly complex seasonal component that fits
to the noise rather than capturing the underlying seasonal pattern.

A seasonally simplified exponential smoothing model substitutes a carefully constructed

event model for the seasonal component. The event schedule is constructed to map each

Statistical Description of Exponential Smoothing 15

period of the year into a seasonal “bucket.” For example if you wanted to reduce the
number of seasonal indices for a weekly data set from 52 to 26 you would map weeks 1
and 2 of each year into event type 1, weeks 3 and 4 into event type 2, etc. The resulting
model would calculate 26 event indices to capture the seasonal pattern. We would refer to
the model as having a “bucket size” of 2 since two weeks each year were put into each of
the 26 buckets. If we used a bucket size of 4 then the weeks 1-4 of each year would map
into event type 1, weeks 5-8 into event type 2, etc. and the resulting model would
calculate 13 event indices to capture the seasonal pattern.

Forecast Pro includes the ability to consider seasonally simplified models as part of
expert selection as well as allowing you to build them as a customized exponential
smoothing model.

Discrete Distributions
Most statistical forecasting models are based on interval data, i.e., data for which zero has
no special meaning. Forecasts and data can be negative as well as positive, and the
interval from zero to one is statistically equivalent to the interval from 100 to 101. Very
little business data are interval in nature but, for the most part , interval data forecast
models still perform well.

But there are exceptions. For instance the data might consist entirely of zeroes and small
integers. Infrequently used spare parts often fall into this class. The forecasts from simple
exponential smoothing for such items may be perfectly reasonable and useful, but the
confidence limits are usually unusable.

This is due to the confidence limits from a standard model being symmetric. They do not
take into account that sales of these types of items cannot go negative but might become
very large. The discrete distributions forecast model produces the same point forecasts
but produces much more accurate confidence limits.

Forecast Pro tries two different discrete distributions to fit the datathe Poisson
distribution and the negative binomial distribution. Forecast Pro selects the distribution
that fits the data better and uses that distribution to compute the forecasts.

Poisson Distribution
The Poisson distribution ranges over integers in the range {0,1,2,...}. It applies to such
processes as the number of customers per minute who arrive in a queue, the number of
auto accidents per month on a given road, or sales of a particular spare part per month.

The probability that exactly x events occur is given by the following formula.

e −λ λx
f ( x) =
x!

16 Discrete Distributions
The Poisson distribution has a single parameter λ that equals both the mean number of
events per unit of time, and the variance around the mean. This parameter λ is a positive
real number. Forecast Pro chooses the Poisson distribution when the ratio of the sample
mean to the sample variance is near unity.

It is likely that the mean number of events per unit of time is actually changing over time.
Therefore we must estimate λ as a time series in its own right. It has been shown by
Harvey [1989] that this is optimally done via simple exponential smoothing. The current
estimate of the level is also an estimate of the current value of λ.

Therefore Forecast Pro performs the following steps.

Use simple exponential smoothing to estimate and forecast λ. The forecasts are
equal to the value of λ at the end of the series.

Use the final value of λ to determine from the equation for the distribution the
probabilities of 0, 1, 2, ... events per unit of time. These in turn are used to
compute integer confidence limits.

The advantage to using a discrete distribution is not an improved point forecast but
improved confidence limits, and the availability of a formula to compute the probability
of zero events, one event, etc.

Negative Binomial Distribution

The variance of many integer series runs higher than can be modeled by the Poisson
distribution, where the ratio of the variance to the mean is unity. For instance, the
mechanical failures that require a certain spare part may be accurately modeled by a
Poisson distribution, but orders for the part may not reflect current failures. The parts may
be inventoried by an intermediate distributor, or the end user may buy more than is
needed immediately. The result is that, while the mean of the parts orders matches the
mean of the failures, the orders are more variable. The negative binomial distribution
allows us to model such data.

The negative binomial distribution ranges over the integers {0,1,2,...}. The probability
that exactly x events occur is given by the following formula.

( x + y − 1)! y
f ( x) = p (1 − p) x
x !( y − 1)!

This is the formula for the probability of x failures before the yth success in a sequence of
Bernoulli trials in which the probability of success at each trial is p. In Forecast Pro, we
regard the negative binomial distribution in a more empirical way. It is flexible enough to
model many discrete business series that are not modeled well by the Poisson
distribution.

The parameters to be fitted to the data consist of y, an integer which assumes values in
the range {1,2,3,...}, and p, which lies in the interval from 0.0 to 1.0. These two
Negative Binomial Distribution 17
parameters are fitted to the data by using the facts that the mean of the distribution is y(1-
p)/p while its variance is y(1-p)/p2. Thus the ratio of the mean to the variance is p. The
mean of the series is estimated via simple exponential smoothing, as with the Poisson
distribution. We assume that the ratio of the variance to the mean is a constant, which we
also estimate from the data.

This gives us the distribution of x at the end of the historic data. The point forecasts equal
the final estimate of the mean. The confidence limits are computed by using the formula
for f(x) as a function of x and the two fitted parameters.

Croston’s Intermittent Demand

Model
Description

Much product data, especially for lower volume items, are intermittent in nature. For
many, or even most periods, there is no demand at all. For periods where there is some
demand, it is randomly distributed independently or nearly independently of the demand
interval. This might be the case for spare parts that are usually ordered in batches to
replenish downstream inventories. The Poisson and negative binomial distributions do
not usually fit such data well because they link the zeroes and non-zeroes as part of the
same distribution.

Croston [1972] proposed that such data be modeled as a two-step process. He assumed
that the demand intervals were identically independently distributed (iid) geometric. This
is equivalent to assuming that the probability that a non-zero demand occurs in any given
period is iid Bernoulli, as though by the flip of an unfair coin. He further assumed that the
distribution of successive demands was iid normal.

The alternative model for data this messy is usually simple exponential smoothing. This
yields horizontal forecasts at a level fitted adaptively to the data. Willemain et al. [1994]
examined the performance of a variant of the Croston model relative to that of
exponential smoothing, and found it markedly superior in forecast accuracy, both for
simulated and real data, even in the presence of autocorrelation and cross-correlation
between demand size and demand interval.

The variation that Willemain et al. introduced was the substitution of the log normal
distribution for the normal distribution for successive order sizes. This is sensible for
most data because the probability of non-positive demand size is zero. However, it cannot
be applied to demand data that sometimes goes negative, as it sometimes does when a
company registers returns as negative demand.

Implementation

Two basic models are implemented in Forecast Pro—the Croston Model as originally
implemented and the Willemain variant. The Willemain variant is always selected unless
there are occasional negatives in the historic data. If numerous data points are negative,

18 Croston’s Intermittent Demand Model

the Croston model is tagged as inappropriate, as it also is when the vast majority of the
historic periods have a non-zero demand.

The quantities that must be estimated from the data include the following.

Probability of a demand in a given period, adaptively estimated to reflect

conditions at the end of the historic data via simple exponential smoothing of the
demand interval. The smoothing parameter is optimized, as is that for the mean
order size, by minimizing the sum of squared fitting errors.

Mean order size, adaptively estimated in the same way.

Standard deviation of the demand size, estimated globally over the historic data
set.

The forecasts are computed as the product of demand probability and demand size. All
three of the estimated quantities are used to compute the overall distribution function,
from which the confidence intervals are computed.

Curve Fitting
Curve fitting is generally used to model the global trend of a data set. Although curve
fitting is not as sophisticated as some of the other Forecast Pro forecasting
methodologies, it can still be quite useful. Unlike some of the other methods, curve fitting
may be used with short time series (the suggested minimum length is ten data points).
Also, the program provides a quick and easy way to identify the general form of the curve
your data are following. Be aware however, that curve fitting methods do not
accommodate for or project seasonal patterns in a series.

The curve fitting routine supports four types of curves—straight line, quadratic,
exponential and growth (s-curve). You can let Forecast Pro choose the form of the curve
or select it yourself.

The automatic tries the four curves and selects the one that minimizes the BIC for the
historic series. The equations for each curve are shown below (t=time). All of the
coefficients of the model are selected to minimize the sum of the squared errors.

Straight line: Y = a + bT

Quadratic: Y = a + bT + cT 2

Exponential: Y = e a +bT

a
Growth Curve: Y=
1 + e − b (T − c )

Negative Binomial Distribution 19

Box-Jenkins Statistical Models
Box-Jenkins is a powerful forecasting technique which, for appropriate data, often
outperforms exponential smoothing. Traditionally, however, Box-Jenkins models have
been difficult and time consuming to build. This has kept them from widespread
acceptance in the business community.

However, automatic algorithms such as those found in Forecast Pro, now allow
forecasters to build Box-Jenkins models quickly and easily. As a result, they are now
candidates for more widespread use.

In the largest empirical studies of forecasting accuracy to date (Makridakis [1982],

Makridakis [2000]), exponential smoothing outperformed Box-Jenkins overall, but in
many specific cases, Box-Jenkins outperformed exponential smoothing. Ideally, a
forecaster would switch between Box-Jenkins and exponential smoothing models,
depending on the properties of the data. This is precisely what the Forecast Pro expert
system is designed to do.

Box-Jenkins models are built directly on the autocorrelation function (ACF) of a time
series variable. Therefore, a prerequisite for Box-Jenkins is that the data should possess a
reasonably stable autocorrelation function. If the autocorrelations are not stable, or the
data are too short (say, fewer than 40 points) to permit reasonably accurate estimates of
the autocorrelations, then exponential smoothing is the better choice. This avoids the
principal pitfall of Box-Jenkins: fitting a complex model to chance historic correlations or
outliers.

Univariate Box-Jenkins cannot exploit leading indicators or explanatory variables. If

these are important, then a multivariate method such as dynamic regression is a better
choice.

Forecast Pro implements the univariate ARIMA (AutoRegressive Integrated Moving

Average) procedure described by Box and Jenkins [1976]. The models can be identified
completely automatically by the program, or the user can interactively build a model, or
test variations on the model selected by the Forecast Pro expert system. The program
supports the multiplicative seasonal model described by Box and Jenkins.

This section is intended to provide a background to the statistical methodology used in

the program. Those who would like a more thorough coverage should consult Box and
Jenkins’ classic theoretical textbook or Applied Statistical Forecasting [Goodrich 1989].

Implementation of Box-Jenkins in
Forecast Pro
Automatic identification

The program begins by performing a range-mean test to determine whether the log or
square root transform should be applied to the data.

20 Box-Jenkins Statistical Models

Next the program determines the simple and seasonal differencing necessary to render the
data stationary. It uses an adaptation of the Augmented Dickey-Fuller test (see Goodrich
[1989]). Then it computes approximate values for the parameters of a group of candidate
models. Forecast Pro tests each model, and selects the one that minimizes the BIC
criterion.

Exhaustive fitting and examination of all low order ARIMA models would take an
inordinate amount of computer time. Forecast Pro actually overfits a state space model,
and uses it to generate approximate Box-Jenkins models quickly. Sometimes this method
misses the minimum BIC by a slight amount, but it virtually never selects a bad model.

Business Forecast Systems has compared its Automatic Box-Jenkins models with the
published results from the M-competition, where an expert spent 20 minutes to identify
each ARIMA model manually. Forecast Pro outperformed the Box-Jenkins expert at
every forecast horizon.

Business Forecast Systems recommends that you use Automatic identification routinely.
Use Custom identification only when the program so suggests, or when you have a strong
reason to reject the automatic model.

Initialization

Forecast Pro uses the method of back-forecasting to initialize Box-Jenkins models. This
technique is described in Box and Jenkins [1976].

Parameter estimation

Forecast Pro uses the method of unconditional least squares to obtain final parameter
estimates. If necessary, the parameters are adjusted to ensure stationarity or invertibility

Constant term

By default, Forecast Pro uses a constant term only when an ARIMA model does not
involve differencing. This is to avoid imposition of deterministic trends, which can lead
to large forecast errors at longer horizons. You can, however, override the default if you
want. In that case Forecast Pro will estimate the constant as though it were another
parameter, so that you can check its statistical significance.

Box-Jenkins Background
Two statistical concepts are pivotal for understanding the Box-Jenkins modeling and
dynamic regression, stationarity and autocorrelation.

Stationarity

A time series process is stationary (wide sense) when it remains in statistical equilibrium
with unchanging mean, unchanging variance and unchanging autocorrelations. A

Box-Jenkins Background 21
stationary process can be represented as an optimal autoregressive moving average
(ARMA) model.

Unfortunately, most business and economic time series are not stationary. There are many
forms of nonstationarity, but the following forms are especially important.

Nonstationarity in the mean.

The mean is not constant but drifts slowly, without consistent direction.

The time series is trended or cyclical. The trend is not constant but slowly drifts
up and down.

Nonstationarity in the variance.

The time series is heteroscedastic, i.e. the variance of observations around the
mean is changing.

One treats these cases by transforming the data to stationarity. Nonstationarity in the
mean is removed by differencing. Nonstationarity in the variance is removed by applying
a Box-Cox power transform.

Autocorrelation function

According to ARIMA statistical theory, a time series can be described by the joint normal
probability distribution of its observations Y1, Y2, ... , YN. This distribution is
characterized by the vector of means and the autocovariance function.

The autocovariance of Yt and its value Yt + m at a time m periods later is defined by

[ ],
γ µ = cov(Yt , Yt + µ ) = E (Yt − µ )(Yt + µ − µ )

where the operator E denotes statistical expectation, cov denotes the covariance, and µ is
the expected value of Yt. Notice that the autocovariance function is a function of the time
separation m, not the absolute times. This is an implicit assumption that the
autocovariance function does not depend on the time origin t. In other words, the time
series is stationary. If it is not, then its autocovariance function is not defined.

Notice that γ 0 is the same as the variance σY . The autocorrelation function is computed
2

by dividing each term of the autocovariance function by the variance σY :

2

E (Yt − µ )(Yt + µ − µ )
ρµ =
σ Y2

The autocovariance function is a theoretical construct describing a statistical distribution.

In practice, we can only obtain estimates of the true values. The generally accepted
formula is

22 Box-Jenkins Statistical Models

  1 T −m 
ρm = E 
T
∑ (Y
t =1
t − Y )(Yt + m − Y )


where is the sample mean. The sample autocorrelation function is then given by

cm
rm =
co

The sampling error of this estimate can be large, especially when the autocorrelations are
themselves substantial. The estimates are also highly intercorrelated. Because of this, one
must use caution in labeling particular correlations significant by visual examination of
the sample autocorrelation function.

The sample autocorrelation function displayed in Forecast Pro includes dashed lines at
the 2σ limits, where σ is the approximate standard error of the sample autocorrelation
coefficient, computed via the Bartlett [1946] approximation. The rate at which σ expands
depends on the sample values of lower order autocorrelations.

Description of the ARIMA Model

Box-Jenkins1 models the autocorrelation function of a stationary time series with the
minimum possible number of parameters. Since the Box-Jenkins dynamic model includes
features (moving average terms) that dynamic regression does not, Box-Jenkins
theoretically will produce the optimum univariate dynamic model. Therefore, even when
a dynamic regression model might ultimately be selected, a preliminary Box-Jenkins
analysis provides a useful benchmark for model dynamics. Since the procedure is quick
and automatic, this puts very little analytic burden on the user.

The Box-Jenkins model uses a combination of autoregressive (AR), integration (I) and
moving average (MA) terms in the general AutoRegressive Integrated Moving Average
(ARIMA) model. This family of models can represent the correlation structure of a
univariate time series with the minimum number of parameters to be fitted. Thus these
models are very efficient statistically and can produce high performance forecasts.

The notation we will use is consistent with that used for exponential smoothing.

N Number of historic data points

m Forecast lead time (horizon)

p Number of periods in a year

Yt Observed value at time t

1Properly,Box-Jenkins refers to the modeling procedure that these two statisticians devised to fit ARIMA
models to data, and not the model itself. In this document, however, we use the two terms almost
interchangeably.

Description of the ARIMA Model 23

 ∇t Differencing operator

∇s Seasonal differencing operator

B Backward shift operator

φi Autoregressive coefficient (lag i). In Forecast Pro this term is

displayed as a[i].

φ(B) Autoregressive polynomial of order p

Φi Seasonal autoregressive coefficient (lag i) In Forecast Pro this

term is displayed as A[i].

Φ(Bs) Seasonal autoregressive polynomial of order ps

qi Moving average coefficient (lag i). In Forecast Pro this term is

displayed as b[i].

q(B) Moving average polynomial of order q

Θi Seasonal moving average coefficient (lag i). In Forecast Pro this

term is displayed as B[i].

Θ(Bs) Seasonal moving average polynomial of order qs

Yt (m) Forecast for time t+m from origin t

et One-step forecast error Y t −Yt −1

εt Normally independently distributed random shock.

Differencing

If a time series is not stationary in the mean, then the time series must first be transformed
to stationarity by the use of differencing transforms. To describe differencing transforms
we use the backward shift operator B, defined as follows.

BYt = Yt −1

B mYt = Yt − m

This operator will be used in our discussion of ARMA processes. For instance, the
differencing operator is defined as follows.

∇ = (1 − B )

24 Box-Jenkins Statistical Models

Autoregressive processes

The AR(p) model is specified by the equation

Yt − φ 1Yt −1 − φ 2 Yt − 2 −...−φ p Yt − p = ε t (1)

in which the dependent variable appears to be regressed on its own lagged values. This
equation can also be represented in terms of the backward shift operator B as

(1 − φ1B − φ 2 B2 −... − φ p B p )Yt = ε t (2)

and, if we adopt the notation φ(B) for the polynomial in B, it can be written succinctly in
the form

φ( B)Yt = ε t (3)

Moving average processes

The Moving Average process MA(q) is given by

Y = ε t − q 1ε t −1 − q 2 ε t − 2 − q q ε t − q (4)

or, alternatively, in the polynomial form

Yt = θ ( B)ε t (5)

The pure moving average process MA(q) is virtually never observed in real world data. It
describes the unlikely process whose autocorrelations are nonzero for q lags, and zero
thereafter.

Moving average terms are, in practice, used only in conjunction with differencing or
autoregressive terms. In that case, they are invaluable. They induce data smoothing just
like that of exponential smoothing.

ARMA and ARIMA processes

The AutoRegressive Moving Average process ARMA(p,q) combines the features of the
AR(p) and MA(q) processes. In polynomial form, it is given by

φ( B)Yt = θ( B)ε t (6)

Thus the AR(p) process is the same as ARMA(p,0) and the MA(q) process is the same as
ARMA(0,q).

Any stationary time series can be modeled as an ARMA(p,q) process. Any time series
that can be made stationary by differencing d times can be modeled as an ARIMA(p,d,q)
process. The ARIMA(p,d,q) model is given by the following equation.

φ( B)(1− B)d Yt = θ( B)ε t (7a)

Description of the ARIMA Model 25

This is the most general nonseasonal Box-Jenkins model.

Deterministic trends

By default, Forecast Pro does not include a constant term in an ARIMA model except
when the model does not involve differencing. If you dictate that a constant term be used
the equation for the model now takes the form shown in equation (7b).

φ( B)(1− B)d Yt = θ( B)ε t + c (7b)

The effect of a constant term is to introduce deterministic trending into your model, in
addition to its other properties. If you have differenced once, the trend is linear; if you
have differenced twice, it is quadratic.

This is usually undesirable because it extrapolates the global trend of the historic data
indefinitely into the future, even when the current trend is slight. This usually produces
poor forecast accuracy for longer horizons. Business Forecast Systems has confirmed this
effect by testing over the 111 Makridakis time series.

Seasonal Models
Equation (7a) is adequate to model many seasonal series, provided that the polynomials
“reach back” one or more seasonal periods. This means that either p or q (or both) must
equal or exceed the seasonal period s. Since all intervening terms would also be included,
such a model is not parsimonious, i.e., it would contain unnecessary coefficients to be
estimated. This is often damaging to predictive validity of the model.

On the other hand, we might consider a seasonal version of equation (7) in which the
backward shift operator B is replaced by its seasonal counterpart Bs. The resulting
equation is

Φ ( Bs )(1− Bs ) D Yt = Θ ( Bs )ε t (8)

where the polynomials q and Θ are of orders P and Q respectively. This is the
ARIMA(P,D,Q)s model. It relates the observation in a given period to those of the same
period in previous years, but not to observations in more recent periods.

The most general seasonal model includes both seasonal and simple ARIMA models at
once. The following equation describes the multiplicative seasonal ARIMA model.

φ( B)Φ ( Bs )(1 − B)d (1 − Bs ) D Yt = θ( B)Θ ( Bs )ε t (9)

It is usually symbolized as ARIMA (p,d,q)(P,D,Q).

26 Box-Jenkins Statistical Models

Selecting Model Orders
The hardest part of Box-Jenkins modeling is deciding which ARIMA(p,d,q) model fits
the data best, i.e. in identifying the degree of differencing d, the AR order p, and the MA
order q. Much of Box and Jenkins [1976] is devoted to this so-called “identification”
problem. The Forecast Pro expert system, in fact, identifies the model automatically .
Therefore, the remainder of this section, in which the original Box-Jenkins procedure is
presented, is inessential.

The original Box-Jenkins procedure depends upon graphical and numerical analysis of
the autocorrelation function and the partial autocorrelation function. It is a pattern
recognition procedure that requires skill and patience to learn. We will discuss only the
nonseasonal case.

Degree of differencing

The identification procedure begins by determining the degree of differencing d that is

required to make the original data Yt stationary. This is done through examination of the
autocorrelation function rk.

The first few lags of the autocorrelation function of the raw data Yt are inspected; if these
die out relatively quickly, then no differencing is required, i.e. d=0. If not, then the
original data are replaced by its first difference ∇Yt and the process is repeated. If the
autocorrelation function of the differenced data dies out quickly, d=1. If not, the data are
differenced a second time to obtain ∇ 2Yt . This process is repeated until, for some d, the
autocorrelation function of the multiply differenced data does die out quickly. In practice,
d is rarely greater than 2.

Once the degree of differencing is determined, the remainder of the analysis deals with
the stationary series ∇ dYt . If d is zero, these are the original data.

Autoregressive order

The autoregressive order p is determined by inspection of the sample partial


autocorrelation function φ kk . We will motivate the odd notation and the definition of this
function through a thought experiment.

Suppose that the process is thought to be purely autoregressive (q=0). Then a rational
strategy to determine p would be to compute a regression of Yt on its first lag, then on its
first two lags, and so on until the last lag introduced into the regression turns out not to be
statistically significant. This is determined by a statistical test on φ kk , which is defined as
the coefficient of Yt − k in a regression on Yt −1 , Yt − 2 ,..., Yt −1 , i.e. the k’th coefficient of the
k’th regression.

Actually, a fast recursive algorithm is used instead of performing so many regressions. A

graph is presented of the first forty-eight lags so that the AR order can be determined. If

Selecting Model Orders 27

the process is ARIMA(p,d,0), then the partial autocorrelation function dies abruptly after
p lags.

Moving average order

A pure moving average process ARIMA(0,d,q) exhibits the same behavior in the
autocorrelation function that the autoregressive process ARIMA(p,d,0) does in the
partial autocorrelation function. In other words, if the process is ARIMA(0,d,q), then the
sequence rk is large for k<q+1 and small for k>q. Thus the autocorrelation function is
used for MA processes in the same manner as the partial autocorrelation function for AR
processes.

The functions rk and φ kk also exhibit similar behavior for ARIMA(p,d,0) and
ARIMA(0,d,q) processes, respectively. Instead of abruptly cutting off at p and q,
respectively, these functions tail off smoothly in exponential decay or exponentially
damped sine waves. By examining both functions, the forecaster can determine the orders
of pure AR and MA processes.

Mixed processes ARIMA(p,d,q) are more complex. Neither the partial autocorrelation
function nor the autocorrelation function abruptly dies out. Instead, the autocorrelations
remain large for k ≤ q+1 and die out exponentially thereafter. The partial autocorrelations
remain large for k<p+1 and die out for k>p. Manual identification of mixed ARIMA
processes is often very difficult.

There are two severe problems with this procedure for order identification.

First, even when the data really does fit an ARIMA process, the sample
autocorrelations used to identify the process can be very different from the
theoretical ones due to sampling variation.

Second, the actual data usually contain outliers and other unmodelable features
that can significantly distort the autocorrelation and partial autocorrelation
functions. It is our judgment that the Box-Jenkins procedure should be used only
as the very roughest guide.

We recommend that you fit an automatic Box-Jenkins model first. Then, if you suspect
that you can find a better model, you can try variations of the automatic model. You can
use the BIC criterion to make a final decision. Note that the Forecast Pro automatic
identification method has bested human experts in several academic studies.

Dynamic Regression
Forecast Pro dynamic regression supports the development of forecasts that combine
time-series-oriented dynamic modeling and the effects of explanatory variables or leading
indicators. The conventional regression model is enhanced by including support for an
extension of the Cochrane-Orcutt autoregressive error model, and for the use of lagged
dependent and independent variables. Forecast Pro does not support the development of
simultaneous equation models.

28 Dynamic Regression
Dynamic regression should be used when (1) the data are long enough and stable enough
to support a correlational model (2) explanatory variables result in a definite increase in
accuracy of fit and (3) reliable forecasts for the explanatory variables are available.
Remember that complex models often produce forecasts that are less accurate than those
from simpler models, even though they may fit the historic data better.

Description of Dynamic Regression

Model
The ordinary least squares dynamic regression model takes the form

P( B)Yt = βX t + et (1)

where the errors et are independently identically normally distributed. The symbols in this
equation and the equations to follow are defined in the table below.

N Number of historic data points

M Forecast lead time (horizon)

s Number of periods in a year

Yt Observed value at time t

Xit Observed value of i’th explanatory variable at time t

B Backward shift operator

φi Autoregressive coefficient of Yt-i

ρi Autoregressive coefficient of et-i

βi Coefficient of Xi

Yt (m) Forecast for time t+m from origin t

et One-step forecast error Y t −Yt −1

The lags of the dependent variable are contained in the polynomial P(B), just as in the
Box-Jenkins model. The dynamic regression model differs from Box-Jenkins in two
important ways:

It includes one or more independent variables, which drive the process. For
example, advertising or promotion usually drive sales.

Description of Dynamic Regression Model 29

 Equation (1) does not include moving average terms, which are often very useful
in Box-Jenkins. Regression models will therefore be less parsimonious than Box-
Jenkins for some processes.

Thus dynamic regression is stronger than Box-Jenkins in one way and weaker in another.

It will often be found that the errors obtained from equation (1) are correlated, contrary to
assumption. This can be determined by examination of diagnostics in the dynamics
module. This may indicate that additional lags of the dependent variable should be
introduced, or additional independent variables or new lags of existing independent
variables should be introduced, or both.

The generalized Cochrane-Orcutt model is an alternative way to improve model dynamics

that often requires estimation of fewer new parameters. In the Cochrane-Orcutt model,
equation (1) is replaced by the pair of equations

P( B)Yt = βX t + ν t (2)

R( B)ν t = et (3)

in which the raw residuals are correlated via an autoregressive process specified by the
polynomial R(B) in the backward shift operator. Equations (2) and (3) can be rewritten as
a single equation

R( B) P( B)Yt = R( B) βX t + et (4)

Dynamic Regression Diagnostics

A regression model is far harder to fit to the historic data than a Box-Jenkins model for
several reasons. First, the dynamic portion of the model (lagged dependent variable and
Cochrane-Orcutt terms) must be determined term by term on the basis of hypothesis
testing, rather than automatically. Second, there are no moving average error terms in
dynamic regression; if they are needed they must be approximated by additional
complexity in the dynamic regression model. Third, the explanatory portion of the model
adds an additional layer of complexity over the univariate case. Moreover, the lag
distribution of the explanatory variables must also be considered. As a result there may be
hundreds of specific terms that should be considered in a particular model.

This complex situation calls for an orderly and systematic strategy. The Forecast Pro
regression diagnostics are modularized into three batteries of tests aimed at two phases of
the model development process. These phases are:

Development of the dynamic model

Development of the explanatory model

The dynamic regression test battery provides specific diagnostics for the current model.
Most of the diagnostics are chi-squared statistics based on Lagrange multiplier tests

30 Dynamic Regression
(Engle [1984]). Lagrange multiplier tests are asymptotically equivalent to the more
commonly used Wald tests and likelihood ratio tests. The following paragraphs describe
the tests.

Each diagnostic tests for a specific deficiency in the model. However, they are not
independent of each other. A deficiency in one specific area can cause several other test
statistics to become significant as well. Because of this it is best to find the test where the
null hypothesis is rejected at the highest probability, and make that one specific change.
Then, reexamine the diagnostics for the altered model.

Dynamics specification

The first group of diagnostics tests for inclusion of Cochrane-Orcutt autoregressive error
terms. The tests are described below.

_AUTO[-n]. The alternative hypothesis is that an error autocorrelation of lag n

should be added to the model. Forecast Pro performs a test for each of the first
twelve lags and the first two seasonal lags. A test is omitted if the term is already
in the model.

The remaining dynamics tests check for inclusion of lagged dependent variables.

Y[-n] test. The alternative hypothesis is that the n’th lag of the dependent variable
should be added to the model. Forecast Pro performs a test for each of the first
twelve lags and the first two seasonal lags. It uses the actual name of the variable.
A test is omitted if the term is already in the model.

The program recommends that some specific new term be added to the model, unless all
tests are insignificant at the level 0.01.

Variable specification

The variable specification tests check for problems in specification of the independent
variables. The tests are described below.

Excluded variables. A Lagrange multiplier test is computed for each inactive

variable on the script.

Time trend. The alternative hypothesis is that a linear time trend improves the
model. A significant test does not necessarily indicate that a time trend variable
should be added. The problem often lies with model dynamics or by the exclusion
of some other variable.

Constant term. The alternative hypothesis is that a constant term improves the
model.

Dynamic Regression Diagnostics 31

 Lagged independent variables. A test is made for each independent variable
now present in the model. The alternative hypothesis is that its first lag should
also be in the model.

Custom excluded variable tests

The alternative hypothesis in the excluded variables test described above is that the model
should include the single additional variable specified. The custom excluded variables
test option allows you to test combinations of excluded variables.

It is not uncommon that combinations of variables will be jointly significant even when
they are separately insignificant.

Bass Diffusion Model

The Bass diffusion model is a new product forecasting technique that can be used with or
without historic demand data. The Bass model is most often used to forecast first time
adoptions of new-to-world products.

The model tries to capture the adoption rates of two types of users—innovators and
imitators. Innovators are early adopters of new products and are driven by their desire to
try new technology. Imitators are more wary of new technology—they tend to adopt only
after receiving feedback from others.

Yt = p(m − ∑ Yt ) + q( ∑ t )(m − ∑ Yt )
Y (1)
m
Yt Number of adopters at time t

m Number of potential adopters over entire life cycle

p Coefficient of Innovation

q Coefficient of Imitation

The Bass model can be written in several different forms. The form in equation (1) is
adapted from Kahn[2006]. Notice the plus sign on the left hand side of the equation
separates the innovation component from the imitation component. Conceptually
equation (1) can be thought of as:

Yt = (p * Remaining Potential) + (q * Current Adopters * Remaining Potential) (2)

Equation (2) illustrates how p defines the strength of the Innovation Effect and q defines
the strength of the Imitation Effect.

If you have 5 or more historic data points, p, q and m can be fit to the data using
regression. Consult Bass[2004] for details.

32 Bass Diffusion Model

With fewer than 5 historic data points , p, q and m must be input into the model. In these
instances, the coefficients could be set using values from an analogous product’s model.
There is also a considerable body of literature on the Bass model including published
coefficients for different types of technologies. Consult Lilien, Rangaswamy and Van den
Bulte[1999].

Forecasting By Analogy
By Analogy is a new product forecasting technique that can be used with or without
historic demand data. The approach is sometimes also referred to as “looks like” analysis.

The concept is a very simple one. You are launching a new product and you expect the
initial sales pattern to be similar to an analogous product’s initial sales pattern or to a
“launch profile” that you’ve created.

If the product has not yet launched (i.e., there is no historic data available) then you must
supply an estimate of the initial sales over a specific period of time (the “launch total”
over the “launch horizon”). Forecast Pro will then create the forecast by proportionally
allocating the launch total over the launch horizon using the analog series to define the
proportions.

If historic data exists, Forecast Pro will calculate and display an “estimated launch total”.
To do so, it first uses the analog series to determine the cumulative percentage of the
launch total that the available historic data represent, it then assumes that the sum of the
available history equals that cumulative percentage and estimates the launch total. For
example, if there are 5 historic demand observations that sum to 500 and the sum of the
first 5 periods of the analog series corresponds to 40% of the analog series’ launch total,
then 500 is assumed to equal 40% of the estimated launch total and thus the estimated
launch total equals 1,250.

If historic data exists and you specify that the estimated launch total should be used to
generate the forecast, Forecast Pro will create the fitted values and forecasts by
proportionally allocating the estimated launch total over the launch horizon using the
analog series to define the proportions.

If historic data exists and you specify a launch total to be used, Forecast Pro will subtract
the sum of the available history from the specified launch total to ascertain the cumulative
forecast needed so that the sum of the available history and forecast will equal the
specified launch total. It then spreads the needed cumulative forecast value using the
analog series’ forecast values to define the proportions. The same proportionality factors
used to generate the forecasts are then used to generate the fitted values—thus the fitted
values represent the historic volume that would normally be associated with the forecast.

Dynamic Regression Diagnostics 33

Model Statistics
Within-sample statistics are displayed each time a model is fitted to the data. Out-of-
sample statistics are displayed whenever a hold out sample is used. Each statistic is listed
below:

Sample size. The number of historical data points used to fit the model. Operations that
discard data points, (e.g., differencing, inclusion of lagged variables, etc.) can reduce this
statistic.

Number of parameters. The number of fitted parameters (coefficients) in the model.

Mean. The sample mean (average) for the historical data.

1
Y = ∑ (Yt )
n

Standard deviation. A measurement of the dispersion of the historical data around its
mean.

1
∑
2
S= (Yt − Y )
n −1

R-square. R-square is the fraction of variance explained by the model.

R = 1−
2 ∑ (Yt − Ft ) 2
∑ (Yt − Y ) 2
Adjusted R-square. The adjusted R-square is identical to the R-square except that it is
adjusted for the number of parameters (k) in the model.
2 n −1
R = 1 − (1 − R 2 )
n − k −1

Durbin-Watson. The Durbin-Watson d-statistic is used to test for correlation of adjacent

fitted errors, i.e. for first-lag autocorrelation. If T is the number of sample points and et is
the fitted error at point t, then d is computed as follows.
T

∑( et − et − 1 )
2

d= t =2
T

∑
t =1
et2

While the d-statistic is easy to compute, it is hard to interpret.

The null hypothesis is that the first-lag autocorrelation is zero. One looks up the Durbin-
Watson bounds dL and dU for sample size T and significance α in a table. The null is
accepted if d < d L and rejected if d > dU . If d L < d < dU , then the test is inconclusive. Our
recommendation, with which many disagree, is to reject the null only when the test is
conclusive.

34 Model Statistics
Another problem is that the d-statistic is not strictly valid for models with lagged
dependent variables. In that case, many statisticians use the Durbin h-statistic instead. The
Durbin h is not reported in Forecast Pro.

We recommend that you rely on the Ljung-Box test, which is straightforward, and on
visual examination of the error autocorrelation function.

Ljung-Box test. The Ljung-Box Q-statistic, which is used to test for overall
autocorrelation of the fitted errors of a model, is a statistical improvement on the Box-
Pierce (portmanteau) test. If T is the number of sample points, ri is the i’th
autocorrelation coefficient, and L the number of autocorrelation coefficients, then Q is
computed as follows.
L
r2
Q = T ( T + 2 )∑ i
i =1 ( T − i )

The statistic is a weighted sum of squared autocorrelations, so it is zero only when every
autocorrelation is zero. The more autocorrelation, the greater the size of Q. The weights
are selected to make Q approximately Χ 2 ( L − n) , i.e. Chi-square with L-n degrees of
freedom.

Forecast error. The standard forecast error is the root mean square of the fitted errors
adjusted for the number of parameters (k) in the model. It is used to compute the
confidence limits of the forecasts, but, realistically, it is usually an overly optimistic
estimate of true out-of-sample error.

FE =
∑ (Y t − Ft ) 2
n−k
BIC. The AIC (Akaike Information Criterion) and the BIC (Bayesian Information
Criterion) are the two order estimation criteria in most common use. A specific model is
selected from a model family by finding the model that minimizes the AIC or BIC.

Either statistic rewards goodness-of-fit, as measured by the root mean square error s, and
penalizes for complexity, i.e. the number of parameters n. Koehler and Murphree [1986]
showed that, for series from the M-competition, the BIC leads to better out-of-sample
forecast performance and, for this reason, Forecast Pro uses and displays the BIC.

There are several equivalent versions of the BIC, related to each other by transforms. In
Forecast Pro, we use the following equation, in which T represents the sample size.
n
BIC = sT 2T

This version of the BIC is scaled the same as the standard forecast error. It can very
loosely be interpreted as an estimate of out-of-sample forecast error.

The BIC can be used to compare different models from the same family, and for the same
data. Since it is scaled to the standard forecast error, it is meaningless as an absolute
criterion of merit.

Dynamic Regression Diagnostics 35

MAPE. The MAPE (Mean Absolute Percentage Error) is used to measure within sample
goodness-of-fit and out-of-sample forecast performance. It is calculated as the average of
the unsigned percentage errors.
1 | Y − Ft |
MAPE = ∑ t
n | Yt |
SMAPE. The SMAPE (Symmetric Mean Absolute Percentage Error) is a variation on the
MAPE that is calculated using the average of the absolute value of the actual and the
absolute value of the forecast in the denominator. This statistic is preferred to the MAPE
by some and was used as an accuracy measure in several forecasting competitions.
1 | Yt − Ft |
SMAPE = ∑
n (| Yt + Ft |) / 2
RMSE. The RMSE (Root Mean Square Error) is used to measure within sample
goodness-of-fit. It is calculated as the square root of the average of the squared errors.
1
RMSE =
n
∑ (Yt − Ft ) 2
MAD. The MAD (Mean Absolute Deviation) is used to measure within sample
goodness-of-fit and out-of-sample forecast performance. It is calculated as the average of
the unsigned errors.
1
MAD = ∑ Yt − Ft
n
MAD/Mean Ratio. This MAD/Mean ratio is an alternative to the MAPE that is better
suited to intermittent and low-volume data. Percentage errors cannot be calculated when
the actual equals zero and can take on extreme values for low volume data. These issues
become magnified when you start to average MAPEs over multiple time series. The
MAD/Mean ratio tries to overcome this problem by dividing the MAD by the Mean—
essentially rescaling the error to make it comparable across time series of varying scales.
The statistic is calculated exactly as the name suggests—it is simply the MAD divided by
the Mean.

GMRAE. The GMRAE (Geometric Mean Relative Absolute Error) is used to measure
out-of-sample forecast performance. It is calculated using the relative error between the
naïve model (random walk) and the currently selected model. A GMRAE of 0.54
indicates that the size of the current model’s error is only 54% of the size of the error
generated using the naïve model for the same data set.

Box-Cox Power Transforms

It is assumed in both Box-Jenkins and dynamic regression that the error process et is
independently identically normally distributed. Heteroscedasticity of the error process, i.e.
changing variance, is a violation of this assumption. The effect of heteroscedasticity is not
so damaging as autocorrelation because it does not bias estimates of the coefficients. Its
main effect is to reduce statistical efficiency, so that the effect of sampling errors is
greater.

36 Box-Cox Power Transforms

Sometimes heteroscedasticity can be eliminated or reduced by transforming the
dependent variable. The transformed variable is forecasted, then back transformed to the
original distribution. The most important are the power transforms analyzed by Box and
Cox [1964]. The following equations are used to transform the original data Yt to the
transformed data Yt(λ).

Yt λ − 1
Yt ( λ) = (λ ≠ 0 )
λ
Yt(0)=ln(Yt)

The parameter λ specifies the power to which the data are raised, except when it is zero.
In that case, Yt is replaced by its logarithm. The first of the two equations includes
constant terms to make the transform a continuous function of λ.

The Box-Cox power transform can be applied only to positive data.

Safety Stocks
Forecast Pro generates safety stock calculations in addition to point forecasts and
confidence limits. This capability is most often used in setting inventories which are
replenished only at certain variable or fixed intervals.

Inventory control analysis requires the manager to balance inventory holding costs,
reorder costs and other factors to determine economic order sizes and reorder points to
maintain a desired service level at minimum cost. The analysis must take into account the
lead time between placing an order and placing the units in stock. Although such analyses
can become very complex, most of them require answering a question similar to the
following.

How much stock do I need to maintain a service level of 95% if the reorder lead time is
four weeks?

At each point of time, the manager needs enough stock so that the total sales for the next
four weeks will exceed the stock level only five percent of the time. It is easy to calculate
the expected demand over the four week period—just add the forecasts over the four
weeks. The difficulty lies in computing the probability that sales will exceed the
cumulative forecast by some certain amount. To determine this mathematically is a
complex problem that depends upon the details of the statistical forecast model. Most
MRP systems use a very crude approximation to solve this problem and really must, since
the system does not know anything about the statistical forecast model. The difficulty of
the calculation lies in taking into account the serial correlations of sales from point to
point over the reorder cycle.

Forecast Pro is unique in providing a rigorous statistical solution to this problem. It does
so by converting the model to an equivalent but different form called the Wold
representation. This is the key to determining the statistical distribution of the cumulative
forecast. Consult Wold[1938] for details of the computation.
Dynamic Regression Diagnostics 37
But there is one caveat, the computation assumes that the statistical distribution of future
sales is in fact correctly captured by Forecast Pro. This is never absolutely true, so the
safety stocks will be in error to the extent that the Forecast Pro model does not actually
capture the true model.

Consider the following example:

Date 5% Lower Forecast 95% Upper

2009-35 640 757 875
2009-36 654 774 895
2009-37 642 761 880
2009-38 574 680 787
2009-39 572 679 786
2009-40 531 631 731

The forecast for each week represents the mean of all possible futures. Put another way, it
is equally likely, according to the model, that the actual value will be above or below the
forecast.

The upper and lower confidence limits provide information about the spread around the
forecast for a given period. The 95 percent upper confidence limit for week 35 is 875.
Thus, according to the model, actual sales for week 35 should fall at or below 875, 95%
of the time. The 95 percent upper confidence limit for week 38 is 787. Thus, according to
the model, actual sales for week 38 should fall at or below 787, 95% of the time.

Notice that forecasts and confidence limits do not take into account lead times. Therefore,
they cannot be used to answer our question, “How much stock do I need to maintain a
service level of 95% if the reorder lead time is four weeks?”

Lead time DDLT 95% Safety Reorder Point

1 757 117 875
2 1,532 174 1,706
3 2,292 215 2,507
4 2,973 244 3,217
5 3,652 269 3,920
6 4,283 289 4,572

The expected Demand During Lead Time (DDLT) is the cumulative forecast. Thus, for a
lead time of 4 weeks, the DDLT is 2973 (757+774+761+680).

The safety stock is the excess stock needed, above and beyond the DDLT, to maintain the
service level specified for the upper confidence limit percentile. The safety stocks are
output for each lead time up to and including the forecast horizon. Thus, to determine the
stock required for a four-week lead time, you would add together the DDLT and Safety
Stock values for lead time 4 (2973+244=3217). This quantity is known as the Reorder
Point. If your stock falls below the Reorder Point then you do not have enough stock to

38 Safety Stocks
satisfy the expected demand at the specified service level and need to obtain additional
stock (e.g., to reorder).

Outlier Detection and Correction

An outlier is a data point that falls outside of the expected range of the data (i.e., it is an
unusually large or small data point). If you are forecasting a time series that contains an
outlier there is a danger that the outlier could have a significant impact on the forecast.

One solution to this problem is to screen the historical data for outliers and replace them
with more typical values prior to generating the forecasts. This process is referred to as
outlier detection and correction.

Correcting for a severe outlier (or building an event model for the time series if the cause
of the outlier is known) will often improve the forecast. However if the outlier is not truly
severe, correcting for it may do more harm than good. When you correct an outlier, you
are rewriting the history to be smoother than it actually was and this will change the
forecasts and narrow the confidence limits. This will result in poor forecasts and
unrealistic confidence limits when the correction was not necessary.

It is the authors’ opinion that outlier correction should be performed sparingly and that
detected outliers should be individually reviewed by the forecaster to determine whether a
correction is appropriate.

Forecast Pro incorporates an automated algorithm to detect and (optionally) correct

outliers.

The detection/correction algorithm works as follows:

1. The specified forecasting model is fit to the time series, the residuals (fitted errors) are
generated and their standard deviation is calculated.

2. If the size of the largest error exceeds the outlier threshold, the point is flagged as an
outlier and the historic value for the period is replaced with the fitted value.

3. The procedure is then repeated using the corrected history until either no outliers are
detected or the specified maximum number of iterations is reached.

In a multiple-level problem the detection is only performed on the end items (i.e., the
nongroup level). If the correction option has been selected, after all end items are
corrected, the group level totals are reaggregated to reflect the corrected values.

Trading Day Effects

In many cases, the sales or shipments of a product depends upon calendar effects such as
the number of working days, the number of weeks in the period (4-4-5 data), or the
number of weekend days. In more complex cases, each day of the week is associated with

Dynamic Regression Diagnostics 39

a weight that describes its relative contribution to sales or shipments, so the weight for the
month can be computed from the relative contributions of the days that fall into that
particular month.

Trading day effects of this kind can influence sales by as much as five to ten percent.
Since it is one variable that can be computed accurately for the future as well as the past,
it makes a good deal of sense to account for it. This will usually give a boost in accuracy
at very little extra effort or expense.

Forecast Pro supports a weighting transformation (\WGT=) that takes trading day effects
into account in a very simple way.

The trading day weights (both past and future), must be defined as a helper
variable in a file supplied by the user. The data must span from the first historic
data period to the last forecast period..

The actual historic values for a time series are adjusted by dividing through by the
weight for each month. This gives estimates for the sales that would have
occurred in the absence of trading day effects.

Forecasts are prepared from the adjusted historic data.

The forecasts are multiplied by the corresponding future trading day weights.

The user must compute and supply to Forecast Pro Unlimited the appropriate trading day
weights for each month (or quarter)

40 Trading Day Effects

Methodology of Automatic Forecasting

This chapter describes special methodological considerations that apply to automatic

forecasting of hundreds or thousands of items.

Introduction
Much of the forecasting literature (and much of the available software) concentrates on
forecasting time series in an R&D environment. The literature envisions the forecaster as
intensely interested in just one or two complex time series, perhaps for long forecast
horizons, and often with highly significant consequences. For example, one might be
interested in forecasting the economic or social environment for a nuclear power
generation plant. The forecaster is willing, under such circumstances, to invest
considerable time and other resources to obtain the best available forecasting models, and
might be willing to use extremely complex methodology.

This emphasis on the R&D environment ignores the forecaster who needs to forecast
hundreds or thousands of products, on a weekly, monthly or quarterly basis, perhaps for
inventory control or production planning. In this case, the consequences of error in the
forecasts for any particular product may be quite small, although the consequences for
aggregate performance might be large. The methodology of Forecast Pro Unlimited, is
directed squarely at this forecasting environment.

Forecast Pro Unlimited is based on the fairly scanty published research and upon BFS
research and experience. We will summarize some of the facts that have emerged from
the research.

The forecasting methods that succeed best are relatively simple ones. Product data is
often so volatile that more complex models, no matter how well they fit the historical
data, yield inferior forecasts.

41
When historic records are relatively long and not very noisy, there is a substantial payoff
in matching the forecast model to each data set individually.

On the other hand, many business time series are extremely noisy. The information in an
individual historic record may not be sufficient to choose the best method reliably.
Frequently, method A may appear to be superior to method B at one time, and inferior to
it at another. In these cases, overall accuracy may be improved by selecting a model at the
group level. This can be done by using the Forecast Pro Unlimited out-of-sample
evaluation procedure.

Forecasting performance and goodness-of-fit of a model to the historic data is certainly

related, but the relationship is much looser than one would expect.

Forecast Pro Unlimited includes extensions (discrete distributions, Croston’s intermittent

data model and multiplicative error model) to the standard confidence limit methodology
to make confidence limits more accurate. Nevertheless, confidence limits are only rough
guides to real out-of-sample forecast performance. You can evaluate real out-of-sample
performance by using the Forecast Pro Unlimited evaluation methodology.

Classification of Time Series

From the earliest days of forecasting systems for production and inventory control, it has
been recognized that there are essentially three types of product time series. This
stratification is based mainly on sales or production volume.

Type A series are very high volume. These series are usually fairly regular, so statistical
forecasting methods like those in Forecast Pro Unlimited perform well. However, these
high volume items are also of great importance to the firm, and the consequences of
forecast error can be significant. Thus, if there are not too many of them, it is wise to
examine them interactively, and to make judgmental adjustments as appropriate.

Type B series are of medium volume. Ordinarily, these series can be forecasted fairly
accurately by the methods in Forecast Pro Unlimited. Since these items are not separately
as crucial to the firm, they lend themselves well to automatic forecasting. Human
intervention is usually required only when the forecasting software marks them as
exceptional.

Type C series are of lowest volume, and may include as many as 50% of the total. Many
of these series will be mostly zeroes, with occasional small sales and, more rarely, a large
sale. The percent error of forecasts of Type C series is often quite large, but the
consequence of error is usually small. When automatic forecasting first emerged, Type C
series were not usually forecasted at all. Instead, a default forecast (say zero or one) was
used, to save computer time (then a scarce resource). Now that computation is cheap,
methods like those in Forecast Pro Unlimited are likely to provide increased accuracy.

Any of these groups can include rogue series, i.e., series that are so irregular as to be
virtually unforecastable. Obviously, most rogue series are of Type C, so that their
practical significance is usually not high. However, their influence on forecast

42 Trading Day Effects

performance evaluations can be great, since evaluations usually depend upon percent
error, not absolute error. The meaningfulness of forecast performance evaluation,
discussed below, will be greatly enhanced if the evaluation is based upon a classification
of series, and the identification of rogue series.

Multiple-level Forecasting
Multiple-level models apply to data which must be dealt with at several levels of
aggregation. Product data, for instance, often involve SKU’s (stock keeping units), brands
and lines. Forecast Pro Unlimited allows you to aggregate data into a hierarchy of groups,
and to produce consistent forecasts at all levels of aggregation.

Consider the product group ABC consisting of the sum of products A, B and C. If one
forecasts each series independently, the forecast of ABC will differ from the sum of the
forecasts of products A, B and C. Often it is essential for the firm, however, to reconcile
such hierarchical inconsistencies.

There are two generally accepted ways to do this—top-down and bottom-up. You will
need to use your knowledge of the products, or testing, to determine which method is
superior.

The top-down method forecasts ABC, A, B and C first, as a preliminary step. Then the
forecasts of A, B and C are adjusted proportionately at each step of time, to insure that
ABC = A+B+C. The bottom-up method is to forecast A, B and C, and to construct the
forecast for ABC by summing the forecasts for A, B and C. Neither method is superior to
the other under all circumstances. If the items are very similar, like sizes and colors of a
product, then the top-down approach is probably more accurate. If they are disparate, like
a household product and a business product, then the bottom-up method is more likely to
succeed.

Many companies employ a disaggregation procedure instead of forecasting each item at

the lowest level. For example, the distribution of shoe sizes and colors is relatively
constant from style to style. It therefore makes sense to forecast at the style level and then
to apply the overall size-color break down to obtain SKU level forecasts. The
disaggregated SKU forecasts would almost certainly be more accurate than those obtained
directly from SKU level data because of its extreme noisiness.

Bunn and Vassilopoulos [1993] showed that it is often more accurate to extract seasonal
indexes at the group level, use them to deseasonalize the item level data, forecast the
resulting nonseasonal series and reseasonalize the forecasts.

For instance, suppose that the lowest level item is a soft drink in a certain size and
container type (glass, plastic, aluminum, etc.). One would extract seasonal indexes after
aggregating the data over size and container type. This is especially useful if the data
includes some new container sizes without adequate history to calculate seasonality
independently. Another situation where this capability is important, is when products are
constantly being replaced with new models. For example, the average life span of a
computer printer model is 18 months, making independent estimation of the seasonal
Multiple-level Forecasting 43
indexes very tenuous. Extracting seasonality at the group level (containing all models past
and present) and applying the indexes to the current models results in more reliable
estimates.

Forecast Pro Unlimited makes it easy to apply top-down seasonality. One merely marks
the group with the modifier \INDEXES to compute both seasonal and event indexes at
that level.

Although the above examples all use two levels of aggregation, Forecast Pro Unlimited
allows you to define as many levels as you desire. Any combination of top-down and
bottom-up reconciliation can be used.

Forecast Pro Unlimited disaggregates in a series of steps. First the forecasts for a top-
down group are frozen. Then the first-generation component forecasts are adjusted so that
they sum to their parent forecasts. Then those forecasts are frozen and used to adjust their
component forecasts. This process continues until all item-level forecasts have been
adjusted.

Incorporation of Additional
Information
It will often be appropriate to adjust forecasts from Forecast Pro Unlimited judgmentally.
This may be done to include information about orders that have been received, or which
are expected; the effects of promotions; knowledge about external economic trends;
product mix changes, etc. These are particular cases where the information available to
the user is greater than that available to Forecast Pro Unlimited.

In other cases, i.e., when the additional information possessed by the manager is not clear,
the effect of subjective intervention by the user may be counterproductive. Studies have
shown that, although most managers believe that their subjective assessment is superior to
quantitative projection, that is not always the case. As Fildes [1990] has put it, manager
intervention is a “mixed blessing.” We advise the user to be cautious.

In any case, user intervention is time consuming. In those research studies where
subjective intervention provided improved accuracy, the managers typically used
graphical analysis and fairly intensive consideration as their tools. Quick “eyeball”
adjustments are not likely to contribute much to accuracy.

Forecast Pro Unlimited offers a forecast adjustment facility to permit subjective

intervention by the user. Both individual and group level forecasts can be adjusted in a
variety of ways.

Selection of Forecasting Method

Forecast Pro Unlimited offers the user five basic forecasting methods, and a wide range
of variations on these methods. The methods that are included were selected from among
those few univariate forecasting models that are supported by the research, and which can

44 Trading Day Effects

be automated without excessive computational burden. Although true multivariate
methods are not feasible in automatic systems, Forecast Pro Unlimited does support event
modeling which can accommodate promotional schedules and business interruptions.

The five basic methods are simple moving average, exponential smoothing, Croston’s
intermittent demand, discrete distributions and Box-Jenkins (ARIMA). Simple moving
average is included only for use on very short data series, where it is infeasible to fit more
complex models to the data. The Croston’s model is designed for data with numerous
zeros. Discrete distributions (negative binomial and Poisson distributions) are for use on
data whose values are small integers. The other two “methods,” exponential smoothing
and ARIMA, are not actually single methods but, rather, families of methods. The
member methods differ mainly in their accommodation of structural characteristics in the
data like trend, seasonality and random noise. Choosing a method thus involves two
steps: choosing a family and choosing a method from within the family.

Empirical research studies such as the M-Competition have shown that there is no one
single forecasting methodology that is most accurate in all cases. Fildes [1990] has shown
that improvements in accuracy of 20% or more can be obtained by selecting the method
that is most appropriate for a given data set. Thus the task of selecting a method is crucial
for accuracy.

Features such as data length, stability, trending and seasonality will often lead the
experienced forecaster to favor one method over another. These factors may help the
experienced forecaster form a “hunch” about the best method for a particular group of
data. However, as Fildes [1990] has demonstrated, the best and most reliable way to fit
the method to the data is through testing. That is why Forecast Pro Unlimited includes an
expert selection algorithm to automatically select a method for each series and a facility
for testing over a hold-out sample.

Thus there are three approaches to selecting a method. The first is to allow the program’s
expert selection algorithm to choose the models. The second is to allow the program to
select a model after you have decided on a family of models. The third is to make the
selection yourself after hold-out testing over your data (preferably all of it).

Forecast Pro Unlimited’s expert selection is easy to use and works extremely well. The
only disadvantages is that the algorithm is time consuming. Manual model selection
allows you to consider more models during the experimentation stage. Although this step
is time consuming, your forecast production runs will be substantially quicker.

The following sections describe these processes in greater depth. However, they do not
cover the statistical foundations of the forecast methods themselves. They concentrate
instead on implementation as automatic methods, and on implications for use of Forecast
Pro Unlimited. Background material on the five methodologies (simple moving average,
exponential smoothing, Croston’s intermittent demand, discrete distributions and Box-
Jenkins) is presented in the next chapter.

Selection of Forecasting Method 45

Model Selection via Out-of-Sample
Testing
Instead of letting Forecast Pro Unlimited decide which forecast model to use for each
series separately, you can specify it yourself by adding the appropriate modifiers to your
script file. For example, \SIMPLE forces use of simple exponential smoothing, \HOLT
forces use of the Holt model, and \EXSM=LA evokes additive Winters. These three
options might be appropriate for nontrended, trended and seasonal data, respectively.

The modifiers on a particular line of the script file apply to all items in the data file (or
ODBC table) cited on that line. Therefore you should classify your items into different
groups (and data files) with similar properties. In this way you can avoid fitting seasonal
models to nonseasonal data, trended models to nontrended data, etc. Each data file is
cited on a different line of the script, along with the model specification.

At first thought, this procedure seems inferior to that of selecting a model separately for
each item. In fact, however, overall forecasting performance may be markedly improved.
That is because business data is often so irregular that the statistical information in a
single series may not be sufficient to make a reliable choice of model.

After we have classify the items into groups of like items, we must decide on a model that
is best overall. This is done by using out-of-sample testing.

To test a particular model, define a holdout sample on the dialog bar before creating the
forecasts. This directs Forecast Pro Unlimited to withhold the specified number of time
points from the end of the data, and to fit the model to the remaining data, which we call
the “fit set.” The withheld data is called the “check set.”

Forecast Pro Unlimited first forecasts the check set data from the last point of the fit set.
Then it moves to the first point in the check set as a forecast base, and forecasts the
remaining n-1 values. This process continues over all but the last point in the check set.
Forecast errors are computed by subtracting the known true values of the check set from
their forecasts.

By rolling forward in this way, Forecast Pro Unlimited accumulates a total of n one-step-
ahead forecast errors from n different forecast bases, n-1 two-step-ahead errors from n-1
different forecast bases, etc. Since forecast performance can change radically from one
forecast base to another, the rolling forecast errors provide a much better picture of true
out-of-sample performance than a “snapshot” taken at only one forecast base.

The line-level summary statistics of the evaluation are written to the text window, where
you can print it if you wish. The results from each model are summarized, so you can
choose the model that performed best overall. You will probably make repeated runs,
making small changes between each run. You might, for instance, reclassify your
seasonal and nonseasonal series. Finally, of course, you will have the information you
need to select the best methodology for your data. While this approach requires more
initial work than using expert selection, it could pay off in terms of improved accuracy.

46 Trading Day Effects

While this example focuses on different exponential smoothing models, you can also use
the Forecast Pro Unlimited out-of-sample evaluation methodology to choose between
automatic exponential smoothing and automatic Box-Jenkins, or to fine-tune some of the
Box-Jenkins options.

Model Selection via Out-of-Sample Testing 47

Glossary

This glossary contains definitions of the technical terms used in the main body of the text.
A particular definition may involve terms that are defined elsewhere in the glossary.

The ACF consists of the autocorrelations for lags 1, 2, 3, ... N.

ACF Forecast Pro displays the ACF as a correlogram, i.e. a bar chart of
(autocorrelation the autocorrelations arranged by lag.
function)
(AutoRegressive Integrated Moving Average) model. A family of
ARIMA model sophisticated statistical models used by Box and Jenkins to
describe the autocorrelations of a time series data. The symbol
ARIMA(p,d,q) indicates a model involving p autoregressive
terms and q moving average terms, applied to data that have been
differenced d times. The Box-Jenkins technique involves (1)
Identification of a particular ARIMA model to represent historic
data; (2) Estimation of ARIMA model coefficients, (3) Statistical
validation of the model; and (4) Preparation of forecasts.

The correlation of a variable and itself N periods later, and hence

Autocorrelation a measure of predictability.

The forecast base is the time point from which forecasts are
Base prepared.

49
 A model selection criterion proposed by Schwarz [1978]. Within
BIC (Bayes a model family (e.g. exponential smoothing or Box-Jenkins), the
information model that minimizes the BIC is likely to provide the most
criterion) accurate forecasts. Since models with many parameters often fit
the historical data well, but forecast poorly, the BIC balances a
reward for goodness-of-fit with a penalty for model complexity. If
your current model yields the lowest BIC out of the models you
have tested, Forecast Pro marks it with “Best thus far.”

Logarithmic or power transform of the data. Used to reduce or

Box-Cox power eliminate dependence of the local range of a time series on its
transform local mean.

Strictly speaking, the statistical technique developed by Box and

Box-Jenkins Jenkins to fit ARIMA models to time series data. More loosely,
the term refers to the ARIMA models themselves.

A forecast is generally produced along with its upper and lower

Confidence confidence limits. Each confidence limit is associated with a
limits certain percentile. If the upper confidence limit is calculated for
97.5% and the lower for 2.5%, then actual values should fall
above the upper confidence limit 2.5% of the time, and below the
lower confidence limit 2.5% of the time. These are often called
the 95% confidence limits to indicate that the actual value should
fall inside the confidence band 95% of the time. In practice,
confidence limits tend to overstate accuracy. You can set the
confidence limit percentiles in Configure.

The variable you want to forecast. Strictly speaking this term only
Dependent applies to regression modeling, where there are independent
variable variables as well, but it is sometimes convenient to use it for the
variable in univariate models as well.

To difference a time series variable is to replace each value

Differencing (except for the first) by its difference from the previous value.
The seasonal difference replaces each value (except for those in
the first year) by its difference from the value one year previously.

This statistic checks for autocorrelation in the first lag of the

Durbin-Watson residual errors. It should be about 2.0 for a perfect model.
test Forecast Pro computes the Durbin-Watson d-statistic, which is,
strictly speaking, applicable only for regressions that include a
constant intercept term, but do not include lagged dependent
variables.

An exogenous variable is an explanatory variable that can be

Exogenous treated as a time series of ordinary numbers. Practically speaking,
variable independent variable means the same thing.

50 Glossary
 A robust forecasting method that extrapolates smoothed estimates
Exponential of level, trend, and seasonality of a time series.
smoothing
The historic data set used to fit the parameters of a model, and as
Fit set the base of extrapolation for the forecasts.

Standard error of the within-sample forecasts, computed by

Forecast error running the forecast model through the historic data. Used as an
estimate of the one-step forecast error.

Number of periods you wish to forecast.

Forecast horizon
A forecast scenario extends the historic series of independent
Forecast variables into the future. Dynamic regression forecasts are
scenario dependent on the forecast scenario.

The time difference between a time series value and a previous

Lag value from the same series.

Checks for autocorrelation in the first several lags of the residual

Ljung-Box test errors. If the Ljung-Box test is significant for a correlational
model (Box-Jenkins or dynamic Regression) then the model
needs improvement. The test is significant if its probability is >
.99, in which case it is marked with two asterisks in the standard
diagnostic output.

See local mean.

Local level
The average level of a time series in the general neighborhood of
Local mean a given point in time. Sometimes called the local level.

The average rate of increase of a time series in the general

Local trend neighborhood of a given point in time.

Mean Absolute Deviation. This measure of goodness-of-fit is

MAD calculated as the average of the absolute values of the errors. It is
an important statistic in rolling simulation analysis.

Mean Absolute Percentage Error. A statistic used to measure

MAPE within sample goodness-of-fit and out-of-sample forecast
performance. It is calculated as the average of the unsigned
percentage errors.

A forecasting model is an equation, or set of equations, that the

Model forecaster uses to represent and extrapolate features in the data.

Glossary 51
 Model complexity is measured by the number of parameters that
Model must be fitted to the historic data. Overfitting, i.e., using too many
complexity parameters, leads to models that forecast poorly. The BIC can
help to find the model that properly trades off goodness-of-fit in
the historic fitting set, and its model complexity.

Involving more than one variable at a time. Dynamic regression is

Multivariate a multivariate technique.

The difference between a predicted value and a true value in the

Residual error fitting set, i.e. the fitted error.

A robust method is insensitive to moderate deviations from the

Robust underlying statistical assumptions.

A statistic that is used as an indication of model fit. It is

Root mean calculated by taking the square root of the average of the squared
squared error residual errors.
(RMSE)
Periodic patterns of behavior of the series. For instance, retail
Seasonality sales exhibit seasonality of period 12 months. Usually the
forecaster must take seasonality explicitly into account during the
model fitting process.

A process is said to be stochastic when its future cannot be

Stochastic predicted exactly from its past. In a stochastic process, new
uncertainty enters at each point in time.

Involving only one variable at a time. Exponential smoothing and

Univariate Box-Jenkins are univariate techniques.

52 Glossary
Bibliography

J. S. Armstrong [1983] Relative Accuracy of Judgmental and Extrapolative Methods in

Forecasting Annual Earnings, Journal of Forecasting, 2, pp. 437-447.

J. S. Armstrong [1985] Long Range Forecasting From Crystal Ball to Computer (2nd
ed.), New York: Wiley.

J. S. Armstrong [2001] Principles of Forecasting: A Handbook for Researchers and

Practitioners, Norwell MA: Kluwer Academic Publishers.

R. Ashley [1983] On the Usefulness of Macroeconomic Forecasts as Inputs to

Forecasting Models, Journal of Forecasting, 2, pp. 211-223.

R. Ashley [1988] On the Relative Worth of Recent Macroeconomic Forecasts,

International Journal of Forecasting, 4, pp. 363-376.

F. Bass [2004] A New Product Growth Model for Consumer Durables, Management
Science, 50, pp.1825-1832.

B. L. Bowerman and R. T. O'Connell [1979] Time Series and Forecasting: An Applied

Approach, Boston: Duxbury Press.

G. E. P. Box and G. M. Jenkins [1976] Time Series Analysis: Forecasting and Control,
Revised Edition, San Francisco: Holden-Day.

R. G. Brown [1963] Smoothing, Forecasting and Prediction of Discrete Time Series,

Englewood Cliffs, NJ: Prentice-Hall.

D. W. Bunn and A. I. Vassilopoulos [1993] Using Group Seasonal Indices in Multi-

item Short-term Forecasting, International Journal of Forecasting, 9, 4, pp. 517-526.

Caceci and Cacheris [1984] Fitting Curves to Data, Byte Magazine, May, 1984.
53
C. Chatfield [1978] The Holt-Winters Forecasting Procedure, Applied Statistics, 27, pp.
264-279.

C. Chatfield and M. Yar [1988] Holt-Winters Forecasting: Some Practical Issues, The
Statistician, 37, pp. 129-140.

C. Chatfield and M. Yar [1991] Prediction Intervals for the Holt-Winters Forecasting
Procedure, International Journal of Forecasting, 6, 1, pp. 127-137.

G. C. Chow [1960] Tests of Equality between Subsets of Coefficients in Two Linear

Regressions, Econometrica, 37, pp. 591-605.

D. Cochrane and G. H. Orcutt [1949] Application of Least Squares Regression to

Relationships Containing Autocorrelated Error Terms, Journal of the American
Statistical Association, 44, pp. 32-61.

J. D. Croston [1972] Forecasting and Stock Control for Intermittent Demands,

Operational Research Quarterly, 23(3), pp. 289-303.

E. B. Dagum [1982] Revisions of Time Varying Seasonal Filters, Journal of

Forecasting, 1, 2, pp. 173-187.

D. A. Dickey, W. R. Bell, and R. B. Miller [1986] Unit Roots in Time Series Models:
Tests and Implications, The American Statistician, 40, pp. 12-26.

D. A. Dickey and W. A. Fuller [1979] Distribution of the Estimators for Autoregressive

Time Series with a Unit Root, Journal of the American Statistics Society, Vol. 74, pp.
427-431.

D. A. Dickey and W. A. Fuller [1981] The Likelihood Ratio Statistics for

Autoregressive Time Series with a Unit Root, Econometrica, 49, pp. 1057-1072.

R. F. Engle [1984] Wold, Likelihood Ratio, and Lagrange Multiplier Tests in

Econometrics, Handbook of Econometrics Vol. II, Eds. Z. Griliches and M.D. Intriligator,
North-Holland Publishers BV.

R. Fildes [1979] Quantitative forecasting -- the state of the art: extrapolative methods,
Journal of the Operational Research Society, 30, pp. 691-710.

R. Fildes and S. Howell [1979] On Selecting a Forecasting Model, Forecasting, Studies

in the Management Sciences, Vol. 12, S. Makridakis and S. C. Wheelwright (Eds.),
Amsterdam: North-Holland.

A. D. Flowers and M. H. Finegan [1988] A Comparison of Microcomputer Forecasting

Packages, ORSA/TIMS.

W. A. Fuller [1976] Introduction to Statistical Time Series, New York: Wiley.

54 Bibliography
E. S. Gardner, Jr. [1983] Automatic Monitoring of Forecast Errors, Journal of
Forecasting, 2, 1, pp. 1-21.

E. S. Gardner, Jr. [1985] Exponential Smoothing: The State of the Art, Journal of
Forecasting, 4, 1, pp. 1-38.

R. L. Goodrich [1989] Applied Statistical Forecasting, Belmont, MA: Business Forecast

Systems, Inc.

C. W. J. Granger [1989] Combining Forecasts—l Twenty Years Later, Journal of

Forecasting, 8, pp. 167-173.

C. W. J. Granger and G. McCollister [1978] Comparison of Forecasts of Selected

Series by Adaptive, Box-Jenkins, and State Space Methods, ORSA/TIMS Meeting, Los
Angeles, Summer, 1978.

C. W. J. Granger and P. Newbold [1977] Forecasting Economic Time Series, New

York: Academic Press.

G. W. C. Granger and R. Ramanathan [1984] Improved Methods of Combining

Forecasts, Journal of Forecasting, 3, pp. 197-204.

A. C. Harvey [1981] Time Series Models, Oxford, England: Phillip Allen.

G. Hill and R. Fildes [1984] The Accuracy of Extrapolation Methods; An Automatic

Box-Jenkins Package Sift, Journal of Forecasting, 3, pp. 319-323.

C. C. Holt [1957] Forecasting trends and seasonals by exponentially weighted moving

averages, O .N. R. Memorandum, No. 52, Carnegie Institute of Technology (now
Carnegie-Mellon University).

K. Kahn [2006] New Product Forecasting: An Applied Approach, Armonk NY: M.E.
Sharpe.

A. B. Koehler and B. L. Bowerman [1989] Prediction Intervals for ARIMA Models,

Ninth International Symposium on Forecasting, Vancouver.

A. B. Koehler and E. S. Murphree [1986] A Comparison of the AIC and BIC on

Empirical Data, Sixth International Symposium on Forecasting, Paris.

R. Lewandowski [1982] Sales Forecasting by FORSYS, Journal of Forecasting, 1, pp.

205-214.

J. Ledolter and B. Abraham [1984], Some Comments on the Initialization of

Exponential Smoothing, Journal of Forecasting, 3, pp. 79-84.

Bibliography 55
G. Libert [1984] The M-Competition with a Fully Automatic Box-Jenkins Procedure,
Journal of Forecasting, 3, pp. 325-328.

G.L. Lilien, A. Rangaswamy and C. Van den Bulte [1999] Diffusion Models:
Managerial Applications and Software, Institute for the Study of Business Markets, 7,
University Park, PA.

G. M. Ljung and G. E. P. Box [1978] On a Measure of Lack of Fit in Time Series

Models, Biometrika, 65, pp. 297-303.

E. J. Lusk and J. S. Neves [1984] A Comparative ARIMA Analysis of the 111 Series of
the Makridakis Competition, Journal of Forecasting, 3, pp. 329-332.

S. Makridakis and M. Hibon [1979] Accuracy of Forecasting: An Empirical

Investigation (with discussion), Journal of the Royal Statistical Society, A 142, Part 2,
97-145. Reprinted in Makridakis et al. [1982].

S. Makridakis and M. Hibon [1989] Exponential Smoothing: The Effect of Initial

Values and Loss Functions on Post-sample Forecasting Accuracy, Ninth International
Symposium on Forecasting, Vancouver.

S. Makridakis and M. Hibon [2000] The M3-Competition: Results, Conclusions and

Implications, International Journal of Forecasting, 16, 4, pp. 451-476.

S. Makridakis and S. C. Wheelwright [1979] Interactive Forecasting, Second Edition,

San Francisco: Holden-Day.

S. Makridakis, S. C. Wheelwright and R.J. Hyndman [1998] Forecasting Methods

and Applications, Third Edition, New York: Wiley.

S. Makridakis et al. [1982] The Accuracy of Extrapolation (Time Series) Methods:

Results of a Forecasting Competition, Journal of Forecasting, 2, pp. 111-153.

S. Makridakis et al. [1984] The Forecasting Accuracy of Major Time Series Methods,
Chichister: Wiley.

E. McKenzie [1986] Error Analysis for Winters' Additive Seasonal Forecasting System,
International Journal of Forecasting, 2, pp. 373-382.

D. C. Montgomery and L. A. Johnson [1976] Forecasting and Time Series Analysis,

New York: McGraw-Hill.

J. S. Morrison [1995] Life-Cycle Approach to New Product Forecasting, The Journal of

Business Forecasting, 14, 2, pp. 3-5.

56 Bibliography
H. L. Nelson and C. W. J. Granger [1979] Experience with Using the Box-Cox
Transformation when Forecasting Economic Time Series, Journal of Econometrics, 10,
pp. 57-69.

C. R. Nelson and H. Kang [1981] Spurious Periodicity in Inappropriately Detrended

Time Series, Econometrica, 49, pp. 741-751.

C. R. Nelson and H. Kang [1983] Pitfalls in the Use of Time as an Explanatory Variable
in Regression, Journal of Business and Economic Statistics, 2, pp. 73-82.

P. Newbold [1983] ARIMA Model Building and the Time Series Analysis Approach to
Forecasting, Journal of Forecasting, 2, 1, pp. 23-35.

P. Newbold and T. Bos [1990] Introductory Business Forecasting, Cincinnati: South-

Western.

K. Ord and R. Fildes [2013] Principles of Business Forecasting, Cengage.

A. Pankratz and U. Dudley [1987] Forecasts of Power-transformed Series, Journal of

Forecasting, 6, pp. 239-248.

A. Pankratz [1991] Forecasting with Dynamic Regression Models, New York: Wiley.

R. Ramanathan [1992] Introductory Econometrics with Applications, Second Edition,

Fort Worth: Harcourt Brace Jovanovich College Publishers.

G. Schwarz [1978] Estimating the Dimension of a Model, Ann. Statist. 6, 2, pp. 461-464.

J. Shiskin, A. Young, and J. Musgrave [1967] The X-11 Variant of the Census Method
II Seasonal Adjustment Program, Technical Paper No. 15, Bureau of the Census.

E. Stellwagen [2012] Exponential Smoothing: The Workhorse of Business Forecasting,

Foresight: the International Journal of Applied Forecasting. 27, pp. 23-28.

P. A. Texter and J. K. Ord [1989] Forecasting Using Automatic Identification

Procedures, International Journal of Forecasting, 5, 2, pp. 209-215.

H. Theil [1971] Principles of Econometrics, New York: Wiley.

N. T. Thomopoulos [1980] Applied Forecasting Methods, Englewood Cliffs, NJ.:

Prentice-Hall.

S. C. Wheelwright and S. Makridakis [1985] Forecasting Methods for Management,

New York: Wiley.

Bibliography 57
T. R. Willemain, C. N. Smart, J. H. Shockor and P. A. DeSautels [1994] Forecasting
Intermittent Demand in Manufacturing: a Comparative Evaluation of Croston's Method,
International Journal of Forecasting, 10, 4, pp.529-538.

G. T. Wilson [1979] Some Efficient Computational Procedures for High Order ARMA
Models, J. Statist. Comput. Simul., 8, pp. 301-309.

P. R. Winters [1960] Forecasting Sales by Exponentially Weighted Moving Averages,

Management Sci., 6, pp. 324.

H. O. Wold [1938] A Study in the Analysis of Stationary Time Series, Uppsala: Almquist
and Wicksell.

M. Yar and C. Chatfield [1991] Prediction Intervals for Multiplicative Holt-Winters,

International Journal of Forecasting, 7, 1, pp. 31-37.

G. U. Yule [1927] On a Method of Investigating Periodicities in Disturbed Series, with

Special Reference to Wolfer's Sunspot Numbers, Phil. Trans., A226, pp. 267.

58 Bibliography
Index

Adjusted R-square, 34 diagnostics, 30

ARIMA dynamics, 31
processes, 25 variable specification, 31
Autocorrelation, 49 Event effects, 4
function, 22, 27 Event indexes, 4
Autoregressive Exogenous variable, 50
order, 27 Exponential smoothing
processes, 25 description, 3
terms, 23 Holt, 6, 13
Backward shift operator, 25 simple, 5
Bayes information criterion, 50 Winters, 6, 14
Bayesian information criterion, 35 Fit set, 46, 51
BIC, 8, 35 Forecast error, 35
Box-Cox power transform, 50 GMRAE, 36
Box-Cox power transforms, 36 Holdout sample, 46
Box-Jenkins, 20 Judgmental forecasting, 44
identification, 27 Lagged dependent variables, 29
initialization, 21 Level, 4
mixed processes, 28 Ljung-Box, 35
parameter estimation, 21 Ljung-Box test, 51
seasonal models, 26 MAD, 36, 51
Check set, 46 MAD/Mean Ratio, 36
Cochrane-Orcutt, 30, 31 MAPE, 36, 51
Confidence limits, 8, 50 M-Competition, 45
Croston’s model, 18 Mean, 34
Curve Fitting, 19 Model complexity, 52
Dependent variable, 50 Model selection, 8
Differencing, 24, 27, 50 Moving average
Durbin-Watson, 34 order, 28
Durbin-Watson test, 50 processes, 25
Dynamic regression, 28 terms, 23

59
Multiple-level forecasting, 43
Multivariate, 52
NA-constant level model, 15
Negative binomial distribution, 17
Number of parameters, 34
Outlier detection, 39
Parameter optimization, 8
Partial autocorrelation function, 27
Poisson distribution, 16
Random events, 4
Random shocks. See Random events
Residual error, 52
RMSE, 36
R-square, 34
Safety stocks, 37
Sample size, 34
Seasonal effects, 4
Seasonal indexes, 4
Seasonal simplification, 15
Seasonality
additive, 6
multiplicative, 6
Selecting a method, 44
SMAPE, 36
Standard deviation, 34
Standard forecast error, 35
Stationarity, 21, 24
Stochastic, 52
Time series
classification of, 42
hierarchies, 43
rogue, 42
types A,B,C,, 42
Trading day effects, 39
Trend, 4
Univariate, 52

60 Index