You are on page 1of 22

International Journal of Forecasting 31 (2015) 11051126

Contents lists available at ScienceDirect

International Journal of Forecasting


journal homepage: www.elsevier.com/locate/ijforecast

Review

Forecasting in telecommunications and ICTA review


Nigel Meade a, , Towhidul Islam b,1
a

Imperial College Business School, Imperial College London, Exhibition Road, London SW7 2AZ, UK

Department of Marketing and Consumer Studies, University of Guelph, Guelph, Ontario, N1G 2W1, Canada

article

info

Keywords:
Diffusion models
Time series
Technology forecasting
Mobile telephony
Internet

abstract
Given the length of time that has elapsed since the IJF Special Issue on Telecommunications
Forecasting in 2002 and our reliance on information and communications technology (ICT),
it is now appropriate to review the flow of benefits from forecasting to ICT and from
ICT to forecasting. The importance of ICT is demonstrated by its accounting for 8.2% of
the value added and for over 20% of employment in the OECD countries. The literature
reviewed is categorised by both the ICT area of application and the modelling approach.
The ICT application areas are: mobile telephony, internet usage or provision, and other ICT
related products and services. The main modelling and forecasting approaches are diffusion
modelling and forecasting, time series forecasting and technological forecasting, and this
review devotes a section to each of these approaches. Most of the research activity in the
field (measured by numbers of papers) has occurred in the modelling of diffusion in ICT,
particularly mobile telephony, producing beneficial cross-fertilisation between forecasting
and ICT applications; examples are multi-generational modelling and choice modelling.
Although call centre manpower planning has led to innovative forecasting models, other
analyses of clusters of ICT time series data sets have been less innovative. Technological
forecasting papers tend to be exercises based on expert opinion.
2014 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

Contents
1.
2.

Introduction..........................................................................................................................................................................................
Diffusion models ..................................................................................................................................................................................
2.1.
Single country diffusion analyses ...........................................................................................................................................
2.1.1.
Example analyses and comparisons of forecasting accuracy.................................................................................
2.1.2.
Analysis of the generality of the conclusions from the Frank analysis................................................................
2.1.3.
Definition of market potential .................................................................................................................................
2.1.4.
Summary of similar studies .....................................................................................................................................
2.1.5.
Other single-country analyses related to ICT diffusion..........................................................................................
2.2.
Multi-country analyses of ICT diffusion .................................................................................................................................
2.2.1.
Multi-country analyses with short time series.......................................................................................................

Corresponding author. Tel.: +44 0 20 7594 9116; fax: +44 0 20 7823 7685.
E-mail addresses: n.meade@imperial.ac.uk (N. Meade), islam@uoguelph.ca (T. Islam).

1 Tel.: +1 519 824 4120x53835; fax: +1 519 823 1964.


http://dx.doi.org/10.1016/j.ijforecast.2014.09.003
0169-2070/ 2014 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

1106
1107
1107
1108
1110
1110
1111
1112
1113
1113

1106

3.

4.

5.

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

2.2.2.
Multi-country diffusion analyses using covariates ................................................................................................
2.2.3.
Other multi-national analyses related to ICT diffusion ..........................................................................................
2.3.
Choice modelling and other extensions .................................................................................................................................
2.4.
Summary: the flow of benefits between ICT and forecasting methodologies for diffusion ...............................................
Time series............................................................................................................................................................................................
3.1.
Forecasting for call centres......................................................................................................................................................
3.2.
ICT time series..........................................................................................................................................................................
3.2.1.
Univariate telecommunications time series ...........................................................................................................
3.2.2.
Internet time series: forecasting usage and provision ...........................................................................................
3.2.3.
Forecasting for other ICT products ..........................................................................................................................
3.3.
Summary: the flow of benefits between ICT and time series forecasting methodology ....................................................
Technological forecasting ....................................................................................................................................................................
4.1.
Mobile telephony .....................................................................................................................................................................
4.2.
The internet..............................................................................................................................................................................
4.3.
Summary: the flow of benefits between ICT and technological forecasting methodology ................................................
Summary, conclusions and suggestions for further research ...........................................................................................................
Acknowledgments ...............................................................................................................................................................................
Appendix A.
Supplementary data ..................................................................................................................................................
References.............................................................................................................................................................................................

1. Introduction
In this paper, we review applications of modelling
and forecasting in information and communications technology (ICT). We understand the term ICT as including
telecommunications and referring to both hardware and
software. To put the importance of ICT in context, an analysis of 28 OECD countries showed that, in 2008, the ICT
sector accounted for between 3.7% (Switzerland) and 13.9%
(Finland) of the value added in manufacturing and business
services (the average proportion was 8.2%). For the same
countries, ICT-using occupations (including specialists, advanced and basic users) account for over 20% of total employment on average, ranging from 10.9% (Turkey) to 35.3%
(Luxembourg) (see OECD, 2011).
It is over a decade since the International Journal of Forecasting published a Special Issue on Telecommunications
Forecasting in 2002 (see Fildes, 2002, 2003). Given the time
that has elapsed since then, and the almost total dependence of life in developed economies on ICT, we feel that
it is now appropriate to review the flow of benefits from
forecasting to ICT and from ICT to forecasting.
Our preliminary literature search found that studies focussing on hardware (television, cellular phones, computer
and network equipment) predominated. ICT software applications, such as videoconferencing, distance learning
and business to business communications (B2B), do not
seem to have not generated forecasting studies. Within the
constraints of the published forecasting literature, our aim
is to make this review as inclusive as possible. The papers reviewed here are included primarily because they
make some contribution to modelling and forecasting in
an ICT context. To assess the levels of activity in different
research areas, these papers can be categorised by both the
ICT area of application and the modelling approach. The ICT
application areas can be classified broadly into three categories: mobile telephony, some aspect of internet usage
or provision, and other ICT products, such as PCs or television. The topic that has generated the most interest is mobile telephony, representing 27% of the papers reviewed.
Modelling and forecasting in relation to the internet represented 19%, and other ICT products were discussed in

1113
1115
1115
1116
1117
1117
1118
1118
1120
1121
1121
1121
1121
1121
1122
1122
1123
1123
1123

39% of papers. Forecasting theory and other related issues


were the theme of the remaining 15%. Three main modelling and forecasting approaches have been identified in
our review of forecasting in ICT and telecommunications.
These are diffusion modelling and forecasting, time series
forecasting, and technological forecasting. Considering the
more recent papers (from 2000 onwards), we find that 55%
can be classified as diffusion modelling, 34% can be classified as time series modelling, and 11% focus on technological forecasting. The review considers each of these
approaches in turn: Section 2 discusses diffusion modelling, Section 3 covers time series modelling and Section 4
looks at technological forecasting. Within the discussion
of each approach, we discuss the ICT categories identified,
and also summarise the flow of benefits between forecasting and the relevant ICT applications for each approach.
We have not attempted to use a common framework to
analyse the three approaches described in Sections 24.
The problems addressed differ between sections, and the
degree of homogeneity among the applications of each
approach differs considerably. For example, the studies
modelling diffusion in a single country are comparatively
homogeneous, and some general points about the modelling approach can be, and are, made. In contrast, the studies in technological forecasting have few issues in common,
and no conclusions about the usefulness or appropriateness of a technique can be drawn.
The consequence of conducting a review with the selfimposed broad remit of forecasting in ICT is that a wide
range of analyses is included. Several factors determine
the supply of academic analyses, of which we will identify two. Firstly, where quantitative analysis is crucial to
the understanding of a topic, the availability of data attracts academic analysis. In the absence of easily available data, academic analyses are harder to execute and less
common. The availability of data has facilitated the supply of studies in Sections 2 and 3. In contrast, there are areas of ICT activity in which forecasting studies are either
rare or absent. For example, the probable effects of developments in ICT on the relative popularity and future usage of the different modes of delivery of music, films and

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

books to consumers have received little attention. A second important factor is the filter on the supply of studies,
namely the threshold for publication. The papers considered in this review come from a large number of different academic fields of study, including forecasting, time
series analysis, operations research, operations management, computer science, telecommunications, and information technology, to name the more important examples.
The threshold for the publication of a given study varies between the journals in these different academic fields. Consider, for example, a study which models and forecasts the
diffusion of mobile telephony in Ruritania. The height of
the threshold for publication in a forecasting or time series journal will be influenced by the level of innovation in
its modelling and forecast evaluation. In a telecommunications journal, the height of the threshold will be influenced
more by the strategic implications of the forecasts for the
industry. As we will see in this review, this variety of publication thresholds inevitably means that some studies that
have satisfied one set of criteria will be considered deficient according to a different set of criteria.
The distinction between ICT application areas such as
internet and mobile telephony will be increasingly difficult to make in future reviews, as the convergence of these
technologies progresses. Internet usage is now the main
revenue source for network operators in developed markets, rather than voice calls or messaging. Furthermore,
global internet access via mobile devices was forecast to
exceed the access via fixed broadband by 2014 (Price
Waterhouse Coopers http://www.pwc.com/gx/en/globalentertainment-media-outlook/segment-insights/internetaccess.jhtml). It is not very risky to forecast that ICT will
continue to generate interesting modelling and forecasting
problems for some time to come.
2. Diffusion models
Diffusion is the process by which an innovation is
adopted by a population. Relevant examples of innovations
are, historically, fixed line telephony, or, currently, mobile
telephony. The diffusion process is characterised initially
by the introduction of the innovation, followed by a slow
growth in adoption as awareness increases. The growth accelerates to a point where adoptions per period peak, then
adoptions decelerate as the population becomes saturated
with the innovation. The diffusion models used for the
adoption of an innovation in a single population are necessarily a simplified representation of this process, because
the data set available is typically a short time series, probably containing fewer than 20 observations. In order to gain
further insights into the determinants of the diffusion process, multi-country studies use other variables to explain
the differences in diffusion rates between countries.
Meade and Islam (2006) reviewed studies on the
modelling and forecasting of the diffusion of innovations. They cover a large body of literature which looks
at many different mathematical formulations of an Sshaped diffusion, where the cumulative adoption of an
innovation moves from launch to a saturated market.
We do not intend to duplicate this work, but will use

1107

it for a general reference, and specifically for its appendix, which lists fifteen S-shaped growth curve equations. Thus, if no formulation for a diffusion model is
given, one should refer to the appendix mentioned. In
addition to the diffusion of an innovation, researchers
also consider technological substitution, where an existing
technology is replaced by a newer one, such as the replacement of rail steam locomotion with diesel locomotion, see
for example Blackman (1972), Fisher and Pry (1971) and
Sharif and Kabir (1976). One issue that will be discussed
here is the question of whether mobile telephony (or mobile internet) is a substitute for fixed line telephony (or
fixed broadband). Researchers also consider multiple generations of the same (or similar) technology; this issue is
of particular interest in telecommunications and ICT. Examples of this include personal computers (PC, XT, AT, 386,
486, Pentium, etc.) and mobile telephones (analog, digital,
2G, 3G, 4G), both of which have been through several generations of technology. This is an area of research in which
ICT applications (in computing and telephony) have stimulated novel developments in modelling. Norton and Bass
(1987) extended the use of the Bass diffusion model to successive generations of a technology. Restrictions on the parameters of the multiple generation model were relaxed
successively by Islam and Meade (1997) and Mahajan and
Muller (1996).
We will discuss the modelling of the diffusion of ICT
products and services in the following three sub-sections.
The first deals with single country diffusions, the second
deals with multi-country analyses, and the third looks at
modelling market share and choice models.
2.1. Single country diffusion analyses
Several authors have used a range of models for the
diffusion of fixed line telephony, with examples ranging
from Bewley and Fiebig (1988), Chaddha and Chitgopekar
(1971), Lee and Lu (1987) and Lee, Lu, and Horng (1992)
to Islam and Meade (1996). Similarly, and more recently,
authors have compared subsets of models on data sets
relating to the diffusion of mobile telephony and other
ICT in particular countries. The evolution of a diffusion
process over time can be regarded as either a cumulative
process, represented by an S-shaped curve, or a noncumulative process, represented by a bell-shaped curve.
In the context of ICT, different actors are concerned with
different aspects of forecast diffusion. Service providers
are concerned with satisfying the cumulative number of
adopters, whereas handset providers are concerned with
satisfying new adopters (the non-cumulative number of
adopters per time period). Diffusion models are most
appropriate for using with adoption data. For example, if
one is modelling the diffusion of television in the United
Kingdom, the number of households with a television
license describes the adoption process well (since each
household is legally obliged to buy a license if it has
at least one television). In contrast, television sales data
are less useful, as they confound the adoption process
with replacement sales and purchases of multiple sets per
household. If the data used contain extra components, such

1108

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

as multiple sales per adopter, the components need to be


included in the modelling process.
In this section, since there are many studies with very
similar structures and objectives, we begin by looking at
an example analysis from the literature. In the first subsection, we describe the steps in an appropriate modelling methodology, covering both the way in which the
diffusion is modelled and the structure of the error process. Secondly, we study the importance of the estimation
procedure used in conjunction with the diffusion models.
Thirdly, we discuss the implications of the inclusion of
extra components in the data set being modelled, in addition to the adoption process. Fourthly, we review and summarise the range of studies on single country ICT/telecoms
diffusion in the context of the steps identified in the first
subsection.
2.1.1. Example analyses and comparisons of forecasting
accuracy
In this modelling and forecasting exercise, there are a
series of choices to be made. The choices of the diffusion
model and parameter estimation method are among the
more important choices. An early study is that by Frank
(2004), who models and forecasts the diffusion of wireless subscribers (GSM) in Finland. This paper has been cited
widely and its approach has been replicated broadly by researchers considering various different national data sets.
For this reason, we use this paper as a case study, to demonstrate the choices made and the consequences of these
choices. The variable being modelled and forecast is the
number of mobile telephone subscriptions per 100 people,
and the data are taken from the ITU and Statistics Finland;
they are shown in Fig. 1. The data are not necessarily identical to those used by Frank. Following Frank, the estimation period is from 1980 to 1997; forecasts are evaluated
from 1998 to 2004. This limit is chosen because there was
an upward surge in subscriptions in Finland after 2004, due
to the launch of 3G mobile technology.
The model chosen by Frank is the logistic. He uses the
cumulative proportion of the Finnish population subscribing to a mobile in year t as the dependent variable. This
proportion is calculated as the number of subscriptions
divided by the size of the population; the assumed identity between subscribers and subscriptions is discussed in
Section 2.1.3. We denote the dependent variable as F (t ),
where
F (t ) =

(1 + exp ( (a + bt )))

(1)

and m, a and b are to be estimated from the data. The parameter m is called the market potential, the maximum
proportion of the population who will become subscribers.
Frank effectively assumes that no more than 100% of the
population will subscribe (i.e., m 1). This restriction will
be discussed further in Section 2.1.3. Frank chooses to estimate these parameters, where there are Yt adopters at time
t, using non-linear least squares, in conjunction with the
assumption that
Yt = F (t ) + t .

(2)

We label this approach as Method 1. Frank introduced


some covariates into the model by parameterising b:
b = b0 + b1 GDP t .

(3)

The idea is that a more buoyant economy, meaning a higher


GDP, will lead to a higher rate of diffusion. Other variables
were tried (the proportion of the population with fixed
line telephony, and a dummy variable indicating the switch
from analog to digital) but were not significant. We denote
this approach Method 1a.
Method 2, which was not used by Frank, considers
new subscriptions per period as the random variable to be
modelled, rather than cumulative subscriptions; that is:
yt = f (t ) + t ,

(4)

where yt = Yt Yt 1 and f (t ) = F (t ) F (t 1). The idea


here is that Eq. (4) represents a more credible stochastic
model, since t is likely to be closer to the assumption of
independent, identically distributed errors than the t that
is implicit in the minimisation of sums of squares. It has
long been known that t is prone to autocorrelation (see
Mar-Molinero, 1980).
Method 3, which was not used by Frank either, is a
straightforward application of the Bass (1969) model using
the coefficient of innovation, p, and the coefficient of imitation, q, to describe the sub-processes within the adoption
process:
probability of adoption at time t , Pt

Y(t 1)
x (t ) ;
= p+q
m

(5)

The expected number of adoptions at time t is (m Yt 1 )


Pt . The function x(t ), which is not used here, reflects the
effects of marketing variables; this development is due to
Bass, Krishnan, and Jain (1994). Here, the parameters are
again estimated using non-linear least squares. This approach is similar to Method 2, as it focuses on new subscriptions per period. The method is included because it is
a discrete representation of the logistic.
To summarise the choices that arise when modelling
diffusion: the first choice is the functional form of the
model of the cumulative proportion of adopters, F (t ); the
second choice is between F (t ) and f (t ) as the dependent
variable. As was discussed in Section 2.1, this choice will
depend on the decision makers role in the market. Frank
chose Method 1 (and Method 1a), where F (t ) is defined
in Eq. (1) and the dependent variable is defined in Eq. (2).
Using the Finnish data, described above, we use Methods
1, 1a, 2 and 3 to estimate the model parameters. For each
method, we compute the root mean square errors (rmse)
in cumulative proportions of adopters (compatible with
Methods 1 and 1a) and the rmse in new adopters per period
(compatible with Methods 2 and 3). This facilitates comparisons of the goodness of fit and forecasting accuracy.
Parameter estimation in the model is typically achieved
by non-linear least squares regression. The constraint, m
1, was applied because the solution was unstable without
it. In the absence of this constraint, for Method 1, the market potential m increased without a bound and a and b decreased; that is, the in-sample fit improved at the expense

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

1109

Fig. 1. Cumulative mobile subscriptions in Finland. The estimation region is 19801997, and the forecast region is 19982004. The estimation is via Method
1, using a cumulative logistic model.

Table 1
A summary of the different estimation methods for the diffusion of mobiles in Finland.
Method
1
Cumulative
logistic
Fitted region 19801997

m
a/a/a/p
b/b0 /b/q
b1

1.00

7.83
0.44

1a
Cumulative logistic
with GDP covariate

2
Non-cumulative
logistic

0.74

1.00

4.85
0.046

10.26

3
Bass
model

0.58

1.00
0.00
0.56

0.003

Cumulative
forecast
19982004

rmse (in)
rmse
mape

1.32
2.19
2.70

0.20
12.57
11.72

2.87
4.16
5.15

5.75
21.72
26.11

Non-cumulative
forecast
19982004

rmse (in)
rmse
mape

0.89
1.45
14.60

0.20
3.64
53.36

0.44
2.58
27.48

0.96
3.21
36.60

The parameter estimates are given in the top half of the table. The in-sample fit and the out-of-sample forecasting accuracy
are shown in the bottom half of the table, measured in terms of both cumulative subscriptions and subscriptions per annum.
The objective used by each method is in bold.

of out-of sample accuracy. For pragmatic estimation reasons, this constraint is applied to all methods to ensure
consistency. We discuss the interpretation of the value of
m in Section 2.1.3.
All four methods are used to estimate the relevant
model parameters, and the results are summarised in Table 1. We report forecasting accuracies using the rmse and
the mean absolute percentage error (mape), both in cumulative terms, consistent with Methods 1 and 1a, and in
terms of annual subscriptions, consistent with Methods 2
and 3. We note firstly that Method 1 produces very accurate forecasts and has the lowest error measures under
both the cumulative subscriptions and subscriptions per
annum criteria. Frank produced forecasts using Method 1a.
In our analysis, the in-sample fit is almost perfect for the
cumulative subscriptions, but the forecasting is strikingly
poor for both cumulative and annual subscriptions. Here, it
appears that the use of GDP as an explanation for the fluctuations in cumulative subscriptions has over-fitted the
data at the expense of estimating the long-term trend captured by Method 1. Method 2 over-estimates the cumulative subscriptions, leading to a poorer forecasting accuracy
than for Method 1. The Bass model is estimated in Model

3, giving the coefficient of innovation, p, as zero, which indicates that the Bass model has been reduced to a logistic.
The Bass model produces the least accurate forecasts, while
Method 1, the cumulative logistic, produces the most accurate forecasts, regardless of whether the accuracy is considered for cumulative or non-cumulative data. The data,
fitted and forecast cumulative subscriptions for Finland,
are shown in Fig. 1.
Franks out-of-sample forecast (using GDP as a covariate
and assuming 2.5% growth per annum) is accompanied
by a prediction interval. Unfortunately, the interval seems
to relate only to uncertainty in parameter estimation,
ignoring the uncertainty due to the error process. Thus,
the interval is optimistically narrow: for 2004, the forecast
is 90.6% 1.2%. In Fig. 1, we show a 95% prediction
interval using the approach described by Meade and Islam
(1995); the forecast is 92.6% and the prediction interval is
(89.0%95.5%). In 2004, the actual cumulative proportion
of the population with mobile subscriptions was 95.4%.
In our analysis of the Finnish data, our results were similar but not identical to Franks, due to the use of different software and slightly different data. Our most accurate
forecasts were obtained using the cumulative logistic without explanatory variables. A side benefit of this is that the

1110

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

extrapolation of the cumulative logistic does not require


the forecasting of GDP growth. We also demonstrate that,
for forecasts over long horizons, prediction intervals that
reflect the uncertainty due to both the error process and
parameter estimation add valuable information.

Minimise

2
i

i =1

Minimise

(t + 1 i) i2

i =1

2.1.2. Analysis of the generality of the conclusions from the


Frank analysis
In the previous section, for the Finnish data on mobile
diffusion, we found that the most accurate forecasts were
generated by Method 1, fitting a logistic to the cumulative
adoption data. The next most accurate method was Method
2, fitting the non-cumulative logistic to the annual adoptions. The Bass model, Method 3, also fitted to annual adoptions, was the least accurate. To test the generality of these
findings, we carried out a comparison using the ITU data
for other countries. In order to get well behaved data, we
applied two filters: the country must have at least 10 nonzero observations up to and including 1997, and the country must have a penetration of at least 10% by 1997. There
are 25 countries which meet these criteria. The Finland
analysis was therefore repeated using the data from these
countries; the data available up to and including 1997 were
used for model estimation, and the forecasting accuracy
was measured using the data from 1998 to 2004. The results are summarised in Table 2. As was done for Finland,
the forecasting accuracy for each country in the data set is
measured by the rmse and mape for both cumulative and
annual subscriptions. The relative forecasting accuracy of
each method is examined by ranking the methods according to these four error measures. The cumulative logistic
has the lowest average rank under each of the four criteria.
The non-cumulative logistic is second most accurate according to three criteria and of equal accuracy with the
cumulative logistic according to the mape of annual subscriptions. The Bass model is least accurate according to all
four criteria. Depending on the accuracy criterion, the cumulative logistic was most accurate for 1215 countries,
the non-cumulative logistic was most accurate for 89
countries, and the Bass model was most accurate for 24
countries. In terms of the median mape for the cumulative
data, there is little difference between the methods, but the
Bass again performs poorly for the more volatile median
mape for annual subscriptions (due to the smaller divisors).
In summary, this analysis suggests that the methodology employed by Frank in fitting the cumulative logistic
using Eq. (2) is more likely to produce more accurate forecasts than the non-cumulative logistic or the Bass model
fitted using Eq. (4) for mobile diffusion data.
If we look in more detail at the objective function asso-

ciated with Eq. (2), Minimise Yt F (t )

, this is equiva-

2
lent to Minimise t . If we express these deviations of the
cumulative subscription data from the cumulative logistic,
t , in terms of the deviations of the annual subscriptions
from the non-cumulative logistic, t , we find the following.

Minimise t

+2

t 1
t

(t + 1 k) j k .

j=1 k=j+1

It is clear that minimising the cumulative errors is similar


to minimising the weighted sum of the annual, period to
period, errors, with the highest weighting on the earliest
errors and the lowest weighting on the deviation between
the fitted non-cumulative model and the most recent
observation. This strong anchoring of the fitted (bellshaped) curve to the earlier data seems to be a successful
strategy for the mobile diffusion data.
2.1.3. Definition of market potential
Market potential, m, is defined in Section 2.1.1 as the
proportion of a population who will (eventually) become
subscribers. Being defined as a proportion, m is clearly
bounded above by unity. In these studies, the problem
arises that, typically, the data available are the numbers
of subscriptions rather than of subscribers. As the mobile
market has developed, the average number of subscriptions per subscriber has reached two or three in some
European countries. Looking at the diffusion of 2G mobiles in Greece, Michalakelis, Varoutas, and Sphicopoulos
(2008) found that all of the models considered estimated
the market potential to be in excess of 100%, which is not
surprising, as the later data points exceed 100%. As more
people take out multiple subscriptions, the upper bound
on market potential begins to exceed 100%; for example,
in 2009 Q1 in Finland, penetration was 141.2% (subscriptions/person).
The application of an adoption model to subscription
data worked satisfactorily when most adopters took out
only one subscription. However, post-2007, the average
number of subscriptions per person exceeded unity in
many countries. This means that an upper bound on the
market potential, m, can no longer be identified. Since the
estimated market potential is a crucial determinant of long
term forecasts, and particularly of the forecasts of cumulative adopters that are of interest to service providers, the
accuracy of its estimation is key. Further research is necessary here, but one possible approach would be to model
the taking out of the first subscription as an adoption process (where m is bounded by unity), and to model the
acquisition of further subscriptions as a separate process.
However, it is possible that the phenomenon of multiple
subscriptions per person may disappear or at least diminish in future. The main motivation for multiple subscriptions is to game the tariffs available. For example, tariff A
may be cheaper for domestic calls, tariff B may be cheaper
for international calls, and tariff C may be cheaper for data.
If tariffs become simpler, for example with fixed monthly
payments regardless of usage, then the motivation behind
multiple subscriptions is reduced considerably.
In the context of fixed line telephony, the issue of multiple subscriptions has been a minor issue; however, extra

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

1111

Table 2
A summary of the relative forecasting accuracies of three methods using data describing the diffusion of mobiles in 25 countries.
Method

1. Cumulative logistic
2. Non-cumulative logistic
3. Bass

Cumulative errors

Non-cumulative errors

Ranking by rmse

Ranking by mape

Median mape

Ranking by rmse

Ranking by mape

Median mape

1.76
2.04
2.20

1.72
2.08
2.20

29.6
30.0
29.4

1.64
1.76
2.60

1.72
1.72
2.56

78.0
86.3
242.7

lines were acquired by some households during the 1990s


to facilitate internet use. Duffy-Deno (2001) studied the
determinants of an additional fixed line using US data from
90 cities in 1998. Using a logistic regression, the significant
variables were price (the price elasticity was greater for an
additional line than for the initial line) and the age of the
subscriber (those aged 5564 were most likely to acquire
a further line). In a similar vein, Eisner and Waldon (2001)
used a binomial probit model with US data from 1995; they
found that households that already subscribed to online
services were more likely to subscribe to a further line.
2.1.4. Summary of similar studies
Several authors have followed Franks example in developing single-country forecasts for the diffusion of mobile telephony. These analyses are summarised in Table 3 in
order of country. The technology considered and the time
span of the data used are given. The models used are most
commonly the logistic, Gompertz and Bass models. (Unlike the logistic, the Gompertz represents asymmetric diffusion, with the peak adoption occurring before 50% of the
market potential is reached. Its cumulative proportion of
adoption is:
F (t ) = m exp (c (exp (bt ))) ,

(6)

where m is the market potential, and b and c are positive


parameters to be estimated from the data.)
In some cases, the model was estimated without a forecast being made, with the objective in some cases being
to discover the market potential, and in others to discover
whether or not mobile telephony was acting as a substitute for fixed line telephony. In contrast to Frank (2004),
Gamboa and Otero (2009) provide a plausible prediction
interval for mobile telephony diffusion in Colombia using
a bootstrap approach.
Islam and Meade (1996) explored several formulations
of the market potential for UK business telephones using GDP-related variables. They found that, although more
managerial insights are achieved using explanatory variables, these additional insights do not necessarily lead to
more accurate forecasts, due to the need to forecast the
explanatory variables (an explanation of the poor performance of Method 1a described in Section 2.1.1). In another
study with an emphasis on the estimation of market potentials, Chen and Watanabe (2006) looked at diffusion within
the Japanese ICT market, considering both mobile and internet access data sets. They linked the drop in cumulative
fixed line subscriptions during 1998 to substitution by mobile telephony. Evidence that mobile telephony is acting as
a substitute for fixed line telephony has also been found by
Frank (2004) in Finland, Lee and Cho (2007) and Sung and
Lee (2002) in Korea, and Chu, Wu, Kao, and Yen (2009) in
Taiwan.

Wu and Chu (2010) use Taiwanese data to test their


hypothesis that the most appropriate diffusion model
changes as diffusion progresses through the stages of slow
start, take-off, post-inflection and saturation. The authors
define the beginning of the take-off stage as the time when
adoption reaches 10%20%, following Rogers (2003). More
explicit definitions of sales take-off are given by Agarwal and Bayus (2002), Lim, Choi, and Park (2003), Tellis,
Stremersch, and Yin (2003), and Wiorkowski and Gylys
(2006). After considering the logistic, Gompertz, Bass and
ARMA models, Wu and Chu find that the Gompertz gives
a greater forecast accuracy in the pre-take-off stage, while
the logistic is more accurate in the post-take-off stage. As
Wu and Chu note, the appropriate choice of model is casedependent (see also Meade & Islam, 2001). However, the
idea of stage-dependence (within each case) needs further research, particularly considering the width of the
prediction intervals associated with the point forecasts
considered by Wu and Chu. In the early stages, the most
important finding is likely to be the scale of the uncertainty
in the forecast, whichever model is used.
Some authors discuss network effects or network externalities in mobile telephony: that is, they consider the increase in benefits to mobile users as the size of the network
increases. In terms of the Bass model, this effect is captured
by the coefficient of imitation, and there is a corresponding
parameter in the logistic model. Doganoglu and Grzybowski (2007) find that, in Germany, the mobile subscription demand depends on lagged cumulative subscribers
and service prices. In a micro-level study on the adoption of
PCs, Goolsbee and Klenow (2002) found that a high fraction
of surrounding PC owners was the most important determinant of buying a PC. Primarily, this is evidence of a high
coefficient of imitation, although there may be some network effect related to internet access.
Summarising the accuracy of the analyses identified in
Table 3 is not straightforward. In only six of the 14 analyses identified is it possible to find a measure of the outof-sample forecasting accuracy. When found, these
measures may be averaged over different origins and/or
different horizons. Taking these reservations into account,
the mapes found for the most accurate forecasting method
considered are between 5% and 10%; these measures refer to horizons of three years or less. However, we see in
Table 2 that, for the average mape over horizons of one to
seven years, the median value over 25 countries is 30%. The
mape value obtained for Finland (2.7%) is atypically low. In
general, the more typical pattern is that the forecasting accuracy deteriorates sharply for horizons in excess of four
or five years.

1112

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

Table 3
A summary of single-country analyses of ICT diffusion.
Authors

Country

Gamboa
and Otero
(2009)
Frank
(2004)

Technology

Time period

Models used

Market
Potential

Significant
covariates

Forecast (Y/N)

Colombia Mobile phone

1995 Q42008
Q1

Logistic,
Gompertz

Unconstrained

NA

Evaluation for 1
year out-of-sample

Finland

Mobile phone

19811998

Logistic

Proportion of
population

Fixed line
(substitution)

Michalakelis
et al. (2008)

Greece

Mobile phone

19942005 Q3

Unconstrained

NA

Dergiades
and Dasilas
(2010)

Greece

Mobile phone

19932005

Versions of
Bass, Gompertz
and logistic
Gompertz,
logistic and
ARMA

NA

Singh
(2008)

India

Mobile phone

19952006

Logistic,
Gompertz

Massini
(2004)

Italy

Mobile phone

19902001

Logistic,
Gompertz

Chen and
Watanabe
(2006)

Japan

19912002
(mobiles)

Versions of
logistic,
choice/diffusion

Sung and
Lee (2002)

Korea

Fixed line
telephony,
mobile
telephony,
internet access
Fixed line
telephony
connections
and
disconnections

Time varying
and
determined
exogenously
Constrained
and
unconstrained
Either rate of
diffusion or
saturation level
function of
economic
covariates
Unconstrained

12-year
extrapolation and
prediction interval
Evaluation for
3 years
out-of-sample
Prediction interval
by bootstrap

19911998 for 8
regions

Log-linear

NA

Lee and Cho


(2007)

Korea

Mobile phone

19842003

Logistic, ARMA

Botelho and
Pinto (2004)

Portugal

Mobile phone

19892000 Q1

Wu and Chu
(2010)

Taiwan

Mobile phone

19892007

Chu et al.
(2009)

Taiwan

Mobile phone

Exponential
growth,
logistic,
Gompertz
Bass,
Gompertz,
logistic and
ARMA
Bass, Gompertz
and logistic

Massini
(2004)
Hwang,
Cho, and
Long (2009)

United
Kingdom
Vietnam

Mobile phone

19902001

Mobile phone

19952006

Logistic,
Gompertz
Bass, Gompertz
and logistic

2.1.5. Other single-country analyses related to ICT diffusion


In an early study, Jeffres and Atkin (1996) looked at
the factors affecting the diffusion of Integrated Systems
Digital Networks (ISDN), which was perceived to be slower
than expected. They found that the more educated were
less likely to use ISDN for entertainment purposes, and
that a heavy news consumption was positively related
to a desire for ISDN. In a more recent study, Robertson,
Soopramaniem, and Fildes (2007) considered the diffusion
of residential broadband in the United Kingdom. A postal
survey was used to gather data about the length of a

NA

Various
extrapolations

Digital dummy

No

Fixed line
(substitution)

Focus on market
potentials

No: looking for


substitution by
mobiles

Proportion of
population
(same
approach as
Frank, 2004)
Unconstrained

Mobile stock has


negative effect
on connections
and positive
effect on
disconnections
Digital
changeover,
GDP, fixed line
telephony
(substitution)
NA

Unconstrained

NA

Evaluation for
3 years
out-of-sample

Unconstrained

Deregulation,
fixed line
(substitution)
Digital dummy,
price, tariff
NA

Focus on market
potentials

As Italy
Unconstrained

Evaluation for 1
year out-of-sample

Extrapolation only

No
Evaluation for 1
year out-of-sample

householders subscription to broadband and income. They


considered the market for household broadband as a
group of segments determined by income, where each
segment was modelled as a Gompertz diffusion process.
Using data up to 2005, their combined forecast, across all
segments, is for a broadband coverage of 95% by 2015. This
forecast is looking a little optimistic, with Ofcom (the UK
telecommunications regulator) reporting in 2013 that 75%
of UK adults had a broadband connection (fixed or mobile).
We conclude this section with a consideration of two
studies relating to topics that have received little cover-

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

age to date. In a relatively early study of the number of internet providers, Rai, Ravichandran, and Samaddar (1998)
considered the following diffusion models: exponential
growth, logistic and Gompertz. The choice of models here
is questionable, as there is no obvious upper bound on the
number of providers. Exponential growth was shown to be
the better predictor, as the quarterly data showed no sign
of an inflection. They accept that their modelling approach
does not include any of the policy variables that may affect
diffusion. Looking ahead, one consequence of the growth in
the usage of ICT hardware, mobile telephones, desktop and
laptop PCs, and tablets is the issue of the sustainable and
environmentally acceptable recycling or disposal of these
items as they become obsolete. Yang and Williams (2009)
use a logistic model to forecast the generation of obsolete
PCs in the US. Their forecast depends on forecasts of demand for PCs and assumptions about the average lifetime
of the PC. They predict that there will be between 115 and
144 million obsolete PCs in 2050, indicating a potential major shortfall in US recycling facilities. Although their forecast focuses on PCs (where demand has stalled at the time
of writing, 2014), alternative technologies, e.g. tablets, will
substitute for PCs, possibly with shorter lifetimes.
2.2. Multi-country analyses of ICT diffusion
Researchers have analysed ICT multi-country data sets
with a variety of objectives. We begin by describing the
case where the objective is to provide forecasts from short
time series by exploiting the availability of a cross-section
of data from several countries. Secondly, we consider a
set of broadly homogeneous studies. These use a multinational data set to describe the diffusion of fixed or mobile
telephony or the internet, and evaluate the power of a set
of covariates to explain the differences in country diffusion
rates. Thirdly, we describe other multi-country analyses
with various different objectives that are not covered in the
other sections.
2.2.1. Multi-country analyses with short time series
In the analyses covered in this section, the researchers
use cross-sectional data to compensate for a lack of time
series data. Dekimpe, Parker, and Sarvary (1998) draw attention to the weaknesses of using the Bass or similar models with short data series. They emphasise the importance
of sample-matching when using cross-sectional data, in
order to make comparisons across countries meaningful.
They propose a staged estimation procedure using sample matching to allow the use of multi-national data for
modelling and forecasting diffusion. The covariates that
they consider important in matching include income, population heterogeneity, demographics and politics. Islam,
Fiebig, and Meade (2002) use multinational cross-sectional
data to compare pooling methods in the production of
growth forecasts for digital cellular phones, ISDN lines and
fax connections. The pooled forecasts tended to be more
accurate than the available alternatives. Using a similar
approach, Hamoudia and Islam (2004) model and forecast the demand for SMS (short message service, i.e. texts).
They pool annual data from seven European countries over
five years from 1998 to 2003 using a linearised Gompertz

1113

model. Using the mape as a measure of accuracy, their median (over the countries considered) values are 9.1%, 20.1%
and 25.5% over one, two and three years, respectively.
2.2.2. Multi-country diffusion analyses using covariates
In order to enable more general conclusions about the
drivers of diffusion to be drawn, several authors have carried out analyses of mobile diffusion over many countries,
where the coefficients of a diffusion model are parameterised to relate the growth rate or market potential to
economic or infrastructural variables. We summarise these
studies in Table 4. This table is laid out in a similar way to
Table 3, with two differences: the papers appear in chronological order, and there is no column indicating whether
forecasting was carried out, as these studies generally did
not have forecasting as an objective. One of the most cited
papers is that of Gruber and Verboven (2001), who use a
logistic model for the diffusion of mobile telephony in 15
EU states from 1984 to 1997. They parameterise the logistic
using a range of economic variables; these effects are estimated using non-linear least squares on the multi-country
time series data. Their main conclusion is that the prime
cause of an increased diffusion was the transition from
analog to digital; increased competition was also found to
increase the diffusion rate. Their analysis also suggests that
mobile telephony is a substitute for fixed line telephony.
Other researchers have developed these themes, namely
the transition between technological generations, the substitution between fixed and mobile telephony, and the effects of competition and other covariates.
Several studies consider the effect of the transitions
between generations of mobile telephony. Grajek and
Kretschmer (2009) study multiple generations of mobile telephony and extend the information set from subscription data to usage data (average minutes of use per
month). They find that successive generations are substitutes when analysed via subscription data, but act as complements when considered via usage data. Bohlin, Gruber,
and Koutroumpis (2010) examine the factors affecting the
diffusion of newer generations of mobile phones. They find
that income/head, urbanisation and internet penetration
have a positive impact on diffusion across all generations.
However, although the diffusion of the first generation
(analog) stimulated the diffusion of the second generation
(digital), the diffusion of the second generation did not
affect the diffusion of the third generation. The effect of
inter-firm competition, although important for earlier generations, diminishes by the third generation. A number of
covariates are identified as determinants of the diffusion of
mobile telephony, including national wealth (using GDP as
a proxy), regulation, competition and price. Kiiski and Pohjola (2002) and Liikanen, Stoneman, and Toivanen (2004)
find that an increase in GDP per capita is associated with
a faster diffusion. Comer and Wikle (2008) study the reasons for the differences in the rates of diffusion of mobiles
between different countries. Using the growth rate of subscriptions between 1995 and 2005 in each country as the
dependent variable, they find that the most significant explanatory variable is GNP. Using a worldwide data set, 75%
of the variation is explained by this variable; using only
data from Asia, this rises to 90%. Koski and Kretschmer

1114

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

Table 4
A summary of multi-country analyses of ICT diffusion.
Authors

Countries

Technology

Time period

Models used

Market potential

Significant covariates

Gruber and
Verboven (2001)

15 EU countries

Mobile phone

19841997

Logistic

Too early to
estimate accurately

Kiiski and Pohjola


(2002)

23 OECD
countries

Internet

19952000

Gompertz

Time dependent
function of cost and
lagged adoption

Madden and
Coble-Neal (2004a)

56 countries

19952000

80 countries

19921998

Dynamic
panel
regression
Logistic

NA

Liikanen et al.
(2004)

Fixed line
and mobile
phone
Mobile phone

Jang et al. (2005)

29 OECD countries + Taiwan

Mobile phone

19802001

Logistic

Estimated as
proportion of
population
Unconstrained

Sundqvist, Frank,
and Puumalainen
(2005)
Koski and
Kretschmer (2005)

64 countries
(results based on
25)
32 industrialised
countries

Mobile phone

19812000

Bass

Not clear

Digital
mobiles (2G)

19912000

Log
(diffusion) as
dependent

Rouvinen (2006)

78 countries

Mobile phone

19912000

Gompertz

Constrained

Bagchi et al. (2008)

4 multi-country
regions + India
and China
206 countries

Fixed line
and mobile
technology
Mobile phone

19752003
and
19852003
19952005

Price
adjusted
logistic
Regressions

Price dependent

Digital/analog
changeover; number of
competing suppliers.
Diffusion associated
positively with
GDP/capita and
negatively with access
cost
Evidence of substitution
for fixed line, significant
income and price effects
GDP has positive effect,
evidence of fixed line
substitution
Digital/analog
changeover; number of
competing suppliers,
fixed line (substitution)
Cluster analysis used to
relate diffusion with
cultural dimensions
Standardisation
accelerates diffusion,
price competition greater
between standards than
within
Comparison of developed
and developing countries
competition speeds
diffusion, conflicting
standards retards
diffusion
Price and social/political
covariates

Not clear

GDP (log and growth rate)

Up to 157
operators in 41
countries

Multiple
generations
of mobile
phones
3 generations
of mobile
phones

19982004

Logistic

Estimated as
proportion of
population

Heterogeneity of adopters
masked network effects
in usage data

19902007

Logistic

Used rate rather


than penetration

Income/head,
urbanisation, internet
penetration

Comer and Wikle


(2008)
Grajek and
Kretschmer (2009)

Bohlin et al. (2010)

177 down to 67
countries

(2005) look at the effects of regulation and competition


on digital (2G) mobiles. They find that a greater standardisation accelerates both the market entry and the
subsequent diffusion; in addition, they also find that the
liberalisation of fixed line telephony accelerated the entry
of digital. Jang, Dai, and Sung (2005) and Rouvinen (2006)
find that the number of suppliers, a measure of competition, speeds diffusion, while Rouvinen (2006) also finds
that competing technical standards retard diffusion. Chinn
and Fairlie (2007) use panel data from 161 countries over
the period 19992001 to compare the determinants of the
penetration of the PC and of the internet. They find that
the income differential is an important variable in explaining the penetration variability for both technologies. For
the internet, the quality of regulation has an important influence on penetration, followed by the telephone density.
For both technologies, they find that the influence of education is swamped by those of the variables mentioned.
To discover the reasons for the differences in the rate of
technology diffusion of 3G mobile phones across countries, Islam and Meade (2012) investigate the impacts of

market factors, measured by competitive fractionalization,


and economic globalization across 35 countries using a
multi-country diffusion model. They use non-linear mixed
modelling with pooled multi-country data to estimate a
generalized Bass model, taking into account the unobserved heterogeneity in market saturation levels. They find
that increasing the competition within the market tends to
increase the market potential, and also find that countries
with a higher economic globalisation index (e.g. a greater
openness to trade) are associated with higher rates of diffusion.
The effect of prices on the rate of diffusion of mobile
telephony has been approached in several different ways.
Madden and Coble-Neal (2004a) use an economic agent
approach to model the substitution effect between fixed
line and mobile telephony. Their data covers 56 countries
over the period 19952000. They find a significant substitution effect, together with a significant price effect,
although the network effect was greater in terms of causing increased subscriptions. Bagchi, Kirs, and Lopez (2008)

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

study the impact of price decreases on the diffusion of mobiles and fixed telephony in six large regions. They use the
price-adjusted logistic model of Gurbaxani and Mendelson
(1990), and essentially modify the potential by introducing a multiplicative factor reflecting the price; thus, Eq. (1)
becomes
F (t ) =

m exp ( ln (Pt /P0 ))

(1 + exp ( (a + bt )))

(7)

where Pt reflects the price at time t. If the coefficient


is found to be significantly greater than zero, this indicates that the potential level of subscriptions has been increased due to the price decreases. They found that price
decreases had a greater effect in lower income nations, especially if the countries are politically stable, with high education levels. Islam and Meade (2011) used a parametric
hazard model and discrete-time survival mixture analysis
(DTSMA) approaches to model the hazard probability of
the time to sales takeoff for cellular analog telephony in
70 countries. They determine the impacts of three market
factors: price, the number of competitors, and the number of competing standards. The authors showed the
relative advantage of DTSMA in its ability to recognize
unobserved heterogeneity using latent classes. A failure
to account for unobserved heterogeneity can cause an
underestimation of the hazard probabilities. They found
that relatively falling prices, relatively greater numbers of
competitors, and relatively fewer competing standards are
each associated with relatively higher hazard probabilities.
2.2.3. Other multi-national analyses related to ICT diffusion
Restrictions on supply have been an obstacle to the diffusion of fixed line telephony in many countries. Islam and
Fiebig (2001) carried out a multinational study to estimate
saturation levels for fixed line telephony in 46 supplyrestricted countries. They used pooled cross-sectional estimates to forecast the supply-restricted demand where
little or no data were available. Bartels and Islam (2002)
analyzed data from 28 countries to investigate the causes
of supply restrictions for main telephones, and found that
technical efficiency is the major determinant of supply
restrictions.
Looking at a different aspect of mobile telephony, Shin
and Bartolacci (2007) consider the diffusion of mobile
virtual network operators (MVNO). A MVNO provides users
with mobile services without its own government-issued
bandwidth. The analysis is a cross-sectional regression of
the MVNO market share, explained by factors that describe
the market structure. They find that diffusion is retarded by
vertically integrated markets, which are common in Asia,
and enhanced by horizontal market structures, which are
common in Europe.
Gruber and Koutroumpis (2013) examine the effects
of regulation on the adoption of broadband in a country.
The regulation of former national telecommunications monopolies is studied: to encourage market entry, incumbent
operators were instructed to share, sell or split their infrastructure. Data for 167 countries over an11-year period
are used to relate the diffusion (using a logistic model)
of broadband to regulatory interventions. They found that

1115

competition between firms on the former monopolys network tended to accelerate the adoption of broadband,
whereas competition between different access technologies did not.
2.3. Choice modelling and other extensions
In circumstances where time series data are in short
supply, or absent, as in the case of a new product, it may
be appropriate to survey consumer intentions. Survey data
are then used, with or without time series data, to estimate
the parameters of a diffusion model. Kumar, Nagpal, and
Venkatesan (2002) combine a market share survey method
with an overall Bass-type projection of market size to
predict the demand for mobile phones at a firm level.
Jun and Park (1999) propose a choice-based approach
where there is a utility associated with each available
choice. In cases where several generations of a technology
are available simultaneously, the choice could be between
each of these generations, or non-adoption. At time t, the
expected utility for choice k is Vkt , and a multinomial logit
is used to give the probability, Pkt , of an individual making
exp(V )
choice k at time t, where Pkt = exp ktV . The expected
j

( jt )

utility is parameterised, for example Vkt = ak (t k ),


where k is the time elapsed since the launch of the technology. The
expected number of adoptions of choice k at
time t is mk Yk,t 1 Pkt . The information source for the
utility function can be a survey-based approach if no data
are available, using conjoint analysis, for example. Alternatively, if adoption data are available, the utilities may be
estimated from these data. This model is developed further
by Jun, Kim, Park, Park, and Wilson (2002), who extend the
choice-based diffusion model to include both substitution
and competition. Ida and Kuroda (2006) study the demand
for broadband in Japan using a discrete choice model. They
find that the demand for the main source of broadband,
ADSL (asymmetric digital subscriber line), is price inelastic,
in contrast to the demand for cable TV and fibre to home
services, which have been found to be price elastic.
An example of forecasting for a new product is provided by Jun et al. (2000), who forecast the demand for
a new service: a low earth orbit mobile satellite service
(LEO). They use a mixture of analogy and surveys to forecast demand. The analogy is the use of mobile phone penetration data in Korea to provide estimates for the Bass
model describing the LEO service penetration. The survey
data were used to estimate the market potential of LEO service and the subscription probability using a logit choice
model. The authors used the diffusion parameters of the
existing mobile phone service and the market potential derived from the survey to make market predictions using the
diffusion model. The forecast is for approximately 280,000
subscribers in 2005; however, it is undermined by the authors comments that the survey is probably too optimistic,
given that it had been taken five years prior to publication
and during a period when the Korean economy was much
more buoyant. Their reservations appear to have been justified, as there were only 20,000 subscribers in 2009 (Korea
IT Times, 7 April 2009).

1116

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

In forecasting the demand for large screen televisions,


Lee, Cho, Lee, and Lee (2006) adopt the following sequential approach. They model utility using conjoint analysis;
model the price of television as a decreasing function over
time; produce a dynamic random utility function; and use
a Bass model via the estimation of p, q and m based on time
series data from Q1 1998 to Q2 2003 for the whole large
screen market. Finally, forecasts for each television type
are derived by multiplying the total demand by the dynamic choice probability. Kreng and Wang (2009) use the
Jun and Park model on shipment data to model the substitution of liquid crystal displays (LCD) for cathode ray televisions (CRT). The choice considered is between a CRT, a
medium-size screen LCD and a large screen LCD. In addition to a fixed market potential, they also consider a variable market potential that changes with exp(g time),
where g is an estimated parameter. They demonstrate that
the introduction of the variable market potential improves
the forecasting accuracy. This modification of the market
potential can be considered as a proxy for the changing relative prices of CRTs and LCDs, and the move from analog
broadcasting to digital broadcasting (which is inaccessible
for CRT).
Jeon, Kim, and Sohn (2010) look at the convergence of
digital products that incorporate the following into mobile
telephones: a PC, a camera, a music player (MP3), and
medical devices. The study used conjoint analysis with a
Korean panel, and found that the only converged product
for which a premium could be charged was computing
power (essentially faster data reception/transmission).
In the last decade, there has been a high growth in
mobile broadband services3G systems and wireless local area networks (WLAN). The former typically have good
coverage but are limited by low transmission speeds and
high usage costs. WLAN have good transmission but limited coverage. An emerging technology is portable internet services (PIS), which are characterised by high speed
transmission, portability, mobility and multi-media communication. An example of this is WiBro, which was introduced in Korea in 2006 (see Nam, Kim, & Lee, 2008). Nam
et al.s study is in two parts. Firstly, the characteristics of
mobile broadband used by consumers, choosing between
3G, WLAN and WiBro, are identified and survey data collected. A conjoint analysis is then carried out to predict
relative market shares. Secondly, a Bass model is used to
forecast diffusion. In the absence of time series data, the
Bass coefficients are chosen by an expert panel.
VoIP (voice over internet protocol) telephony is a technology that gained acceptance during the 1990s, initially
as a novel use of the internet and later as a serious contender with other forms of telephony. Ida, Kinoshita, and
Sato (2008) used conjoint analysis with Japanese data, and
found that VoIP was considered an additional option for internet use rather than a substitute for fixed line telephony.
However, in a later study in Korea Kwak and Lee (2011)
found that the picture was evolving. Due to an increased
investment in VoIP by telecoms suppliers, they found that
the call rates for VoIP and fixed lines were the factors that
affected the VoIP demand the most. Thus, at the level of
deciding which medium to choose for making a call, VoIP
is a substitute for fixed lines. We did not find any studies

more recent than that of Ida et al. (2008) that have looked
at the decision at subscription level, investigating whether
internet users are prepared to forego a fixed line. The outcome of these national analyses will be affected by the level
of competition between the different service suppliers and
the effect of this on the pricing of both the internet and
fixed line telephony. For example, in the United Kingdom,
both services are often supplied by the same company. A
similar argument can be used with VOIP over the mobile
internet and mobile telephony, in that differences in tariffs
will tend to determine preferences.
Lee, Lee, and Kim (2008) look at the diffusion of home
networking. This is the distribution of data between ICT
appliances for providing communications, multimedia and
automation services anywhere in the home. Conjoint analysis is used with survey data from both consumers and
construction companies (who install the system in newly
built housing). The estimated utilities are used with a logistic model, and incorporate expert judgement into their
forecast via Bayesian updating.
2.4. Summary: the flow of benefits between ICT and forecasting methodologies for diffusion
Here, we look at the flow of benefits between ICT applications and diffusion modelling and forecasting methodologies.
In the case of studies of a single series, their ICT applications, mainly the forecasting of national mobile telephone
diffusion, have led to some methodological developments,
mainly in the area of parameterising market potential. On
a related topic, we note that there is an opportunity for further research in the appropriate modelling of subscription
data, rather than treating them as if they were adoption
data. In general, studies have focussed on the trend represented by the S-curve, with little attention being paid to the
associated error process. The analysis in Sections 2.1.1 and
2.1.2 shows that a consideration of the error process has a
significant effect on the estimated trend, and consequently
on the forecast accuracy. In addition, the error process
feeds into the construction of a prediction interval (used by
only two of the 14 studies in Table 3), giving valuable information about the uncertainty associated with the forecast.
In general, the multi-country diffusion studies focus on
identifying the determinants of diffusion, rather than on
direct forecasting. Although no particular methodological
innovations have been driven by the use of ICT data,
the findings of these studies successfully illuminate the
factors that determine the diffusion of ICT products. The
identitifcation of these factors allows business strategists
to choose appropriate timetables for their entry into
international markets.
ICT has provided an important impetus to the development of choice modelling. The models developed have
provided practicable solutions to the problem of forecasting with little or no data. The nature of ICT devices, such as
mobile telephones or telephones, is well suited to choice
modelling. These devices are available in a small number
of basic types, and each type is available with a range of
extra capabilities. This device structure facilitates the construction of consumer intention surveys that will not only

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

provide the basis for modelling and forecasting, but also


feed into product design. However, the Korean LEO example shows the importance of timeliness and relevance in
the survey questions; any compromise on either of these
issues is likely to undermine the forecasting accuracy. Furthermore, in none of the cases in which forecasts were
based on survey information and little or no time series
information was any attempt made to quantify the uncertainty in the forecast. Further research in this area would
add to the value of this methodology.
3. Time series
The forecasting of time series is undoubtedly the
most heavily researched area in forecasting; see De Gooijer and Hyndman (2006) for a review. The question
we ask here is: is there anything distinctive about an
ICT/telecommunications time series that justifies special
consideration? This section attempts to answer this question. In Section 3.1, we discuss a clearly defined and relevant activity, forecasting in call centres. In Section 3.2, we
review the literature on the forecasting of ICT-related time
series. We draw our conclusions in Section 3.3.
3.1. Forecasting for call centres
Scheduling the manpower available to staff call centres over short, medium and long horizons is an ongoing problem for many service organisations. For further
background to this topic, Gans, Koole, and Mandelbaum
(2003) provide a good overview of call centre operations
and their management; a bibliography of call centre related studies was compiled by Koole (2008); and Mandelbaum (2003) summarises some more recent advances in
management science applied to call centre management.
Since personnel costs account for around two-thirds of call
centre operating costs, effective manpower planning relies on a good forecasting accuracy. The relevant forecasts
are over a range of horizons, from two or three months,
to one to three weeks, to updating within-day forecasts in
real time. We will begin by discussing the different types
of call centres that have been examined, and then consider
the modelling approaches used.
The modelling and forecasting of call frequencies has
been studied for several different types of centre. In Table 5, we list studies on forecasting for call centres categorised by call centre type, where the modelling approach
used for each study is identified. The call centre forecasting
studies fall into four categories. The first is emergency call
centres, which require a zero or very short waiting time
(fire, police, ambulance, and health-related enquiries). The
second is enquiries, and includes directory enquiry calls
to telephone companies. The third category is telemarketing, where sales of goods or services are made following
the receipt of a catalogue or the sight of an advertisement.
The fourth category is call centres run by financial institutions. These categories are not mutually exclusive; for example, some financial institutions may run telemarketing
institutions. The practical value of accuracy is clear in many
types of call centre forecasting: an under-estimation of
call frequencies leads to under-staffing, which may lead to

1117

either delays in dealing with emergencies or lost sales;


over-estimation leads to over-staffing, which equates to
expensive wasted capacity. In general, the earlier data sets
used were confidential. This is unfortunate, as alternative
modelling approaches to those used in the original studies
cannot be evaluated or compared. However, there has been
some comparison of different forecasting methods over the
same data sets in the more recent studies.
According to several authors, including Shen and Huang
(2008a) and Weinberg, Brown, and Stroud (2007), the most
commonly used model for the number of calls arriving at a
centre is

Nij + 1/4 = + i + j + ij ,

(8)

where Nij calls arrive during the jth sub-section of day i.


The error term, ij , is assumed to be normally distributed,
implying that the number of calls arriving in a particular
time slot is a non-central chi-square random variable. Separate considerations of the inter-day and intra-day components, i and j respectively, are common to most of the
modelling approaches used. The differences between approaches lie in the choice of the random variable and the
estimation procedure.
Early analysts, such as Andrews and Cunningham
(1995), Bianchi, Jarrett, and Hanumara (1993, 1998), and
Mabert (1985), used well-known time series methods
such as HoltWinters and ARIMA for forecasting daily call
counts. Later, more customised modelling was carried out
by Tych, Pedregal, Young, and Davies (2002), who developed an unobserved components model for calls to a financial services call centre. With the availability of more
detailed data, Avramidis, Deslauriers, and LEcuyer (2004)
identified four properties of the arrivals of calls at call centres. These are: (a) the variance of the total number of calls
per day is greater than its mean; (b) call arrival rates vary
during the day; (c) there is a positive correlation between
call counts in different periods of the day; (d) there are
significant correlations between call counts on successive
days. To capture properties (a), (b) and (d), they develop an
inhomogeneous Poisson process model of call arrivals, using data from a Bell Canada call centre. Brown et al. (2005)
carry out a statistical analysis of three aspects of call centre
management, using data from an Israeli bank call centre.
Forecasters main preoccupation is with the arrival process,
which Brown et al. consider as an inhomogeneous Poisson
process. Furthermore, Brown et al. also consider the other
two components of a queuing system: the customer waiting time and the duration of service. Extending the work
of Brown et al. (2005), Weinberg et al. (2007) develop a
Bayesian approach for estimating the parameters of the
inhomogeneous Poisson process in order to provide density forecasts for daily arrival rates; their data were from
a US bank call centre. Shen and Huang (2008a) consider
the data as a vector time series, with daily observations
of the intra-day profile (represented by the frequency per
quarter-hour slot over the business day). Their data were
gathered from a US financial institution call centre. They
use singular value decomposition to reduce the dimensionality of the vector, then use an autoregressive model to
forecast the resulting set of time series. They demonstrate

1118

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

Table 5
A summary of forecasting-related studies of call centre activities.
Type of call centre

Emergency services

Enquiries

Tele-marketing

Banking/financial services

Authors

Modelling approach

Mabert (1985)
Tandberg (1995)
Klungle and Maluchnik
(1997)
Taylor (2012)

HoltWinters variants and ARIMA


Emphasis on time-of-day and day-of-week effects
Linear regression

Sparrow (1991)
Xu (1999)
Avramidis et al. (2004)

Queuing theory
Trend extrapolation
Poisson based stochastic models

Bianchi et al. (1993)


Andrews and
Cunningham (1995)
Bianchi et al. (1998)
Soyer and Tarimcilar
(2008)

HoltWinters and ARIMA


ARIMA and transfer function

Betts, Meadows, and


Walley (2000)
Antipov and Meade
(2002)
Tych et al. (2002)
Brown et al. (2005)
Shen and Huang (2005,
2008a,b)
Weinberg et al. (2007)
Taylor (2008)

Linear regression

Taylor (2012)

Density forecasts based on Poisson models

HoltWinters and ARIMA


Emphasis on effect of marketing campaigns

Emphasis on advertising and day-of-week effects


Unobserved components/seasonal ARIMA
Queuing theory approach to modelling arrivals, customer patience and service duration
Singular value decomposition is used to condense the dimensionality of the intra-day
profile
Bayesian forecasting of an inhomogeneous Poisson process
Seasonal ARIMA, an extended HoltWinters approach plus some regression based
approaches
See above

that their approach is clearly more accurate than the historical average model (see Eq. (8)) on an intra-day basis.
Furthermore, using the same data set as Weinberg et al.
(2007), Shen and Huang demonstrate that the accuracy of
their method is very similar to that achieved by Weinberg
et al. Shen and Huang (2008b) further refine their theme of
a reduced dimensionality of the daily profile, in conjunction with the assumption of an inhomogeneous process.
The question arises as to whether the conventional
time series approaches used by earlier authors can be
developed usefully to offer similar accuracies to these later
approaches. Broadly speaking, the answer appears to be
in the affirmative. Taylor (2008) compares the forecasting
accuracies of several approaches over six series of intraday call arrival data, five series from a UK bank, and one
from an Israeli bank (as used by Brown et al., 2005),
over horizons ranging from 30 min to two weeks. Models
using two seasonal patterns, intra-day and intra-week, are
used in both an ARIMA framework and a HoltWinters
framework. Using the mean absolute error as a criterion,
the doubly seasonal ARIMA and HoltWinters models were
more accurate for horizons of less than a week, while a
moving average approach was more accurate for longer
horizons. Taylor (2012) develops the doubly seasonal
HoltWinters model further in order to provide density
forecasts for an intra-day call frequency. The forecasting
accuracy of this model was shown to be similar to that
of Shen and Huang (2008b), but the performance of the
exponential smoothing model was more consistent across
the three series analysed.
In some centres, calls are stimulated by marketing
campaigns. Antipov and Meade (2002) model the response of call frequencies to such campaigns in a financial

institution whose marketing relied heavily on newspaper and television advertisements. They develop a model
with a dynamic level, multiplicative calendar effects (both
within-year and within-week), and a multiplicative advertising response. The inclusion of the response to advertising is shown to almost halve the mean absolute forecast
error. Soyer and Tarimcilar (2008) pursue a modelling approach similar to that of Avramidis et al. (2004), but include
variables that describe different advertising strategies and
promotion policies. Their modelling approach allows the
effectiveness of different marketing campaigns to be evaluated.
3.2. ICT time series
We will discuss several sets of time series that describe
several ICT/telecommunications activities. Section 2 was
subdivided according to the modelling approach; in Section 3.2, it makes more sense to subdivide by the application area, defined by a set of time series. The sets of ICT
time series considered do not appear to have much in common, but do tend to be homogeneous within each set. Some
example time series from three data sets of telecommunications series are shown in Figs. 24; analyses of these
series will be discussed below. We discuss telecommunication series first, followed by a discussion of the work on
other ICT series.
3.2.1. Univariate telecommunications time series
An early study of telecommunications data is that of
Grambsch and Stahel (1990); the data set is the number
of circuits in use under a range of Special Services offered
by a Bell operating company in the USA. These 261 telecoms time series are characterised by negative trends, of-

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

1119

Fig. 2. Telecommunications data: a random selection of five series (Grambsch & Stahel, 1990). Note that the time axis is in months, and the volume axis
is the number of circuits in service.

Fig. 3. Telecommunications data: a random selection of five series (Makridakis & Hibon, 2000). Note that the time axis is in months, and the volume axis
is unknown.
80
70

Traffic Index

60
50
40
30
20
10
0
0
00 000 000 000 000 000 2000 000 000 000 000 000 000 0002000 000 0002000 000 000 000 2000000 001 001 01 001 0012001 00/1
2
/
2 2 /
2 2 2 / 2 2 2 0 2 /2 /
2 2 /
2 /2 2 /2 2 / 2 2 /2 2
2
/02 3/ 03 3/ 04 4/ 05 5/ 6/ 06 07 7/2 8/ 8/ 09 9/ 9/ 10 0/ 1/ 1/ 12 2/ 1/ 1/ /2 2/ 03 03 3/
18 3/0 17/ 31/0 14/ 28/0 12/ 26/0 9/0 23/ 7/ 21/0 4/018/0 1/ 15/029/0 13/ 27/1 10/1 24/1 8/ 22/1 5/0 19/02/0216/0 2/ 16/ 30/0
Time

Fig. 4. Internet traffic data (Madden & Coble-Neal, 2005).

ten with large steps (see Fig. 2). Their paper proposes the
robust trend model (see Meade, 2000) and demonstrates
it using their data set. This case is the exception in Section 3.2, in which the modelling approach is developed
specifically for considering the properties of the time series. The subsequent analyses demonstrate how well the
robust trend model achieves its aim. Fildes, Hibon, Makridakis, and Meade (1998) used a variety of models on this
telecommunications data set, including ARIMA and exponential weighted moving average models; they found no
model that outperformed the robust trend. Gardner and
Diaz-Saiz (2008) also analysed this data set; aiming to challenge the robust trend, they showed that simple exponential smoothing with drift was almost as good.
In the M3 competition, a comparison of the accuracy
levels of many different time series modelling approaches,

co-ordinated by Makridakis and Hibon (2000), the data set


of 3003 time series included 149 telecoms series. However,
on further investigation (suggested by Robert Fildes), we
found 120 of the 149 telecommunications series to be from
the Grambsch and Stahel data set. The remaining 29 time
series are identified as monthly. A sample of these series is
shown in Fig. 3; no further information is given about these
series. We see that four of these five plots show downward
trending series and no obvious seasonality. Madden and
Tan (2007) examine the two subsets of the M3 telecommunications data separately. For the monthly data, ARIMA
tends to be more accurate according to most accuracy measures. The 120 time series are noted to be similar to the
Grambsch and Stahel data set, but the analysis is unreliable. For example, the mean absolute percentage errors
quoted (by Madden and Tan) differ both in relative rank-

1120

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

ings and by several orders of magnitude from those quoted


by Fildes et al. (1998), whose results were validated by
Gardner and Diaz-Saiz (2008).
Madden, Savage, and Coble-Neal (2002) compare
econometric and time series models for forecasting IMTS
(international message telephone service) traffic. The data
are six annual time series from 1964 to 1997 describing the outgoing message traffic (in min) from the US to
Australia, Hong Kong, Japan, the Philippines, South Korea
and Taiwan. The other available covariates are price, US
GDP, and the relevant incoming message volume. Vector
autoregression (VAR) and error correction models (ECM)
are compared with univariate models: ARIMA and a linear
trend model. A combined forecast is also computed as an
unweighted average of the ARIMA, VAR and ECM forecasts.
The forecasts are evaluated over the period 19941997; it
is not clear whether the VAR and ECM use covariate values post-1993. Depending on the accuracy measure used,
the combined forecast was most accurate for most countries (mean square error), while ECM was most accurate
for three countries (geometric root mean square error).
Two exceptions to the use of conventional modelling of
ICT time series were found. The first study discussed here is
that of Mastorocostas and Hilas (2012). They consider time
series of monthly call volumes from their university PBX.
An earlier set of these data was considered by Hilas, Goudos, and Sahalos (2006), who used conventional methods.
They used data from 1988 to 2002 for estimation, and forecast the data for 2003; of the eleven methods considered,
linear extrapolation with seasonal adjustment was most
accurate (e.g., a mape of 14.0% for the national calls series),
followed by a damped trend with multiplicative seasonality (e.g., a mape of 19.5% for the national calls series). Using data from 1998 to 2006 for model estimation and 2007
for forecast evaluation, Mastorocostas and Hilas develop a
fuzzy neural network model for forecasting two monthly
call volume time series. A slightly different set of eleven
conventional methods is also used, including two versions
of seasonal adjustment with linear extrapolation. The fuzzy
neural network out-of-sample forecasts are more accurate
(with a mape of 12.1% for the national calls series) than
the best performing conventional method, damped trend
with multiplicative seasonality (with a mape of 19.1% for
the national calls series). However, the linear extrapolation with seasonal adjustment that performed well previously performed comparatively poorly on the later data
(mape = 24.4%), suggesting that the results are sensitive
to the period of data chosen.
3.2.2. Internet time series: forecasting usage and provision
Madden and Coble-Neal (2004b, 2005) collected a data
set comprising 59 time series of 232 daily observations of
an (internet) traffic index; the index ranges from 0 (very
congested) to 100 (no traffic). The data are close to white
noise, with means of around 5060 and occasional severe
downward spikes; an example time series is shown in
Fig. 4. The authors use the following forecasting methods:
ARARMA, ARMA, Holts, Holts damped trend, robust trend,
a random walk and filtered trend. Filtered trend is an extrapolation of a regression of the series against time, after
the removal of outliers. Using the median RAE as a measure

of the forecasting accuracy, the filtered trend was most accurate across all horizons. In the second exception to the
use of general time series models, Madden and Tan (2008)
revisit the traffic data set and apply a feed-forward neural
network; three versions are used, with different parameter selection procedures. Simple exponential smoothing is
also included in the analysis of these data for the first time.
Unfortunately, due to changes in the choices of the horizon and accuracy measure, it is difficult to compare the
results of the 2005 and 2008 papers. However, the neural network methods tended to produce results with accuracy measures very similar to those of simple exponential
smoothing for horizons of six days or more. Using median
absolute percentage errors, simple exponential smoothing
was most accurate for most horizons. Unfortunately, it is
not possible to tell whether these methods outperform the
filtered trend. For this traffic data set, it is clear that the data
are stationary and have very little structure, so perhaps it
is not surprising that simple exponential smoothing is the
least inaccurate forecasting method.
Ilow (2000) develops a model for forecasting Ethernet
packet data. The objective is to provide a workload model
so as to improve the network performance via dynamic
bandwidth allocation. Fractional ARIMA modelling is used.
Although the empirical results are not reported fully, the
benefit of the long memory approach is demonstrated by
the lower mean square error for FARIMA than for the linear AR model. Tzagkarakis, Papadopouli, and Tsakalides
(2009) model and forecast internet traffic on wireless local area networks. They decompose the time series using
Singular Spectrum analysis. In broad terms, the series is decomposed into signal, a non-linear deterministic trend, and
noise; the trend is extrapolated to provide the forecast.
The provision of broadband is modelled and forecast
by Mack and Grubesic (2009), using a range of econometric models with data for Ohio, US. The dependent variable
is the number of broadband providers in a ZIP code area,
where annual data for 20002001 are used to forecast from
20022004. They use the household density, population
growth, an urban/non-urban dummy, the number of businesses in the area, the median income, the area in square
miles, and a (0, 1) dummy variable for if the area is served
by large incumbent suppliers. With the exception of household density, all of these variables were found to be significant; however, they found that the inclusion of a spatial lag
term improved the forecasting accuracy considerably. The
intuition behind the spatial lag is simply that the demand
in neighbouring ZIP code areas will be similar, and thus, so
will the number of providers.
Thompson and Garbacz (2011) investigate the economic impact of mobile broadband (MB) and fixed broadband (FB). While both MB and FB have a significant effect
on GDP, they find that MB has a greater impact in higher
income countries. In a related study, Srinuan, Srinuan, and
Bohlin (2012) use survey data to investigate whether MB is
a complementary service or a substitute for FB. They conclude that, in most areas of Sweden, MB is considered as a
substitute. However, this finding may be influenced by the
fact that FB is subject to greater regulation in Sweden, and
both MB and FB are provided by the same companies.

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

3.2.3. Forecasting for other ICT products


The medium-term demand for PCs is modelled by Kurawarwala and Matsuo (1998). They compare a linear
growth model, a Bass model and a seasonal trend model
with ARIMA models (without and with seasonality). Using
monthly data for the demand for several PC products, they
find that the linear trend and seasonal trend models tend
to produce more accurate forecasts. For new products with
short time series, they use the demand for analogous products to provide initial parameter estimates; they emphasise the importance of peak sales timing and total life cycle
sales in the choice of analogy.
In a relatively early study, a hedonic pricing model for
notebook PCs was constructed by Rutherford and Wilhelm
(1999). Using regression analysis, they forecast competitive selling prices for a given configuration of notebook PC.
In addition to finding variables such as CPU speed and hard
disk capacity significant, they also found that certain manufacturers consistently either charged significant premia
or offered significant discounts.
3.3. Summary: the flow of benefits between ICT and time
series forecasting methodology
In Section 3.1, we saw that call centre applications, in
a similar way to the multi-country diffusion studies, have
produced several substantial analyses which have been
used successfully for forecasting. The availability of intraday data has led to the development of doubly seasonal
versions of ARIMA and exponential smoothing models.
Considering the studies of univariate time series modelling
and forecasting in Section 3.2, the majority are applications of established methodologies. The only example we
found where the application extended the methodology is
the robust forecasting approach of Grambsch and Stahel
(1990). The two applications of neural networks for forecasting show potential with the two or three series they
are used with, but the methods need to be tried on a far
larger set of time series in order to demonstrate that they
are broadly competitive. Thus, the answer to the question
posed at the beginning of Section 3 is that (based on the
studies discussed in Section 2.2, and apart from the exceptions mentioned) there is nothing about telecommunications and other ICT time series that sets them apart. In
other words, the fact that a series is derived from ICT activity is no guide for the choice of a forecasting methodology.
4. Technological forecasting
Technological evolution is an engine of growth that expands market potentials and creates new market leaders
(see Sood & Tellis, 2005). Technological forecasting is defined by Ascher (1978) as the attempt to project technology capabilities and to predict the inventions and spread
of technological innovations. This definition includes diffusion modelling, which we consider separately in Section 2; here, we look at studies using other approaches.
Technological forecasting is important, as it is an input to
strategic management decisions regarding where to invest
for new technologies, and when to shift investment from
the old generation of technology to the new. Technological

1121

forecasting also influences the design of new technologies.


As was mentioned in the introduction, in the context of
ICT, applications of non-diffusion technological forecasting
have received relatively little coverage compared to diffusion or time series modelling. Due to this relatively sparse
body of literature, we review the studies by application
area.
4.1. Mobile telephony
We discuss two examples of long term forecasts, the
first focussing on market size and the second focussing
on the timing of the introduction of a new generation of
technology.
Firstly, we look at forecasting the market size. Jeon,
Hyun, and Granger (2004) perform a long-term technological forecast of the Chinese mobile telephony market.
After a thorough discussion of alternative methods, they
chose forecasting by analogy. The mobile telephony markets of the USA, Japan and Korea are identified as possible
precedents, as they are very large markets and are more
developed than that of China. The appropriateness of the
analogy was measured by the time taken for a particular
growth phase (from 5% to 21%). Using this criterion, Korea
was chosen as the most appropriate precedent, and market
forecasts for China are computed on this basis.
Secondly, we look at two approaches to forecasting the
timing of the introduction of a new technology. Meade and
Islam (2003, 2010) used copulas to model the dependence
between two related products or technologies (e.g., fax machines and cellular phones) using data from 132 countries.
Given that a country has adopted one technology, this dependence relationship is then used to predict the adoption time of another technology. Kim, Daim, and Anderson
(2010) forecast the year of implementation (specifically,
the year of first commercialisation) of a new generation
of mobile telephony. They use regression with the introduction date as the dependent variable; the independent
variables are four technical variables describing the technology, chosen with the help of an expert panel. These variables are: channel bandwidth in kHz; number of channels;
channel bit rate in kbps; and data capacity in kbps. Starting in 1975, they show that the interval between generations is close to 11 years; depending on assumptions about
data capacity, their model predicted that 4G would appear
between 2012 and 2015. The actual first commercialisation occurred at their lower limit, with 4G networks being
available in South Korea and the USA in 2012. By late 2013,
80% of all 4G connections were concentrated in South Korea (where 50% of connections were 4G), Japan and the USA
(where, for both countries, 20% of connections were 4G).
4.2. The internet
Similarly, there is little model-based forecasting in the
context of the internet: instead, forecasts are based on expert opinion. We begin by discussing studies that identify
the drivers of broadband adoption. The change in focus in
these studies from economic variables to applications illustrates the rapid rate of change over the twelve years between the first study and the most recent.
Stordahl and Rand (1999) used a Delphi survey to identify and quantify the drivers of the demand for broadband. They questioned an international panel of experts

1122

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

over two rounds in 1997, and found that households were


prepared to devote up to 2% of their disposable income
to broadband capacity and applications. Numerical forecasts of demand were computed using parameters derived
from the Delphi exercise. Using a similar approach, Stordahl (2004) produced long-term forecasts for the European
broadband market, broken down by country and the category of broadband. Weber and Kauffman (2011) discuss
the drivers of ICT adoption and identify three internet application contexts in which research should be pursued:
location-based services, cloud computing, and social
networking. These application areas are already well established, but the authors see opportunities for further
development.
The second set of studies is more qualitative, looking at
possible developments in internet technology. In a group
of articles on internet prediction, Elliott (2010) discusses
a US National Science Foundation sponsored project, the
Global Environment for Network Innovations, GENI. This
ongoing project (see www.geni.net) plans to carry out experiments to investigate what are perceived to be the more
important developments in the infrastructure of the internet. These developments include content distribution services (the investigation of more scaleable architectures for
distributing high-bandwidth content such as videos and
virtual worlds in gaming); mobility (the development of
novel protocols in order to improve the support for mobile
devices in internet architecture); the robustness of internet architecture (disruption tolerant networks (DTN),
investigating architectures that are independent of the current TCP/IP architecture); and routing (concerns about the
ability of the current routing architecture to scale up to
an increasing traffic). Holzle and Barroso (2010) draw attention to other constraints on the development of internet services, and focus particularly on the development of
warehouse-scale computers that are able to deal with the
storage and transmission of large amounts of data robustly.
Odlyzko (2010) highlights the growth and predominance of mobile voice over the last decade. Looking at
mobile broadband, he sees the growth in the demand for
data services outstripping the capacity, as he expects network managers to favour voice, which is far more lucrative than data transmission. This could lead to a number
of scenarios, such as a growth in applications that exploit
high capacity fixed cable internet, or a growth in services
that are designed for low bandwidth mobile internet. However, the assumptions of this study conflict with current
(2013) anecdotal evidence (as discussed in the introduction), where data transmission is tending to be more lucrative than voice, and the development of 4G is facilitating a
greater use of mobile broadband.
The routine linking of sensors to the internet could lead
to high data flows that could swamp current data storage
(Clark, 2010). Estrin (2010) discusses participatory sensing: the collection of data via mobile phones for collection
and analysis by the subscriber or others. Applications of
this include health monitoring or the tracking of environmental sustainability. These studies are in line with the development of smart phone apps (applications) for health
monitoring.

4.3. Summary: the flow of benefits between ICT and technological forecasting methodology
The brevity of this section reflects the shortage of publications in this area. The majority of the studies are exercises based on expert opinion, which tend not to be capable
of generalisation into a broader forecasting context. An exception is the study by Jeon et al. (2004), who do provide a
very informative case study on the choice and application
of a long-term technological forecast; they forecast the development of the Chinese mobile telephony market using
the Korean market as an analogy.
5. Summary, conclusions and suggestions for further
research
The literature on forecasting applied to an ICT application goes back at least forty years (e.g., Chaddha & Chitgopekar, 1971); in those days, the applications related to
fixed line telephony and main-frame computers. The range
of applications has increased with the development of consumer technology from PCs and basic mobile phones to
tablets and smart phones. The interactions between ICT
application areas and forecasting have led to several developments in forecasting. The literature on the multigeneration diffusion model has included a range of ICT
applications; consumer technology products such as mobile phones have led to a wider use of choice models; and
the special properties of some ICT series have led to the
development of the robust trend time series forecasting
method.
Our review shows that the most activity has occurred in
the modelling of diffusion in ICT, particularly mobile telephony. The level of innovation in the single-country diffusion exercises has generally been low; often, the analysis
is simply a comparison of the logistic and Gompertz models using different national data sets. We do note that fitting the cumulative diffusion seems to work better than the
estimation of diffusion model parameters using adoptions
per period. This experience contrasts with that in the marketing literature, where adoptions per period is the dependent variable that is used most widely in the estimation
of the Bass model; however, this approach tended to be
less successful in the context of mobile telephony. The phenomenon of multiple subscriptions per person makes the
interpretation of market potentials problematic. The issue
of using models that are designed for capturing adoptions
to model and forecast subscriptions needs further investigation. In the context of modelling multiple generations
of ICT innovations, the use of the simultaneous hazard approach (see Lillard, 1993) for predicting the timing of the
next generation of technology is worth exploring.
The analysis of the multi-country diffusion of mobiles is
generally more informative in terms of identifying the relative levels of importance of several determinants of the
diffusion rate. In summary, cross-sectional differences in
the diffusion rate can be attributed to the national GNP.
Over time, the general consensus is that diffusion accelerated due to the transition from analog to digital mobile
phones and to decreases in prices. Other covariates that
have been cited as accelerating diffusion include the number of competing suppliers and the presence of fewer com-

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

peting standards. Although these analyses do not focus on


forecasting directly, their identification of the important
determinants of diffusion facilitates better-informed management decision making.
Modelling the choice between alternatives (possible
competing products or suppliers) has generated some new
approaches, for example that of Jun and Park (1999). Further experience in modelling and forecasting using this
approach is needed to enable us to assess the strengths
and weaknesses of the approach. Another choice-based
methodology that has potential for ICT applications is the
generation of time series forecasts from cross-sectional
discrete choice experiments. This approach was suggested
by Islam (2014) and Islam and Meade (2013). They used
recent developments in discrete choice experimentation
to measure household-level preferences, and established
a causal link between the attributes of the technology and
adoption time intentions using discrete time survival mixture analysis and the Bass diffusion model.
The forecasting of time series for call centres has stimulated several innovations. Examples include the use of
models incorporating multiple seasonal patterns and the
development of inhomogeneous Poisson process models.
We saw that there are several clusters of ICT data sets that
are broadly homogeneous, and thus, tend to be modelled
and forecast better by a single model (or class of models).
An early study of one of these clusters of time series led
to an innovation in forecasting methodology, namely the
robust trend of Grambsch and Stahel (1990).
As we pointed out in the introduction, the supply of academic analyses is affected by factors such as the ease of
access of relevant data and the publication criteria of journal editors. This review has covered the literature available
and only the literature available. The distribution of this
body of literature has influenced both the topics discussed
in this review and the emphasis placed upon them. It does
not necessarily follow that the distribution of academic effort has been optimal for the long term benefit either of
the ICT industry or of forecasting theory and methodology. We will try to bear this in mind in our discussion
of future research. Looking to the future, it is likely that
diffusion modelling will continue to dominate the ICT forecasting literature, due to the increasing number of applications as a result of further developments in consumer
technology. The emphasis is likely to move to higher frequency (than annual) data as the rate of diffusion increases.
These developments are also likely to reinforce the use of
choice modelling for informing product planning and production. For future time series ICT applications, the availability of higher frequency data in call centres is likely to
lead to a greater emphasis on the real-time updating of
intra-day forecasts. More generally, it is likely that ICT will
benefit from general innovations in forecasting methodologies, rather than ICT stimulating time series modelling
innovations. The technological developments in ICT are of
widespread importance and have led to the waxing and
waning of huge corporations. The use of expert opinion is
likely to be the main resource; however, the use of analogous data (as per Jeon et al., 2004) for modelling phenomena such as technological convergence deserves
consideration, and research outside ICT suggests that it
leads to improvements in accuracy (see nkal, Goodwin,
Thomson, Gnl, & Pollock, 2009).

1123

Acknowledgments
We are grateful for the comments on the initial drafts
of this paper by three anonymous referees, and to the
Social Science and Humanities Research Council (SSHRC),
Canada, grant # 430199, for research support. We have also
benefitted from many discussions with Robert Fildes.
Appendix A. Supplementary data
Supplementary material related to this article can be
found online at http://dx.doi.org/10.1016/j.ijforecast.2014.
09.003.
References
Agarwal, R., & Bayus, B. L. (2002). The market evolution and sales take-off
of product innovations. Management Science, 48, 10241041.
Andrews, B. H., & Cunningham, S. M. (1995). L.L. Bean improves callcenter forecasting. Interfaces, 25(6), 113.
Antipov, A., & Meade, N. (2002). Forecasting call frequency at a financial
services call centre. Journal of the Operational Research Society, 53,
953960.
Ascher, W. (1978). Forecasting: an appraisal for policy makers and planners.
Baltimore, MD: The Johns Hopkins University Press.
Avramidis, A. N., Deslauriers, A., & LEcuyer, P. (2004). Modeling daily
arrivals to a telephone call center. Management Science, 50, 896908.
Bagchi, K., Kirs, P., & Lopez, F. (2008). The impact of price decreases on
telephone and cell phone diffusion. Information and Management, 45,
183193.
Bartels, R., & Islam, T. (2002). Supply restricted telecommunications
markets: the effect of technical efficiency on waiting times. Journal
of Productivity Analysis, 18, 161169.
Bass, F. M. (1969). A new product growth model for consumer durables.
Management Science, 15, 215227.
Bass, F. M., Krishnan, T., & Jain, D. (1994). Why the Bass model fits without
decision variables. Marketing Science, 13, 203223.
Betts, A., Meadows, M., & Walley, P. (2000). Call centre capacity
management. International Journal of Service Industry Management,
11(2), 185196.
Bewley, R., & Fiebig, D. (1988). Flexible logistic growth model with applications in telecommunications. International Journal of Forecasting, 4,
177192.
Bianchi, L., Jarrett, J., & Hanumara, R. C. (1993). Forecasting incoming calls
to telemarketing centers. The Journal of Business Forecasting Methods
and Systems, 12, 312.
Bianchi, L., Jarrett, J., & Hanumara, R. C. (1998). Improving forecasting
for telemarketing centers by ARIMA modeling with intervention.
International Journal of Forecasting, 14, 497504.
Blackman, W. A. (1972). A mathematical model for trend forecasts.
Technological Forecasting and Social Change, 3, 441452.
Bohlin, A., Gruber, H., & Koutroumpis, P. (2010). Diffusion of new
technology generations in mobile communications. Information
Economics and Policy, 22, 5160.
Botelho, A., & Pinto, P. C. (2004). The diffusion of cellular phones in
Portugal. Telecommunications Policy, 28, 427437.
Brown, L., Gans, N., Mandelbaum, A., Sakov, A., Shen, H., Zeltyn, S., & Zhao,
L. (2005). Statistical analysis of a telephone call center: a queuingscience perspective. Journal of the American Statistical Association, 100,
3650.
Chaddha, R. L., & Chitgopekar, S. S. (1971). A generalization of the logistic
curve and long range forecasts (19661981) of residence telephones.
Bell Journal of Economics and Management Science, 2, 542560.
Chen, C., & Watanabe, C. (2006). Diffusion, substitution and competition
dynamism inside the ICT market: the case of Japan. Technological
Forecasting and Social Change, 73, 731759.
Chinn, M. D., & Fairlie, R. W. (2007). The determinants of the Global
Digital Divide: a cross-country analysis of computer and Internet
penetration. Oxford Economic Papers, 59, 1644.
Chu, W.-L., Wu, F.-S., Kao, K.-S., & Yen, D. C. (2009). Diffusion of mobile
telephony: an empirical study in Taiwan. Telecommunications Policy,
33, 506520.
Clark, D. (2010). Fighting over the future of the Internet. IEEE Internet
Computing, 10, 2223.
Comer, J. C., & Wikle, T. A. (2008). Worldwide diffusion of the cellular
telephone, 19952005. The Professional Geographer, 60, 252269.

1124

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

De Gooijer, J. G., & Hyndman, R. J. (2006). 25 years of time series


forecasting. International Journal of Forecasting, 22, 443473.
Dekimpe, M. G., Parker, P. M., & Sarvary, M. (1998). Staged estimation
to global cellular telephone adoption. Technological Forecasting and
Social Change, 57, 105132.
Dergiades, T., & Dasilas, A. (2010). Modelling and forecasting mobile
telecommunication services: the case of Greece. Applied Economics
Letters, 17, 18231828.
Doganoglu, T., & Grzybowski, L. (2007). Estimating network effects in
mobile telephony in Germany. Information Economics and Policy, 19,
6579.
Duffy-Deno, K. T. (2001). Demand for additional telephone lines: an
empirical note. Information Economics and Policy, 13, 283299.
Eisner, J., & Waldon, T. (2001). The demand for bandwidth: second
telephone lines and on-line services. Information Economics and
Policy, 13, 301309.
Elliott, J. (2010). GENI: opening up new classes of experiments in global
networking. IEEE Internet Computing, 10, 3942.
Estrin, D. (2010). Participatory sensing: applications and architecture.
IEEE Internet Computing, 10, 1214.
Fildes, R. (2002). Telecommunications demand forecastinga review.
International Journal of Forecasting, 18, 489522.
Fildes, R. (2003). Corrigendum to Telecommunications demand
forecastinga review [International Journal of Forecasting 18
(2002) 489522]. International Journal of Forecasting, 19, 337.
Fildes, R., Hibon, M., Makridakis, S., & Meade, N. (1998). Generalising
about univariate forecasting methods: further empirical evidence.
International Journal of Forecasting, 14, 339358.
Fisher, J. C., & Pry, R. H. (1971). Simple substitution model of technological
change. Technological Forecasting and Social Change, 3, 7588.
Frank, L. D. (2004). An analysis of the effect of the economic situation on
modeling and forecasting the diffusion of wireless communications
in Finland. Technological Forecasting and Social Change, 71, 391403.
Gamboa, L. F., & Otero, J. (2009). An estimation of the pattern of diffusion
of mobile phones: the case of Colombia. Telecommunications Policy,
33, 611620.
Gans, N., Koole, G., & Mandelbaum, A. (2003). Telephone call centers:
tutorial, review, and research prospects. Manufacturing and Service
Operations Management, 5, 79141.
Gardner, E. S., Jr., & Diaz-Saiz, J. (2008). Exponential smoothing in the
telecommunications data. International Journal of Forecasting, 24,
170174.
Goolsbee, A., & Klenow, P. J. (2002). Evidence on learning and network
externalities in the diffusion of home computers. Journal of Law and
Economics, 45, 317344.
Grajek, M., & Kretschmer, T. (2009). Usage and diffusion of cellular
telephony, 19982004. International Journal of Industrial Organization,
27, 238249.
Grambsch, P., & Stahel, W. A. (1990). Forecasting demand for special
telephone servicesa case study. International Journal of Forecasting,
6, 5364.
Gruber, H., & Koutroumpis, P. (2013). Competition enhancing regulation
and diffusion of innovation: the case of broadband networks. Journal
of Regulatory Economics, 43, 168195.
Gruber, H., & Verboven, F. (2001). The diffusion of mobile telecommunication services in the European Union. European Economic Review, 45,
577588.
Gurbaxani, V., & Mendelson, H. (1990). An integrative model of IS
spending growth. Information Systems Research, 1, 2346.
Hamoudia, M., & Islam, T. (2004). Modelling and forecasting the growth
of wireless messaging. Telektronikk, 100, 6469.
Hilas, C. S., Goudos, S. K., & Sahalos, J. N. (2006). Seasonal decomposition
and forecasting of telecommunication data: a comparative case
study. Technological Forecasting and Social Change, 73, 495509.
Holzle, U., & Barroso, L. A. (2010). Warehouse-scale computers. IEEE
Internet Computing, 10, 3335.
Hwang, J., Cho, Y., & Long, N. V. (2009). Investigation of factors affecting
the diffusion of mobile telephone services: an empirical analysis for
Vietnam. Telecommunications Policy, 33, 534543.
Ida, T., Kinoshita, S., & Sato, M. (2008). Conjoint analysis of demand for IP
telephony: the case of Japan. Applied Economics, 40, 12791287.
Ida, T., & Kuroda, T. (2006). Discrete choice analysis of demand for
broadband in Japan. Journal of Regulatory Economics, 29, 522.
Ilow, J. (2000). Forecasting network traffic using FARIMA models with
heavy tailed innovations. In Proceedings of the acoustics, speech, and
signal processing IEEE international conference. Vol. 6 (pp. 38143817).
Islam, T. (2014). Household level innovation diffusion model of photovoltaic (PV) solar cells from stated preference data. Energy Policy, 65,
340350.
Islam, T., & Fiebig, D. (2001). Modelling the development of supply
restricted telecommunications markets. Journal of Forecasting, 20,
249264.

Islam, T., Fiebig, D. G., & Meade, N. (2002). Modelling multinational


telecommunications demand with limited data. International Journal
of Forecasting, 18, 605624.
Islam, T., & Meade, N. (1996). Forecasting the development of the market
for business telephones in the UK. Journal of Operational Research
Society, 47, 906918.
Islam, T., & Meade, N. (1997). The diffusion of successive generations of
a technology: a more general model. Technological Forecasting and
Social Change, 56, 4960.
Islam, T., & Meade, N. (2011). Detecting the impact of market factors on
sales take-off times of analog cellular telephones. Marketing Letters,
22, 197212.
Islam, T., & Meade, N. (2012). The impact of competition, and economic
globalization on the multinational diffusion of 3G mobile phones.
Technological Forecasting and Social Change, 79, 843850.
Islam, T., & Meade, N. (2013). The impact of attribute preferences
on adoption timing: the case of photo-voltaic (PV) solar cells for
household electricity generation. Energy Policy, 55, 521530.
Jang, S. L., Dai, S.-C., & Sung, S. (2005). The pattern and externality effect
of diffusion of mobile telecommunications: the case of OECD and
Taiwan. Information Economics and Policy, 17, 133148.
Jeffres, L., & Atkin, D. (1996). Predicting use of technologies for
communication and consumer needs. Journal of Broadcasting and
Electronic Media, 40, 318330.
Jeon, Y., Hyun, K. R., & Granger, C. J. (2004). Long-term technological
forecasting. Telektronikk, 100, 312.
Jeon, H. J., Kim, M. S., & Sohn, S. Y. (2010). Conjoint and WTP analyses
of future mobile phones for digital convergence. Technological
Forecasting and Social Change, 77, 457465.
Jun, D. B., Kim, S. K., Park, M. H., Bae, M. S., Park, Y. S., & Joo, Y. J. (2000).
Forecasting demand for low earth orbit mobile satellite service in
Korea. Telecommunications Systems, 14, 311319.
Jun, D. B., Kim, S. K., Park, Y. S., Park, M. H., & Wilson, A. R. (2002).
Forecasting telecommunication service subscribers in substitutive
and competitive environments. International Journal of Forecasting,
18, 561581.
Jun, D. B., & Park, Y. S. (1999). A choice-based diffusion model for multiple
generations of products. Technological Forecasting and Social Change,
61, 4558.
Kiiski, S., & Pohjola, M. (2002). Cross-country diffusion of the Internet.
Information Economics and Policy, 14, 287310.
Kim, J., Daim, T., & Anderson, T. (2010). A look into the future of
wireless mobile communication technologies. Technology Analysis
and Strategic Management, 22, 925943.
Klungle, R., & Maluchnik, J. (1997). Call center forecasting at AAA
Michigan. The Journal of Business Forecasting Methods and Systems,
16(4), 813.
Koole, G. (2008). Introduction to the special issue on call center
management. Management Science, 54, 237.
Koski, H., & Kretschmer, T. (2005). Entry, standards and competition: firm
strategies and the diffusion of mobile telephony. Review of Industrial
Organization, 26, 89113.
Kreng, V. B., & Wang, H. T. (2009). A technology replacement model with
variable market potentialan empirical study of CRT and LCD TV.
Technological Forecasting and Social Change, 76, 942951.
Kumar, V., Nagpal, A., & Venkatesan, R. (2002). Forecasting category sales
and market share for wireless telephone subscribers: a combined
approach. International Journal of Forecasting, 18, 583603.
Kurawarwala, A. A., & Matsuo, H. (1998). Product growth models for
medium-term forecasting of short life cycle products. Technological
Forecasting and Social Change, 57, 169196.
Kwak, J. H., & Lee, B. G. (2011). Estimating demand curve in the Korean
VoIP telecommunications market. Technological Forecasting and Social
Change, 78, 713728.
Lee, M., & Cho, Y. (2007). The diffusion of mobile telecommunications
services in Korea. Applied Economics Letters, 14, 477481.
Lee, J., Cho, Y., Lee, J.-D., & Lee, C.-Y. (2006). Forecasting future demand
for large-screen television sets using conjoint analysis with diffusion
model. Technological Forecasting and Social Change, 73, 362376.
Lee, C.-Y., Lee, J.-D., & Kim, Y. (2008). Demand forecasting for new
technology with a short history in a competitive environment; the
case of the home networking market in South Korea. Technological
Forecasting and Social Change, 75, 91106.
Lee, J. C., & Lu, K. W. (1987). On a family of data based transformed
models useful in forecasting technological substitution. Technological
Forecasting and Social Change, 31, 6178.
Lee, J. C., Lu, K. W., & Horng, S. C. (1992). Technological forecasting with
non-linear models. Journal of Forecasting, 11, 195206.
Liikanen, J., Stoneman, P., & Toivanen, O. (2004). Intergenerational effects
in the diffusion of new technology: the case of mobile phones.
International Journal of Industrial Organization, 22, 11371154.

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126


Lillard, L. A. (1993). Simultaneous equations for hazards. Journal of
Econometrics, 56, 189217.
Lim, B.-L., Choi, M., & Park, M.-C. (2003). The late take-off phenomenon
in the diffusion of telecommunications services: network effect and
critical mass. Information Economics and Policy, 15, 537667.
Mabert, V. A. (1985). Short interval forecasting of emergency phone call
(911) work loads. Journal of Operations Management, 5, 259271.
Mack, E. A., & Grubesic, T. H. (2009). Forecasting broadband provision.
Information Economics and Policy, 21, 297311.
Madden, G., & Coble-Neal, G. (2004a). Economic determinants of global
mobile telephony growth. Information Economics and Policy, 16,
519534.
Madden, G., & Coble-Neal, G. (2004b). Internet traffic dynamics.
Telektronikk, 4, 168179.
Madden, G., & Coble-Neal, G. (2005). Forecasting international bandwidth
capability. Journal of Forecasting, 24, 299309.
Madden, G., Savage, S. J., & Coble-Neal, G. (2002). Forecasting United
States international message telephone service. International Journal
of Forecasting, 18, 523544.
Madden, G., & Tan, J. (2007). Forecasting telecommunications data with
linear models. Telecommunications Policy, 31, 3144.
Madden, G., & Tan, J. (2008). Forecasting international bandwidth
capacity using linear and ANN methods. Applied Economics, 40,
17751787.
Mahajan, V., & Muller, E. (1996). Timing, diffusion, and substitution
of successive generations of technological innovations: the IBM
mainframe case. Technological Forecasting and Social Change, 51,
109132.
Makridakis, S., & Hibon, M. (2000). The M3-competition: results,
conclusions and implications. International Journal of Forecasting, 16,
451476.
Mandelbaum, A. (2003). Call centres: research bibliography with abstracts.
Technical Report. http://iew3.technion.ac.il/serveng/References/
ccbib.pdf.
Mar-Molinero, C. (1980). Tractors in Spaina logistic analysis. Journal of
the Operational Research Society, 31, 141152.
Massini, S. (2004). The diffusion of mobile telephony in Italy and the
UK: an empirical investigation. Economics of Innovation and New
Technology, 13, 251277.
Mastorocostas, P., & Hilas, C. (2012). A computational intelligence-based
forecasting system for telecommunications time series. Engineering
Applications of Artificial Intelligence, 25, 200206.
Meade, N. (2000). A note on the robust trend and ARARMA methodologies
used in the M3 competition. International Journal of Forecasting, 16,
517519.
Meade, N., & Islam, T. (1995). Prediction intervals for growth curve
forecasts. Journal of Forecasting, 14, 413430.
Meade, N., & Islam, T. (2001). Forecasting the diffusion of innovations:
implications for time series extrapolation. In J. S. Armstrong (Ed.),
Principles of forecasting: a handbook for researchers and practitioners
(pp. 577595). Norwell, MA: Kluwer Academic Publishers.
Meade, N., & Islam, T. (2003). Modeling the dependence between
the times to international adoption of two related technologies.
Technological Forecasting and Social Change, 70, 759778.
Meade, N., & Islam, T. (2006). Modelling and forecasting the diffusion of
innovationa 25-year review. International Journal of Forecasting, 22,
519545.
Meade, N., & Islam, T. (2010). Using copulas to model repeat purchase
behaviouran exploratory analysis via a case study. European Journal
of Operational Research, 200(3), 908917.
Michalakelis, C., Varoutas, D., & Sphicopoulos, T. (2008). Diffusion
models of mobile telephony in Greece. Telecommunications Policy, 32,
234245.
Nam, C., Kim, S., & Lee, H. (2008). The role of WiBro: filling the gaps in
mobile broadband technologies. Technological Forecasting and Social
Change, 75, 438448.
Norton, J. A., & Bass, F. M. (1987). A diffusion theory model of adoption and
substitution for successive generations of high-technology products.
Management Science, 33, 10691086.
Odlyzko, A. (2010). The Internet and past and future communications
revolutions. IEEE Internet Computing, 10, 2021.
OECD (2011). OECD science, technology and industry scoreboard 2011.
OECD Publishing.
nkal, D., Goodwin, P., Thomson, M., Gnl, M. S., & Pollock, A. (2009).
The relative influence of advice from human experts and statistical
methods on forecast adjustments. Journal of Behavioral Decision
Making, 22, 390409.
Rai, A., Ravichandran, T., & Samaddar, S. (1998). How to anticipate the
Internets global diffusion. Communications of the ACM, 41, 97106.
Robertson, A., Soopramaniem, D., & Fildes, R. (2007). Segmental newproduct diffusion of residential broadband services. Telecommunications Policy, 31, 265275.

1125

Rogers, E. M. (2003). Diffusion of innovations (5th ed.). New York: The Free
Press.
Rouvinen, P. (2006). Diffusion of digital mobile telephony: are developing
countries different? Telecommunications Policy, 30, 4663.
Rutherford, D. P., & Wilhelm, W. E. (1999). Forecasting notebook
computer price as a function of constituent features. Computers and
Industrial Engineering, 37, 823845.
Sharif, M. N., & Kabir, C. (1976). System dynamics modeling for forecasting multilevel technological substitution. Technological Forecasting and Social Change, 9, 89112.
Shen, H., & Huang, J. Z. (2005). Analysis of call centre arrival data using
singular value decomposition. Applied Stochastic Models in Business
and Industry, 21, 251263.
Shen, H., & Huang, J. Z. (2008a). Interday forecasting and intraday
updating of call center arrivals. Manufacturing and Service Operations
Management, 10, 391410.
Shen, H., & Huang, J. Z. (2008b). Forecasting time series of inhomogeneous
Poisson processes with application to call center workforce management. The Annals of Applied Statistics, 2, 601623.
Shin, D. H., & Bartolacci, M. (2007). A study of MVNO diffusion and market
structure in the EU, US, Hong Kong and Singapore. Telematics and
Informatics, 24, 86100.
Singh, S. K. (2008). The diffusion of mobile phone in India. Telecommunications Policy, 32, 642651.
Sood, A., & Tellis, G. J. (2005). Technological evolution and radical
innovation. Journal of Marketing, 69, 152168.
Soyer, R., & Tarimcilar, M. (2008). Modeling and analysis of call center
arrival data: a Bayesian approach. Management Science, 54, 266278.
Sparrow, L. B. (1991). Manning the telephone enquiry bureau at British
Gas West Midland. In S. C. Littlechild (Ed.), Operations research in
management (pp. 167173). New York: Prentice Hall.
Srinuan, P., Srinuan, C., & Bohlin, E. (2012). Fixed and mobile broadband
substitution in Sweden. Telecommunications Policy, 36, 237251.
Stordahl, K. (2004). Long-term broadband technology forecasting.
Telektronikk, 100, 1331.
Stordahl, K., & Rand, L. (1999). Long term forecasts for broadband
demand. Telektronikk, 95, 3444.
Sundqvist, S., Frank, L., & Puumalainen, K. (2005). The effects of country
characteristics, cultural similarity and adoption timing on the
diffusion of wireless communications. Journal of Business Research, 58,
107110.
Sung, N., & Lee, Y.-H. (2002). Substitution between mobile and fixed
telephones in Korea. Review of Industrial Organization, 20, 367374.
Tandberg, D. (1995). Time series forecasts of poison center call volume.
Clinical Toxicology, 33, 1118.
Taylor, J. W. (2008). A comparison of univariate time series methods for
forecasting intraday arrivals at a call center. Management Science, 54,
253265.
Taylor, J. W. (2012). Density forecasting of intraday call center arrivals
using models based on exponential smoothing. Management Science,
58, 543549.
Tellis, J. T., Stremersch, S., & Yin, E. (2003). The international takeoff
of new products: the role of economics, culture, and country
innovativeness. Marketing Science, 22, 188208.
Thompson, H. G., Jr., & Garbacz, C. (2011). Economic impacts of mobile
versus fixed broadband. Telecommunications Policy, 35, 9991009.
Tych, W., Pedregal, D. J., Young, P. C., & Davies, J. (2002). An unobserved
component model for multi-rate forecasting of telephone call
demand: the design of a forecasting support system. International
Journal of Forecasting, 18, 673695.
Tzagkarakis, G., Papadopouli, M., & Tsakalides, P. (2009). Trend forecasting based on singular spectrum analysis of traffic workload in a largescale wireless LAN. Performance Evaluation, 66, 173190.
Weber, M., & Kauffman, R. J. (2011). What drives ICT adoption?
Analysis and research directions. Electronic Commerce Research and
Applications, 10, 683701.
Weinberg, J., Brown, L. D., & Stroud, J. R. (2007). Bayesian forecasting of an
inhomogeneous Poisson process with applications to call center data.
Journal of the American Statistical Association, 102(480), 11851198.
Wiorkowski, J. J., & Gylys, V. A. (2006). An empirical real-time test for
takeoff with applications to cellular telephony. Review of Marketing
Science, 4, 141.
Wu, F.-S., & Chu, W.-L. (2010). Diffusion models of mobile telephony.
Journal of Business Research, 63, 497501.
Xu, W. (1999). Long range planning for call centers at FedEx. The Journal
of Business Forecasting Methods and Systems, 18(4), 711.
Yang, Y., & Williams, E. (2009). Logistic model-based forecast of sales and
generation of obsolete computers in the US. Technological Forecasting
and Social Change, 76, 11051114.

1126

N. Meade, T. Islam / International Journal of Forecasting 31 (2015) 11051126

Nigel Meade is Professor of Management Science, Imperial College


Business School, London, UK. He is an experienced statistical modeler
with a background in operational research and statistics applied to
innovation diffusion and finance. He is an associate editor of the
International Journal of Forecasting and the European Journal of Finance. He
has published over 50 papers and eight book chapters, and successfully
supervised more than 20 Ph.D. students.

Towhidul Islam is Professor of Marketing and Consumer Studies, College


of Management and Economics (CME), and CME Fellow in consumer
insights, consumer well-being and public policy, University of Guelph. His
Ph.D. is in management science from Imperial College Business School,
London, UK. His main research interests are in the areas of innovation
diffusion, consumer choice and choice models. His work has appeared in
the Journal of Consumer Research, Management Science, Journal of Consumer
Psychology, International Journal of Research in Marketing, International
Journal of Forecasting, Journal of Forecasting, and the European Journal of
Operational Research, among other outlets.

You might also like