ARTICLE IN PRESS

Tourism Management 29 (2008) 19–31

www.elsevier.com/locate/tourman

Modelling and forecasting Australian domestic tourism

George Athanasopoulos, Rob J. Hyndman
Department of Econometrics and Business Statistics, Monash University, Melbourne, Vic. 3800, Australia
Received 18 October 2006; received in revised form 18 April 2007; accepted 23 April 2007

Abstract

In this paper, we model and forecast Australian domestic tourism demand. We use a regression framework to estimate important
economic relationships for domestic tourism demand. We also identify the impact of world events such as the 2000 Sydney Olympics and
the 2002 Bali bombings on Australian domestic tourism. To explore the time series nature of the data, we use innovations state space
models to forecast domestic tourism demand. Combining these two frameworks, we build innovations state space models with exogenous
variables. These models are able to capture the time series dynamics in the data, as well as economic and other relationships. We show
that these models outperform alternative approaches for short-term forecasting and also produce sensible long-term forecasts. The
forecasts are compared with the official Australian government forecasts, which are found to be more optimistic than our forecasts.
r 2007 Elsevier Ltd. All rights reserved.

Keywords: Australia; Domestic tourism; Exponential smoothing; Forecasting; Innovations state space models

1. Introduction more likely than an international tourist to visit regional

and remote areas of Australia that are not internationally
The Australian tourism industry can be divided into promoted. Thus, Australian domestic tourism is an
three major segments: (i) international inbound tourism; important topic that is in need of careful study and
(ii) domestic tourism; and (iii) outbound tourism. Of these, analysis.
domestic tourism is the largest financial contributor to the To the best of our knowledge, the only forecasts
Australian economy. In 2005, domestic tourism contrib- available for the Australian domestic tourism market are
uted an estimated $55.5 billion to the Australian economy, those produced by the Tourism Forecasting Committee
more than three times the contribution of international (TFC) and published by Tourism Research Australia
arrivals (Tourism Forecasting Committee, 2005). Despite (TRA). TRA is a business unit under the umbrella of
this, the main focus of Australian academic tourism Tourism Australia, which is an Australian federal govern-
research has been on international tourism (see, for ment statutory authority. Tourism Australia was estab-
example, Kulendran & King, 1997; Kulendran & Witt, lished after the initiatives of the Tourism White Paper
2003; Lim & McAleer, 2001a, 2002; Morley, 1998; Morris, (2003) which intended to strengthen the tourism industry.
Wilson, & Bakalis, 1995); worldwide tourism research has Following the white paper, the TFC was also established
had a similar focus (see Li, Song, & Witt, 2005 for a as an independent body. The committee comprises experts
comprehensive survey). from both the private and government sectors in the
Domestic tourism also plays a significant role in tourism and finance industry. Current members are:
maintaining and improving tourism infrastructure, espe- Tourism Australia, Australian Standing Committee on
cially in regional Australia. An Australian tourist is much Tourism, Australian Tourism Export Council, Department
of Industry Tourism and Resources, Australian Bankers
Corresponding author. Association, Tourism and Transport Forum Australia,
E-mail addresses: [email protected]
Property Council of Australia (representing major property
(G. Athanasopoulos), [email protected] investors), Qantas and Queensland Tourism Industry
(R.J. Hyndman). Council. The TFC produces consensus forecasts for

0261-5177/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tourman.2007.04.009
 ARTICLE IN PRESS
20 G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31

international, domestic and outbound tourism activity is collected by computer-assisted telephone interviews from
which are published by TRA. The forecasts produced by approximately 120,000 Australians aged 15 years and over
the TFC in October 2005 for the third quarter of 2005 and on an annual basis (Tourism Research Australia, 2005).
beyond show steady growth in the domestic market. In We use the number of visitor nights ðVNÞ as the indicator
contrast, our forecasts show that the Australian domestic of tourism activity. We disaggregate the data based on the
market is in decline, and it seems that it will remain this main purpose of travel: holiday ðHolÞ, visiting friends and
way at least in the short-term. relatives ðVFRÞ, business ðBusÞ and other ðOthÞ. The
We develop three different statistical models for fore- available sample spans from the first quarter of 1998 to
casting Australian domestic tourism. First, to help in the second quarter of 2005. Hence, there are a total of
understanding and capturing some of the economic n ¼ 30 quarterly observations (see Fig. 1).
relationships important to the domestic tourism market, Fig. 2 shows the total number of visitor nights (the
we construct a regression model of tourism demand. This aggregate of the series in Fig. 1). Also shown are the
modelling framework identifies some useful economic forecasts produced for this series by the TFC in Tourism
relationships and significant events influencing the Forecasting Committee (2005). The annual TFC forecasts
Australian domestic tourism market. However, as these show a steady average growth of 0.9% per annum from
models are static, they are unable to capture the dynamic 2006 to 2014. In contrast, there is no noticeable trend in the
properties of the data. historical data from 1998 to 2005.
The second approach adopted is to use pure time series
models to capture these dynamics. The models employed
3. Statistical models
are single source of error (or innovations) state space
models (see Aoki, 1987; de Silva, Hyndman, & Snyder,
3.1. Regression models
2006; Hannan & Deistler, 1988; Ord, Koehler, & Snyder,
1997; Snyder, 1985). These have been extremely successful
The proposed tourism demand function is
when applied to data from forecasting competitions (e.g.,
Hyndman, Koehler, Snyder, & Grose, 2002; Makridakis VN it ¼ f ðt; DEBT t ; DPI t ; GDPt ; BALI t ; OLYMPt ; MARt ,
et al., 1982; Makridakis & Hibon, 2000), and have JUN t ; SEPt ; et Þ, ð1Þ
numerous advantages over the more common form multi-
ple source of error structural time series models (STSM) (as where i ¼ fHol; VFR; Bus; Othg; VN it is the number of
outlined in Ord, Snyder, Koehler, Hyndman, & Leeds, visitor nights per capita travelling for purpose i; t the time
2005). However, they have never previously been applied to period; DEBT t is real personal debt (by all lenders) per
tourism data. capita included as a proxy to consumer confidence; DPI t
The third modelling approach we take is to include the price index for domestic holiday travel and accom-
exogenous variables in the innovations state space models. modation reflecting the price movement of domestic travel;
These models combine the advantages of each of the above GDPt the real gross domestic product per capita included
modelling frameworks. Hence, they capture the significant as an income variable; BALI t is a dummy variable
economic relationships and events identified in the regres- capturing the effect of the bombings in Bali (BALI t ¼ 1
sion models, and combine these with the time series during the fourth quarter of 2002 and beyond), OLYMPt is
properties of the innovations state space models. Although a dummy variable capturing the effect of the Sydney 2000
other formulations of state space models with exogenous Olympic games (OLYMPt ¼ 1 in the fourth quarter of
variables exist (see Harvey, 1990), this is the first time that 2000, which is the quarter following the games, and 0
the innovations state space formulation with exogenous otherwise), MARt , JUN t and SEPt are seasonal dummy
variables has been published. A two-step estimation variables, and et is a random error term. Full descrip-
procedure is proposed. The estimated models produce tions of the data, the data sources and projections of
accurate short-term forecasts and sensible long-term the regressors used for forecasting are provided in
forecasts. Appendix A.
The data are introduced in Section 2. Section 3 describes We considered many other economic variables of interest
the three models and compares them based on within- including petrol prices, prices of competing goods (e.g., car
sample fits and out-of-sample forecast performance. The sales), audio equipment and others. However, due to the
out-of-sample forecast evaluation also includes forecasts small sample size and the lack of variation in many of these
produced by the TFC. Section 4 presents and analyses the variables, they were found to be statistically insignificant.
long-run forecasts from the three models and those from The small sample size also prevented us testing for non-
the TFC. We summarize our conclusions in Section 5. linear relationships such as threshold effects for the
variables of interest.
2. Data Visual inspection of the dependent variables (see
Appendix B for plots of the per capita data) suggests that
The Australian domestic tourism data were obtained stationarity tests are required. Table 1 presents the results
from the National Visitor Survey, managed by TRA. Data from three unit root tests: the augmented Dickey Fuller
 ARTICLE IN PRESS
G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31 21

Holiday VFR

45000 30000
28000
40000
26000
35000 24000
22000
30000
20000

1998 2000 2002 2004 1998 2000 2002 2004

Business Other

7000
13000
6000
12000
5000
11000
4000
10000
3000
9000
1998 2000 2002 2004 1998 2000 2002 2004

Fig. 1. Quarterly observations for Australian domestic tourism: visitor nights ðVNÞ.

(2001). Ng and Perron (2001) consider modified versions of

VN Sample TFC forecasts four existing unit root tests. Their Monte-Carlo simulation
90000 results indicate that these tests have superior size and
power to any existing test. The results presented here are
robust to the choice of the test. In general, the hypothesis
80000 testing framework for unit roots can be simply presented
by considering an autoregressive process such as
70000 yt ¼ ryt1 þ z0t d þ et , (2)
where zt is a set of exogenous regressors which may contain
60000
a constant and a time trend. In the ADF and MZa tests, the
null hypothesis is that the series contains a unit root (i.e.,
2000 2005 2010 2015
H0 : jrj ¼ 1 versus H1 : jrjo1). The KPSS tests the null
Fig. 2. Linear trend lines fitted to the sample of total visitor nights and the hypothesis that the series does not contain a unit root (i.e.,
TFC forecasts. H0 : jrjo1 versus H1 : r ¼ 1). Before applying these tests,
each series was seasonally adjusted (see Fig. 3) using an
Table 1 additive moving average method (e.g., Makridakis, Wheel-
Unit root tests wright, & Hyndman, 1998). The test results presented in
Seasonally adjusted series ADF KPSS MZa
Table 1 indicate that none of the four series contains a unit
root.
a a
ln VN Hol
t
5.44 0.13 14.88b However, at least two of the response variables are
ln VN VFR
t
2.47 0.22a 10.98a clearly trending (this is especially apparent in Fig. 3 where
ln VN Bus
t
4.97a 0.10a 16.47b the seasonal variation has been removed). Consequently, a
ln VN Oth
t
6.43a 0.14a 8.14a deterministic trend and the growth rates of the explanatory
The data generating processes assumed for ln VN Hol and ln VN Bus contain
variables are employed in the regression model; the growth
t t
both an intercept and a deterministic trend. The data generating processes rates are calculated as ð100  D lnðZÞÞ for variable Z. We
assumed for ln VN VFR
t and ln VN Oth
t contain only an intercept. include up to one lag of each regressor. Given the quarterly
a
The test finds no unit root at the 5% level of significance. frequency of the data, we have also tested for lags two,
b
The test finds no unit root at the 10% level of significance. three and four of each regressor (one lag of each regressor
at a time due to the small sample size) but these were not
(ADF) test (Dickey & Fuller, 1979, 1981), the KPSS test found to be significant. After eliminating variables found
(Kwiatkowski, Phillips, Schmidt, & Shin, 1992) and the to be statistically insignificant at a 10% level of significance
Modified Phillips–Perron ðMZa Þ test due to Ng and Perron across all four equations, the general demand model
 ARTICLE IN PRESS
22 G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31

Holiday VFR

7.15
7.50

7.05
7.40

7.30 6.95
1998 2000 2002 2004 1998 2000 2002 2004

Business Other

6.45 5.7

5.6
6.35
5.5
6.25 5.4

5.3
6.15

1998 2000 2002 2004 1998 2000 2002 2004

Fig. 3. Seasonally adjusted natural logarithms of visitor nights per capita, i.e., seasonally adjusted ln VN it .

employed is The estimated coefficient for the exponential trend is

negative and statistically significant in the ‘‘Holiday’’ and
1000 ln VN it ¼ c þ dt þ b1 Dt1 þ b2 Pt1 þ b3 Y t ‘‘Business’’ equations. Hence, these results show a sig-
þ b4 BALI t þ b5 OLYMPt þ s1 MARt nificant long-term decline in visitor nights per capita where
þ s2 JUN t þ s3 SEPt þ et , ð3Þ travel is for holiday and business purposes.
The estimated coefficient of the lag of the growth of
where Dt1 is the lag of the growth rate of DEBT t ; Pt1 the DEBT is positive and statistically significant in the ‘‘Holi-
lag of the growth rate of DPI t ; Y t the growth rate of GDPt ; day’’ and ‘‘Business’’ equations. This variable can be
and the remaining variables are as described in Eq. (1). The considered a proxy for consumer confidence. An increase in
response variable was scaled up by 1000 to avoid very small the growth rate of borrowing in the last quarter (i.e., a
estimated coefficients. rapid growth in consumer confidence in the previous
Table 2 shows the demand equations estimated one at a quarter), results in an increase in domestic travel for both
time using ordinary least squares (OLS). In this setting, the ‘‘Holiday’’ and ‘‘Business’’ purposes.
OLS estimator is equivalent to the generalized least squares The lag of the growth in the domestic travel price index,
(GLS) estimator employed in a seemingly unrelated DPI, has mixed results across equations. In the ‘‘Holiday’’
regression (SUR) framework, as the equations contain equation, the coefficient of DPI is negative and statistically
identical regressors (refer to Green, 2000, p. 616, for significant. This suggests that as prices for domestic travel
proof). The equations show a satisfactory fit with R2 values grew faster (or declined more slowly) in the previous
ranging from 0.82 for the ‘‘Other’’ variable to 0.98 for the quarter, domestic holiday travel decreased. However, the
‘‘Holiday’’ variable. The diagnostic tests at the bottom of coefficient in the ‘‘Business’’ equation is positive and
the table show that the residuals of each equation satisfy statistically significant. This may suggest that domestic
the basic OLS assumptions of no serial correlation, travel prices have been driven up by an increase in
homoscedasticity and normality. Also the RESET test economic activity. An increase in economic activity should
results indicate that the model is appropriately specified. also result in an increase in domestic travel for business
We simplified the models by eliminating insignificant purposes, leading to the positive relationship between the
variables one at a time, selecting the coefficient with the lag of the growth rate of DPI and business travel.
highest p-value among all insignificant coefficients at the The coefficient of GDP growth was found to be negative
5% level of significance. Then, because we expect that the across all equations, but was only statistically significant in
errors across equations will be contemporaneously corre- the ‘‘Holiday’’ equation. The negative coefficient suggests
lated and because not all equations include the same that an increase in the growth of GDP results in a
regressors, the system was estimated efficiently using the significant decline in visitor nights for holiday purposes,
SUR estimation method (Zellner, 1963). The estimation and vice versa. Perhaps the explanation for this is that
results are presented in Table 3. during periods of increasing economic activity, domestic
 ARTICLE IN PRESS
G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31 23

Table 2 Table 3
Estimated coefficients of the unrestricted demand model of Eq. (3) Estimated coefficients of the demand model of Eq. (3) after eliminating
insignificant parameters
Regressor Holiday VFR Business Other
Regressor Holiday VFR Business Other
Intercept 7493.33a 7041.23a 6472.51a 5823.92a
a a a
(16.71) (34.61) (30.39) (80.78) Intercept 7505.57 7020.25 6441.09 5771.92a
t 5.14a 2.21 8.04a 6.04 (13.33) (21.03) (22.84) (47.28)
(0.91) (1.88) (1.65) (4.38) t 5.91a 6.17a
Dt1 4.34a 4.84b 7.88a 3.88 (0.50) (0.88)
(1.30) (2.70) (2.37) (6.29) Dt1 4.41a 5.91a
Pt1 5.45a 1.81 8.04a 5.89 (1.23) (2.00)
(1.88) (3.89) (3.42) (9.08) Pt1 4.11a 7.58a
Yt 39.09a 14.80 21.10 67.23 (1.64) (2.89)
(9.61) (19.91) (17.48) (46.48) Yt 43.71a
BALI t 14.20 108.02a 35.70 168.23a (8.14)
(16.31) (33.80) (29.67) (78.88) BALI t 56.61a
OLYMPt 31.60 26.78 118.17b 292.65 (17.75)
(38.99) (80.77) (70.92) (188.52) OLYMPt 148.00a
MARt 342.68a 156.95a 180.97a 563.63a (51.26)
(13.40) (27.76) (24.37) (64.79) MARt 338.09a 170.33a 170.83a 540.23a
JUN t 39.62a 56.84a 40.34 510.12a (13.06) (26.87) (24.28) (64.74)
(13.52) (28.01) (24.59) (65.37) JUN t 43.19a 71.36a 42.57b 460.75a
SEPt 31.13a 48.22 46.07b 141.28a (12.40) (26.87) (24.51) (64.74)
(14.35) (29.74) (26.11) (69.41) SEPt 27.78b 33.73 55.03a 109.13
2
(14.01) (27.84) (25.57) (66.86)
R 0.98 0.82 0.88 0.82
R̄
2 0.97 0.75 0.82 0.73 R2 0.98 0.79 0.86 0.77
2 0.98 0.75 0.82 0.74
QN c 0.23 0.40 0.57 1.07 R̄
QSC lags d 1.122 2.092 4.972 9.613e
QHT f 11.27 15.65 15.34 7.71 Standard errors are shown in parentheses beneath each coefficient.
a
QRRg 0.51 1.19 1.76 1.76 Significant at the 5% level of significance.
b
Significant at the 10% level of significance.
Standard errors are shown in parentheses beneath each coefficient.
a
Significant at the 5% level.
b
Significant at the 10% level. highest holiday and VFR travel, but the lowest business
c
Jarque and Bera (1980) w2 test for normality.
d travel. This is the summer quarter for Australia and
Breusch (1978) and Godfrey (1978) Lagrange multiplier w2 test for
serial correlation. includes the longest period of school holidays. The
e
These residuals showed some weak third order serial correlation which lowest level of holiday and VFR travel is found in the
was ignored. June quarter which includes the first semester of all levels
f
White (1980) w2 test for heteroscedasticity. of schooling.
g
Ramsey (1969) RESET w2 test for misspecification.

3.2. Exponential smoothing via innovations state space

models
holiday travel decreases significantly as Australians choose
to travel to overseas destinations instead. Tourism Fore-
Exponential smoothing was proposed in the late 1950s
casting Committee (2005) shows that for Australian
(see the pioneering works of Brown, 1959; Holt, 1957;
residents, short-term departures to overseas destinations
Winters, 1960) and has motivated some of the most
have grown at approximately 4% per annum on average
successful forecasting methods. Forecasts produced using
between 1995 and 2004.
exponential smoothing methods are weighted averages of
The dummy variable for the Sydney Olympics captures a
past observations, with the weights decaying exponentially
positive and statistically significant increase of business
as the observations get older. In other words, the more
travel in the December quarter of 2000. This is the quarter
recent the observation the higher the associated weight.
following the Sydney Olympic games, which suggests that
For example, consider the simple (or single) exponential
business travel was put on hold for the Olympic games and
smoothing method (Brown, 1959). The recursive formulae
took place immediately afterwards.
for computing a h-step-ahead forecast y^ tþh , using all data
The dummy variable for the Bali 2002 bombings
up to time t, are
captures a positive and statistically significant mean shift
in the VFR series. After the Bali bombings, Australians ‘t ¼ ayt þ ð1  aÞ‘t1 , ð4Þ
reverted to visiting friends and relatives more than before. y^ tþhjt ¼ ‘t , ð5Þ
The coefficient estimates for the seasonal dummy
variables are consistent with what would be expected for where ‘t denotes the level of the series at time t and a is a
Australian domestic travel. The March quarter has the smoothing coefficient usually 0oao1. By repeatedly
 ARTICLE IN PRESS
24 G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31

substituting for ‘t1 in (4) we get theme of these papers is that exponential smoothing
methods perform quite well although the choice of
X
t1
y^ tþh ¼ ð1  aÞt ‘0 þ a ð1  aÞi yti . (6) exponential smoothing methods is very subjective (i.e.,
i¼0 not based on any proper selection criteria) and in some
cases we believe that the selected method is not the most
Hence, the forecasted value y^ tþh is a weighted average of all
appropriate.
observations (for which the associated weights decrease
A more common form of state space model is the
exponentially) and the initial level ‘0 . The weight asso-
formulation of Harvey (1990) known as structural time
ciated with each observation depends on a. If a is large we
series models. These models give approximately the same
place a lot of weight on recent observations and in the
forecasts as the innovations state space models. In contrast
extreme case that a ¼ 1 we do not learn anything from
to the single source of error innovations state space models,
history. If a is small more weight is placed on historical
structural time series models are sometimes referred to as
information and in the extreme case where a ¼ 0 we do not
multiple source of error models, as each equation carries its
learn anything from new information as our forecast is
own independent error term. These types of models
always equal to ‘0 . More general exponential smoothing
have been successfully applied in the tourism literature.
methods include trend and seasonal terms.
González and Moral (1996), Kulendran and King (1997)
Recently, a statistical framework for such forecasting
and Kulendran and Witt (2001) only consider basic
methods has been developed (Hyndman et al., 2002; Ord et
structural models (often referred to as the non-causal
al., 1997). Innovations state space models encapsulate the
model as it does not include any explanatory variables).
notion of exponential smoothing in a state space frame-
González and Moral (1995), Garcı́a-Ferrer and Queralt
work, and allow maximum likelihood estimation, model
(1997), Greenidge (2001), Turner and Witt (2001) and
selection and prediction intervals to be derived. Their
Kulendran and Witt (2003) also consider causal structural
general form is
time series models. The results in these papers indicate that
yt ¼ w0 xt1 þ et , ð7Þ the structural time series models produce quite accurate
xt ¼ Fxt1 þ get , ð8Þ forecasts. So much so that Li et al. (2005) in their extensive
literature survey, list structural time series models along-
where yt is an observation at time t, xt is a vector of side time-varying parameter models as the models that
unobserved components which can be a mixture of a level, perform consistently well. Furthermore it seems that
growth and a seasonal component, et is Gaussian white explanatory variables do not help in forecasting tourism
noise with mean zero and variance s2 , w is a vector demand. In general, the non-causal basic structural models
containing elements of zeros and ones, F is a transition out-forecast the causal structural time series models.
matrix containing zeros, ones and possibly model para- Finally, du Preez and Witt (2003) consider a multivariate
meters, and g is a vector of unknown parameters which basic structural model. This is outperformed by the
determine the impact of the random noise on the univariate model.
unobserved components of the series. Eq. (7) is referred The modelling strategy we follow in this paper is a
to as the measurement (or observation) equation and restricted version of the Hyndman et al. (2002) methodol-
describes the relationship between the unobserved states ogy for innovations state space models, which is based on
and observation yt . Eq. (8) is referred to as the transition the taxonomy proposed by Pegels (1969), extended by
(or state) equation and describes the evolution of the states Gardner (1985) and advocated by Makridakis et al. (1998).
over time. In this study, only additive seasonal models are considered.
Notice that, unlike structural time series models The models considered are listed in Table 4. All three
(Harvey, 1990), innovations state space models only models have additive errors and an additive seasonal
involve a single source of error. For example, the model component and may contain no trend, an additive trend or
that underlies the simple exponential smoothing method of a damped trend. We refer to this class of models as ‘‘ETS’’
Eqs. (4) and (5) is known as the local level model: (for Error-Trend-Seasonal) models. The damped trend
yt ¼ ‘t1 þ et , ð9Þ model was proposed by Gardner and McKenzie (1985) as a
modification to Holt’s linear model. This modification
‘t ¼ ‘t1 þ aet . ð10Þ
comes through the parameter f which dampens the trend.
Hyndman, Koehler, Ord, and Snyder (2005) show that the When 0ofo1, the forecasts produced by the model
optimal forecasts from innovations state space models are converge to ‘n þ bn =ð1  fÞ as h ! 1 where n is the time
identical to those obtained using exponential smoothing of the last observation. Thus, the short-run forecasts are
methods. trended but the long-run forecasts are constant.
Exponential smoothing methods have been previously Treating this as a time series modelling exercise, the
applied to tourism data. Law (2000), Burger, Dohnal, dependent variable considered is visitor nights instead of
Kathrada, and Law (2001), Lim and McAleer (2001b) and visitor nights per capita. The parameters are restricted to
Cho (2001, 2003) directly implement exponential smooth- 0oao1, 0oboa, 0ogo1 and 0ofo0:98. The damping
ing methods to forecast tourism demand. The general parameter f is restricted to a maximum of 0.98 to ensure
 ARTICLE IN PRESS
G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31 25

Table 4
Innovations state space additive models with seasonal component

No trend Additive trend Damped trend

yt ¼ ‘t1 þ stm þ et yt ¼ ‘t1 þ bt1 þ stm þ et yt ¼ ‘t1 þ bt1 þ stm þ et

‘t ¼ ‘t1 þ aet ‘t ¼ ‘t1 þ bt1 þ aet ‘t ¼ ‘t1 þ bt1 þ aet
st ¼ stm þ get bt ¼ bt1 þ bet bt ¼ fbt1 þ bet
st ¼ stm þ get st ¼ stm þ get
y^ tþhjt ¼ ‘t þ stþhm y^ tþhjt ¼ ‘t þ hbt þ stþhm y^ tþhjt ¼ ‘t þ ð1 þ f þ    þ fh1 Þbt
þ stþhm

‘t denotes the level of the series at time t; bt denotes the slope at time t; st denotes the seasonal component at time t; m is the number of seasons in a year;
y^ tþhjt denotes a forecast of ytþh based on all the data up to time t.

that the damped model gives noticeably different forecasts Table 5

from the additive trend model. Section 3.3 includes further Models and estimated parameters
discussion on the damped trend model. Model Holiday VFR No Business Other
The models are estimated by maximising the likelihood parameter Damped trend trend Damped trend Damped trend
function using the R package ‘‘forecast’’ (Hyndman, 2006),
and the Akaike information criterion (AIC) is employed as a 0.10 0.52 0.00 0.00
b 0.08 0.00 0.00
the model selection criterion. The Bayesian information
g 0.00 0.00 0.00 0.00
criterion (BIC) was also considered but was found to be f 0.85 0.98 0.34
too restrictive. The selected models and estimated coeffi-
cients are presented in Table 5. The model selected for the
‘‘Holiday’’, ‘‘Business’’ and ‘‘Other’’ series is the damped
trend model, while a no-trend model is selected for VFR. The dependent variable employed is visitor nights per
The seasonal smoothing parameter, g, is zero for all capita, i.e., yt is VN it as defined in Eq. (1). Because the
models, indicating a fixed and unchanging seasonal seasonal component in each of the time series models was
pattern. deterministic, we model seasonality using seasonal dum-
mies in the set of exogenous variables.
3.3. Innovations state space models with exogenous variables A damped trend model is used for all series. Including
the exogenous variables in the observation equation, the
The two modelling approaches described so far have ETSX model is
contrasting advantages and disadvantages. The regression
yt ¼ ‘t1 þ bt1 þ z0t d þ et , ð11Þ
models identified some very useful economic relationships,
such as the positive relationship between consumer ‘t ¼ ‘t1 þ bt1 þ aet ,
confidence and domestic holiday travel. The effects of bt ¼ fbt1 þ bet ,
significant world events such as the Sydney Olympic games ^
y^ nþh ¼ ‘n þ ð1 þ f þ    þ fh1 Þbn þ z^ 0nþh d, ð12Þ
and the Bali bombings were also highlighted. These
relationships are important to policy makers (such as where zt is a vector of the exogenous variables not
Tourism Australia). However, the regression model does including the constant and the time trend. When
have some disadvantages. For example, if the model is used 0ofo1, the forecasts produced by model (12) converge
for forecasting, forecasts of the regressors are required. to ‘n þ bn =ð1  fÞ þ z^ 0nþh d^ as h ! 1. Thus, the short-term
Furthermore, the regression model is static—it does not forecasts produced by this model are largely affected by the
explore the dynamic properties of the data. In contrast, final trend estimate bn . However, as we forecast further into
time series models such as the ETS capture the dynamic the future, this effect diminishes (i.e., the trend is damped).
characteristics of the data and use these to forecast the This allows long-term forecasts to be largely driven by the
future. forecasts of the exogenous variables. This type of model
In this section, a combination of these two modelling has important policy implications as long-term forecasts
strategies is proposed, giving ETS models with exogenous (in this case tourism demand) can reflect the beliefs/views
variables or ‘‘ETSX’’ models. These models are estimated of policy makers about the future of the exogenous
via a two-step procedure. First, we identify the exogenous variables.
variables to be included in the model. These are the The estimated models are presented in Table 6. The
variables found to be statistically significant for each smoothing parameters are restricted as in Section 3.2. The
equation in the SUR estimation results, as presented in estimates of the coefficients of the exogenous variables
Table 3. In the second step, the fully specified model is seem to be consistent with the corresponding estimates in
estimated by maximising the likelihood function. the regression framework of Section 3.1.
 ARTICLE IN PRESS
26 G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31

Table 6 no deterministic trend was included in the regression

Estimates of the ETSX models model, the fit of the ETSX model was best. In the case of
Holiday VFR Business Other
the ‘‘Other’’ series the ETS model performed best. This is
not surprising because this series is the most difficult to
Parameter model via economic relationships as it contains travel for
a 0.13 0.00 0.47 0.01 very diverse purposes.
b 0.01 0.00 0.00 0.00
f 0.98 0.97 0.98 0.76
Variable 3.5. Forecast evaluation of the three models and TFC
Dt1 6.79 3.78
Pt1 7.25 4.21 The in-sample evaluation of the models has shown that the
Yt 67.67 models fit the data quite well. Although this is useful when
BALI t 132.09
modelling domestic tourism, it does not mean that the
OLYMPt 104.05
MARt 661.69 213.54 95.78 129.18 models can forecast well. In order to get some indication of
JUN t 65.52 72.54 21.25 116.15 the forecasting performance of the models, we also conduct
SEPt 48.64 31.95 32.91 27.51 an out-of-sample forecast performance evaluation.
Due to the short sample size available, the period
September 2004–June 2005 is selected as the ‘‘holdout’’
Table 7 sample. Thus, the models are estimated using the first 26
In-sample accuracy measures for the three models observations (March 1998–June 2004) and 1–4 step-ahead
Regr ETS ETSX Regr ETS ETSX forecasts are produced.
The forecast error measures are presented in Table 8.
RMSE ME For each measure, the first four rows show the error
Holiday 814.0 1173.1 974.5 8.5 244.4 182.0
VFR 1216.7 1266.6 1196.1 36.6 56.7 55.4
produced by each model for each individual series. The
Business 510.4 688.8 541.6 12.5 0.0 28.6 ‘‘Total’’ row gives the forecast error produced by each
Other 583.4 548.3 581.2 34.7 0.0 24.3 model for the total aggregate of visitor nights. The final
Total 1649.2 2264.9 1714.8 92.3 187.7 290.2 ‘‘Average’’ row gives the error of each model, averaged
MAE MAPE across the four individual series.
Holiday 665.7 994.2 757.8 1.9 2.9 2.2 Again, we will focus on the most popular forecast error
VFR 930.8 949.7 899.9 4.0 4.2 3.9
measure, the MAPE, presented in the lower right corner of
Business 395.5 548.8 450.6 3.5 4.9 4.1
Other 481.6 454.4 473.2 10.4 9.5 10.4 Table 8. The three models developed in this study seem to
Total 1352.0 1649.0 1421.4 1.9 2.3 2.0 be competitive in forecasting the individual series. Fore-
casting the total visitor nights, the ETSX models perform
best producing the lowest MAPE of 4.20%.
There is only one instance where the TFC forecasts
3.4. In-sample evaluation of the three models
outperform any of the models. This is for the ‘‘Business’’
series where the TFC forecasts outperform the ETS model.
We evaluate how well the three models fit the data by
The average improvement that can be achieved by tourism
computing some in-sample accuracy measures. The accu-
analysts in implementing our ideas is highlighted in
racy measures employed are the root mean squared error
(RMSE), the mean error (ME), the mean absolute error
(MAE) and the mean absolute percentage error (MAPE) Table 8
(see Makridakis et al., 1998, for definitions). The results are Forecast error measures calculated for the holdout sample: September
2004–June 2005
presented in Table 7.
For each error measure, the first four rows summarize the Regr ETS ETSX TFC Regr ETS ETSX TFC
error produced by each of the three models when fitted to the
RMSE ME
individual series. The final row labelled ‘‘Total’’ summarizes
Holiday 680.8 1633.1 1761.0 2255.6 185.8 383.7 71.7 286.1
the aggregated errors produced by each model for the total of VFR 1925.9 1625.4 1892.0 2449.9 919.5 422.4 1067.2 1718.9
visitor nights. The mean error provides a measure of the bias Business 1787.0 1081.5 857.7 748.2 363.4 919.3 612.6 397.0
in the fitted models, and the other three measures describe the Other 535.4 468.5 536.0 1056.3 122.3 173.6 171.7 73.1
Total 3746.6 3696.0 3826.7 4233.5 1219.5 1898.9 1923.1 2328.9
accuracy of the models in fitting the data. Average 1232.3 1202.1 1261.7 1627.5 304.9 474.7 480.8 582.2
Let us concentrate on the MAPE results presented on the
lower right corner of the table. The regression model MAE MAPE
Holiday 1856.7 1426.8 1528.3 2186.1 5.8 4.8 5.0 7.0
provides the best fit for the ‘‘Holiday’’ and ‘‘Business’’ VFR 954.2 1131.9 1118.8 1882.2 4.8 5.2 5.5 8.5
series. For these cases the regression model included a Business 507.5 919.3 612.6 731.9 5.2 9.5 6.4 7.4
deterministic trend. The fit of the ETSX models appro- Other 380.5 316.1 371.0 906.2 7.7 6.5 7.6 17.6
Total 2960.5 2757.6 2657.6 3126.9 4.5 4.3 4.2 4.9
ached the fit of the regression models as the damping
Average 924.7 948.5 907.7 1426.6 5.9 6.5 6.1 10.1
parameter f approached one. For the ‘‘VFR’’ case, where
 ARTICLE IN PRESS
G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31 27

the ‘‘Average’’ row. Here the TFC average MAPE is between Table 9
60% and 70% larger than that for any of our models. Percentage growth/decline in visitor nights from 1998 to 2014

Actual Regr ETS ETSX TFC

4. Comparing long-run forecasts from the three models and
TFC Period
98–99 0.28
99–00 0.30
Fig. 4 plots the long-term visitor nights forecasts 00–01 1.27
(aggregated to annual), produced by each of the models. 01–02 3.11
Also plotted are the forecasts produced by the TFC 02–03 1.52
(Tourism Forecasting Committee, 2005). The annual 03–04 0.94
percentage growth values in these forecasts are shown in 04–05 2.68 3.74 2.77 3.27
05–06 0.99 0.19 1.13 0.66
Table 9. All three models predict an overall decrease in 06–07 0.43 0.45 0.36 0.95
visitor nights for the year 2005. This is also predicted by the 07–08 0.42 0.30 0.26 1.05
TFC. The largest decline is predicted by the ETS models (a 08–09 0.41 0.22 0.17 1.01
drop of 3.74% in comparison to 2004) followed by the 09–10 0.40 0.17 0.09 0.96
TFC forecast (a drop of 3.27%). However, from 2006, the 10–11 0.38 0.14 0.01 0.93
11–12 0.37 0.12 0.06 0.88
forecasts produced by the models tell a very different story 12–13 0.36 0.11 0.13 0.85
about the future of Australian domestic tourism compared 13–14 0.34 0.10 0.19 0.83
to the TFC forecasts.
Average growth
The regression model driven by deterministic trends 99–04 0.21
(in the ‘‘Holiday’’ and ‘‘Business’’ series) predicts a 05–10 0.56 0.84 0.42 0.23
continuous decline in visitor nights. The average predicted 05–14 0.48 0.55 0.22 0.48
decline over the 2005–2014 period is 0.48% per annum. Total growth
The rate of decline diminishes further the forecasts are into 98–04 1.16
the future, due to the nature of the exponential trend. If the 04–10a 3.34 5.01 2.54 1.29
forecast horizon was extended enough, one would see the 04–14a 4.73 5.45 2.17 4.86
forecasts approach an asymptote parallel to the horizontal 05–14b 2.11 1.78 0.61 8.41
axis. This is an important implication of the exponential a
Total growth from the last actual observation to the forecast period
(instead of linear) trend on the forecasts produced by these stated.
b
models. It shows that this type of model does not forecast Total growth over the forecasted periods.
negative visitor nights at some point in the future, as would
be the case if a linear trend was employed. The most optimistic of our forecasts come from the
The ETS models predict the lowest number of domestic ETSX models. Recall that these models are a combination
visitor nights with not one increase for the period of the ETS damped trend model and the exogenous
2005–2014. Given that three of the series are modelled by variables. The characteristics of the forecast function of
a damped trend model the rate of decline is damped. The this model (discussed in Section 3.3) can be now seen in
average percentage decrease in visitor nights over the operation. From 2006 to 2011, the models predict a decline
period 2005–2014 is 0:55%. However, the average decrease in the visitor nights, but this decline is rapidly damped.
for the period 2005–2008 is 1:17% compared to 0:13% for Beyond 2011, the trend is damped enough for the growth in
the period 2009–2014. the exogenous variables (such as the growth in the
Australian population, the growth in GDP, etc.) to
dominate and produce an overall growth in visitor nights.
Regr
For the whole period 2005–2014 these models predict an
310000
ETS average decline of 0:22%.
ETSX
TFC
In contrast to the forecasts from any of the three models,
the TFC predict a steady growth in visitor nights for the
300000 period 2005–2014. The average predicted growth over
2005–2014 is 0:48%. This means that domestic visitor
nights are predicted to increase by 8:41% over the next 10
290000 years. This is a much greater total increase than that
predicted by any of the three statistical models. The
regression and ETS models predict a total decline of 2:11%
280000
and 1:78%, respectively. The ETSX models predict a small
total growth of 0:61%.
2000 2005 2010 Fig. 5 presents the annual forecasts (from the three
Fig. 4. Long-run annual forecasts for total visitor nights from the three models and the TFC) for the disaggregated series by
models and the TFC. purpose of travel. This figure highlights where the TFC
 ARTICLE IN PRESS
28 G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31

PANEL A : Holiday PANEL B : VFR

145000 110000
140000 105000

135000 100000

130000 95000
90000
125000
85000
120000
80000
2000 2005 2010 2000 2005 2010

PANEL C : Business PANEL D: Other

48000 23000
46000
22000
44000
21000
42000
40000 20000
38000 19000
36000
18000
2000 2005 2010 2000 2005 2010

Regr ETS ETSX TFC

Fig. 5. Long-run annual forecasts for each of the four travel purpose components of visitor nights, from the three models and the TFC.

seem to have been over-optimistic. For two of the four innovations state space models by proposing innovations
series, namely ‘‘VFR’’ and ‘‘Other’’, the forecasts produced state space models with exogenous variables.
by the models and the TFC are fairly similar (the exception All three statistical models are shown to outperform the
being ETS for both cases which has a flat forecast TFC published forecasts for short-term demand of
function). However, there is a noticeable discrepancy Australian domestic tourism. The long-term forecasts
between the models and the TFC forecasts for the other produced by the models indicate that the TFC long-term
two series. The largest component of the total visitor nights forecasts may be optimistic. In particular, the models
is ‘‘Holiday’’ travel. Panel A highlights the significant suggest that TFC forecasts of the ‘‘Holiday’’ and ‘‘Busi-
divergence between the models and the TFC in the long- ness’’ travel components of Australian domestic tourism
term forecasts for this series. There is also a divergence have been optimistic. The statistical models show that
between the models and the TFC forecasts for the Australian domestic tourism is on the decline.
‘‘Business’’ component as shown in Panel C. The forecasts The proposed statistical models are clearly of substantial
of these two components are the primary source of the benefit to policy makers. In particular, we recommend the
overly optimistic TFC forecasts for total visitor nights. use of innovations state space models with exogenous
variables (the ETSX models), which can capture time series
dynamics as well as economic and other relationships, and
5. Conclusion which out-performed the other models based on both the
MAE and MAPE error measures.
We have modelled Australian domestic tourism demand
using three statistical models. The first approach used
regression models. The estimated regression models have Acknowledgments
identified significant economic relationships for domestic
tourism. This analysis has also highlighted the impact of We acknowledge valuable comments and suggestions
world events on Australian domestic tourism such as the from Farshid Vahid, Brett Inder, Chris Ryan (Editor),
increase in business travel immediately after the 2000 Sydney three anonymous referees, seminar participants from
Olympic games, and the significant increase in visiting friends the Department of Econometrics and the Tourism
and relatives after the 2002 Bali bombings. In order to take Research Unit at Monash University, and the 2007 Council
advantage of the time series properties of the data, we also for Australian University Tourism and Hospitality
consider time series modelling and implement innovations Education. We thank Andrew Maurer and Tim Quinn
state space models (for the first time in the tourism literature). at Tourism Research Australia for providing data, sugges-
We combine the properties of the regression and the tions and explanations. We would like to acknowledge
 ARTICLE IN PRESS
G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31 29

the financial support of Tourism Research Australia and aging across the months of each quarter, series 5671-P1A
the Sustainable Tourism Cooperative Research Centre. from the DX database, Australia).
DPI: Domestic holiday travel and accommodation price
index (seasonally adjusted by additive method, Australian
Appendix A. Descriptions and projections of regressors Bureau of Statistics, series ID A2329356K).
GDP: Real Gross Domestic Product per capita in
DEBT: Total real personal finance commitments per constant 03–04 $A billions (seasonally adjusted, series
capita from all lenders in Australia in $A thousands AUS.EXPGDP.LNBQRSA from the DX database,
(seasonally adjusted, aggregated monthly series by aver- Australia).

DEBT DPI
180

160
250000 140

120
150000 100

1985 1990 1995 2000 2005 2010 2015 1995 2000 2005 2010 2015

GDP POP
13
22
11 21

9 20
8
19
7
6 18
1980 1990 2000 2010 1995 2000 2005 2010 2015

Fig. A.1. Levels of the regressors and their projections.

PANEL A PANEL B
7800 7300
7700
7200
7600
7500 7100
7400
7000
7300
7200 6900

1998 2000 2002 2004 1998 2000 2002 2004

PANEL C PANEL D

6500 5800

6400 5600

6300 5400

6200 5200

6100 5000
1998 2000 2002 2004 1998 2000 2002 2004

Fig. B.1. These are the visitor nights per capita series as defined in Eq. (3): Panel A: ln VN Hol
t  1000; Panel B: ln VN VFR
t  1000; Panel C:
ln VN Bus
t  1000; Panel D: ln VN Oth
t  1000.
 ARTICLE IN PRESS
30 G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31

POP: Australian population and population projections Harvey, A. C. (1990). Forecasting, structural times series models and the
as provided by the Australian Bureau of Statistics: Series B Kalman filter. Cambridge: Cambridge University Press.
Population Projections, Australia 2004–2101, ABS cat. no. Holt, C. E. (1957). Forecasting seasonals and trends by exponentially
weighted averages. O.N.R. Memorandum 52, Carnegie Institute of
3222.0. Technology, O.N.R. Memorandum, Pittsburgh USA. Published in
The forecasts for GDP, DPI and DEBT were produced 2004 in the International Journal of Forecasting 20, 1–13, with
using the R package ‘‘forecast’’ (Hyndman, 2006) additional commentaries.
(Fig. A.1). The models selected by minimizing the AIC Hyndman, R. J. (2006). Forecast: Forecasting time series. R package
are: additive trend, additive trend and no-trend, respec- version 1.0. hURL:http://www.robhyndman.info/Rlibrary/forecasti.
Hyndman, R. J., Koehler, A. B., Ord, J. K., & Snyder, R. D. (2005).
tively. The forecasts for POP are the population projec-
Prediction intervals for exponential smoothing using two new classes
tions supplied by the Australian Bureau of Statistics. of state space models. Journal of Forecasting, 24, 17–37.
Hyndman, R. J., Koehler, A. B., Snyder, R. D., & Grose, S. (2002). A
Appendix B. Visitor nights per capita series state space framework for automatic forecasting using exponential
smoothing methods. International Journal of Forecasting, 18, 439–454.
Jarque, C., & Bera, A. (1980). Efficient tests for normality homoskedas-
The visitor nights per capita series according to Eq. (3) is ticity and serial independence of regression residuals. Economic
given in Fig. B.1. Letters, 6, 255–259.
Kulendran, K., & King, M. L. (1997). Forecasting international quarterly
tourist flows using error-correction and time-series models. Interna-
References tional Journal of Forecasting, 13, 319–327.
Kulendran, N., & Witt, S. (2001). Cointegration versus least squares
Aoki, M. (1987). State space modelling of time series. Berlin: Springer. regression. Annals of Tourism Research, 28, 291–311.
Breusch, T. (1978). Testing for autocorrelation in dynamic linear models. Kulendran, N., & Witt, S. F. (2003). Forecasting the demand for
Journal of Australian Economic Papers, 17, 334–355. international business tourism. Journal of Travel Research, 41,
Brown, R. G. (1959). Statistical forecasting for inventory control. New 265–271.
York: McGraw-Hill. Kwiatkowski, D., Phillips, P. C. B., Schmidt, P., & Shin, Y. (1992).
Burger, C., Dohnal, M., Kathrada, M., & Law, R. (2001). A practitioners Testing the null hypothesis of stationarity against the alternative of a
guide to time-series methods for tourism demand forecasting—a case unit root: How sure are we that the economic time series have a unit
study of Durban, South Africa. Tourism Management, 22, 403–409. root? Journal of Econometrics, 54, 159–178.
Cho, V. (2001). Tourism forecasting and its relationship with leading Law, R. (2000). Back-propagation learning in improving the accuracy of
economic indicators. Journal of Hospitality and Tourism Research, 24, neural network-based tourism demand forecasting. Tourism Manage-
399–420. ment, 21, 331–340.
Cho, V. (2003). A comparison of three different approaches to tourist Li, G., Song, H., & Witt, S. F. (2005). Recent developments in
arrival forecasting. Tourism Management, 24, 323–330.
econometric modelling and forecasting. Journal of Travel Research,
de Silva, A., Hyndman, R. J., & Snyder, R. D. (2006). The vector
44, 82–99.
innovation structural time series framework: A simple approach to
Lim, C., & McAleer, M. (2001a). Cointegration analysis of quarterly
multivariate forecasting. Working paper, Department of Econometrics
tourism demand by Hong Kong and Singapore for Australia. Applied
and Business Statistics. Melbourne: Monash University.
Economics, 33, 1599–1619.
Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for
Lim, C., & McAleer, M. (2001b). Forecasting tourism arrivals. Annals of
autoregressive time series with a unit root. Journal of the American
Tourism Research, 28, 965–977.
Statistical Association, 74, 427–431.
Lim, C., & McAleer, M. (2002). Time series forecasts of international
Dickey, D. A., & Fuller, W. A. (1981). Likelihood ratio statistics for
travel demand for Australia. Tourism Management, 23, 389–396.
autoregressive time series with a unit root. Econometrica, 49,
Makridakis, S., Anderson, A., Carbone, R., Fildes, R., Hibon, M.,
1057–1072.
du Preez, J., & Witt, S. F. (2003). Univariate versus multivariate time Lewandowski, R., et al. (1982). The accuracy of extrapolation (time
series forecasting: An application to international tourism demand. series) methods: Results of a forecasting competition. Journal of
International Journal of Forecasting, 19, 435–451. Forecasting, 1, 111–153.
Garcı́a-Ferrer, A., & Queralt, R. A. (1997). A note on forecasting Makridakis, S., & Hibon, M. (2000). The M3-competition: Results,
international tourism demand of Spain. International Journal of conclusions and implications. International Journal of Forecasting, 16,
Forecasting, 13, 539–549. 451–476.
Gardner, E. S., Jr. (1985). Exponential smoothing: The state of the art. Makridakis, S., Wheelwright, S. C., & Hyndman, R. J. (1998).
Journal of Forecasting, 4, 1–28. Forecasting: Methods and applications (3rd ed.). New York: Wiley.
Gardner, E. S., Jr., & McKenzie, E. (1985). Forecasting trends in time Morley, C. L. (1998). A dynamic international demand model. Annals of
series. Management Science, 31(10), 1237–1246. Tourism Research, 25, 70–84.
Godfrey, L. (1978). Testing for higher order serial correlation in regression Morris, A., Wilson, K., & Bakalis, A. (1995). Modelling tourism flows
equations when the regression contain lagged dependent variables. from Europe to Australia. Tourism Economics, 1, 147–167.
Econometrica, 46, 1303–1310. Ng, S., & Perron, P. (2001). Lag length selection and the construction
González, P., & Moral, P. (1995). An analysis of the international tourism of unit root tests with good size and power. Econometrica, 69,
demand of Spain. International Journal of Forecasting, 11, 233–251. 1519–1554.
González, P., & Moral, P. (1996). Analysis of tourism trends in Spain. Ord, J. K., Koehler, A. B., & Snyder, R. D. (1997). Estimation and
Annals of Tourism Research, 23, 739–754. prediction for a class of dynamic nonlinear statistical models. Journal
Green, W. H. (2000). Econometric analysis (4th ed.). Upper Saddle River, of the American Statistical Association, 92, 1621–1629.
NJ: Prentice-Hall International. Ord, J. K., Snyder, R. D., Koehler, A. B., Hyndman, R. J., & Leeds, M.
Greenidge, K. (2001). Forecasting tourism demand. Annals of Tourism (2005). Time series forecasting: The case for the single source of error
Research, 28, 98–112. state space. Working Paper 7/05.
Hannan, E. J., & Deistler, M. (1988). The statistical theory of linear Pegels, C. C. (1969). Exponential smoothing: some new variations.
systems. New York: Wiley. Management Science, 12, 311–315.
 ARTICLE IN PRESS
G. Athanasopoulos, R.J. Hyndman / Tourism Management 29 (2008) 19–31 31

Ramsey, J. (1969). Test for specification errors in classical linear least Turner, L. W., & Witt, S. F. (2001). Forecasting tourism using univariate
squares regression analysis. Journal of the Royal Statistical Society B, and multivariate structural time series models. Tourism Economics, 7,
31, 350–371. 135–147.
Snyder, R. D. (1985). Recursive estimation of dynamic linear models. White, H. (1980). A heteroskedasticity-consistent covariance matrix
Journal of the Royal Statistical Society B, 47, 272–276. estimator and a direct test of heteroskedasticity. Econometrica, 48,
Tourism Forecasting Committee (2005). Tourism Forecasting Committee 817–838.
October 2005 forecasts. Canberra: Tourism Research Australia. Winters, P. R. (1960). Forecasting sales by exponentially weighted moving
Tourism Research Australia (2005). Travel by Australians, September averages. Management Science, 6, 324–342.
Quarter 2005. Canberra: Tourism Australia. Zellner, A. (1963). Estimators for seemingly unrelated regression
Tourism White Paper (2003). A medium to long term strategy for tourism. equations: Some exact finite sample results. Journal of the American
Commonwealth of Australia. Available on hhttp://www.industry.gov.aui. Statistical Association, 58, 977–992.