ELSEVIER International Journal of Forecasting 11 (1995) 447-475

Forecasting tourism demand: A review of empirical research

Stephen F. Witt a'*, Christine A. Witt b
aEuropean Business Management School, University of Wales, Swansea SA2 8PP, UK
bManagement Centre, University of Bradford, Bradford BD9 4JL, UK

Abstract

Accuracy is particularly important when forecasting tourism demand on account of the perishable nature of the
product. The main methods used to forecast tourism demand which are reported in published empirical studies are
discussed, together with the empirical findings. The vast majority of such studies are concerned with econometric
modelling/forecasting, and the most appropriate explanatory variables are examined. Particular emphasis is placed
on empirical comparisons of the accuracy of tourism forecasts generated by different techniques. Considerable
scope exists for improving the model specification techniques employed in econometric forecasting of tourism
demand. Furthermore, no single forecasting method performs consistently best across different situations, but
autoregression, exponential smoothing and econometrics are worthy of consideration as alternatives to the no
change model.

Keywords: Tourism forecasting; Empirical review; Quantitative methods; Qualitative methods; Accuracy comparisons

1. Introduction sations have produced international tourism de-

mand forecasts in the past on a country-to-coun-
The need for accurate forecasts of tourism try level (e.g. USA to Spain) (Brooke et al.,
demand to assist managerial decision making is 1985; Means and Avila, 1986, 1987, 1988).
well recognised. For example, Archer (1987) Edwards (1985) publishes more aggregate fore-
points out that "In the tourism industry ... the casts (e.g. USA to neighbouring countries, USA
need to forecast accurately is especially acute to other short/medium haul destinations, USA
because of the perishable nature of the product. to long haul destinations) at approximately five-
Unfilled airline seats and unused hotel rooms year intervals. The larger companies in the
cannot be stockpiled" (p.77). Clearly, similar tourism industry (such as major airlines) gener-
considerations apply to unfilled coach seats, ate forecasts internally, but these are confiden-
unoccupied apartments, unused hire cars and so tial.
on. However, although the benefits of accurate Various time horizons and levels of aggrega-
forecasts to airlines, hotel chains, etc. are clear, tion are relevant to decision making. For exam-
there are no standard suppliers of forecasts to ple, short term forecasts are required for
the tourism industry. Certain commercial organi- scheduling and staffing, medium term forecasts
for planning tour operator brochures and long
* Corresponding author. term forecasts for investment in aircraft, hotels

0169-2070/95/$09.50 t~) 1995 Elsevier Science B.V. All rights reserved

SSDI 0169-2070(95)00591-9
448 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

and infrastructure. Governments are interested demand forecasting methods. Hence, this paper
in total international inbound and outbound presents an up-to-date and more comprehensive
tourist expenditures at a country level, hotels in review of the tourism demand forecasting litera-
tourism demand by city or region and airlines in ture than has appeared elsewhere, and also
tourism demand by route. However, the bulk of evaluates the accuracy of tourism forecasts gen-
published forecasting studies on tourism demand erated by various models.
concentrates on the short to medium term at a
highly aggregated (country-to-country) level.
The purpose of this paper is to review and 2. Tourism forecasting techniques
evaluate the existing empirical literature on
tourism demand forecasting. It should be noted Previous reviews of tourism demand forecast-
that tourism differs from travel in general in that ing often list many possible forecasting methods.
various sorts of travel are excluded, such as the In practice, however, the set of techniques used
journey to work. Additionally, the focus is not to forecast tourism demand is rather limited. The
just on the travel component of a trip, but also majority of articles which are concerned with
on destination accommodation, attractions and tourism demand forecasting are econometric
so on. Although several tourism forecasting studies. (In fact, although such studies often
reviews exist, most of these are published in suggest that the econometric models developed
tourism journals and specifically aimed at tour- may be used for forecasting purposes, the
ism audiences; hence considerable attention is models are usually not used to generate forecasts
devoted to explaining the various techniques which are then evaluated in terms of accuracy.)
which can be used to forecast tourism demand, Other quantitative forecasting techniques which
together with their advantages and disadvan- actually have been used in tourism situations
tages. Indeed some reviews go no further than include spatial models (particularly gravity
this (Archer, 1980, 1987; Vanhove, 1980; BarOn, models) and time-series models. Empirical re-
1984; Van Doorn, 1984). Other reviews addition- search on qualitative forecasting in tourism has
ally include a brief discussion of empirical results centred on Delphi studies and scenarios.
regarding the relative accuracy of different fore- In reviewing the main empirical studies that
casting methods when applied to tourism de- have been published in the area of tourism
mand (Van Doorn, 1982; Uysal and Crompton, demand forecasting, the studies will be consid-
1985), while yet others include a more com- ered chronologically under each category of
prehensive discussion (Witt and Martin, 1989). forecasting method. This should assist in tracing
Calantone et al. (1987) also incorporate empiri- any major developments that have occurred over
cal material relating to the goodness of fit of time in the application of the techniques to the
regression models explaining tourism demand in tourism situation. In addition, a separate section
their review paper. will analyse those studies which evaluate the
In contrast, a broader review of tourism fore- relative performance of different techniques in
casting by Sheldon and Var (1985) resulted in terms of generating accurate forecasts of tourism
changing the emphasis from explaining forecast- demand.
ing techniques to discussing the application of
these techniques to tourism. However, only 11
studies are considered in their tabular compari- 3. Empirical research on quantitative
son of econometric models, and, even though forecasting
the Sheldon and Var article was published in
1985, only one of these studies was published 3.1. Causal methods
post 1978. Progress has been made in the inter-
vening period, both in econometric modelling of The causal multiple regression approach to
tourism demand and also in empirical assessment forecasting tourism demand involves the use of
of the relative accuracy of different tourism regression analysis to estimate the quantitative
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 449

relationship between tourism demand and its control policies and international political events
determinants; the estimation is carried out using need to be included in the discussion.
historic data, and future values of tourism de- The general form of the tourism demand
mand are obtained by using forecasts of the function is now considered, with particular stress
demand determinants in conjunction with the on the most important explanatory variables.
estimated relationship. (For a discussion of the Table 1 summarises the hypothesised models in
potential inaccuracies involved in forecasting the main empirical econometric studies of tour-
exogenous variables and the resultant impact on ism demand that have been published during the
the variable to be forecast, see Ashley, 1983.) last 30 years. While important studies have, no
Hence, in reviewing the existing literature which doubt, been omitted from the table, it does
utilises multiple regression to estimate the rela- contain those commonly cited in the tourism
tionship between tourism demand and its ex- literature. However, where an author uses the
planatory variables, only those models which are same econometric model in different articles
estimated using time-series (or pooled cross-sec- (possibly applied to different data and hence
tion and time-series) data are considered. Cross- with different empirical results), only one of the
section models do not accommodate the possi- studies is listed in the table. It is clear from the
bility of trend changes and are therefore unlikely table that the models have tended to become
to be stable over time. This is supported by more complex over time in terms of the ex-
Peterson et al. (1985) who find that cross-section planatory variables included.
recreation demand models are not stable over Forecast variable: Tourism demand is general-
time. ly measured in terms of the number of tourist
visits from an origin country to a foreign destina-
tion country or, in terms of tourist expenditures,
3.1. I. Econometric models by visitors from the origin country in the destina-
tion country. Tourist nights spent in the destina-
Tourist visits can take place for various tion country are an alternative measure. Further-
reasons: holidays, business trips, visits to friends more, where the objective is to forecast the
and relatives, pilgrimages and so on. The vast distribution of tourism demand only, the forecast
majority of empirical studies of tourism demand variable is usually the market share of the
examine either total tourist trips (i.e. for all particular destination, expressed either in terms
purposes), or just holiday trips. Only a few are of tourist visits or tourist expenditures/receipts.
concerned with business tourism. As most tourist International tourism demand data are col-
visits take place for holiday purposes, the deter- lected in various ways. Tourist visits are usually
minants of demand for total tourist trips are recorded by frontier counts (inbound), registra-
taken to be the same as those for holiday trips in tion at accommodation establishments (inbound)
most empirical studies. In discussing the ex- or sample surveys (inbound and outbound). A
planatory variables which are appropriate for problem with frontier counts is that in certain
tourism demand functions, this review will there- cases a substantial transit traffic element may be
fore concentrate on those most relevant for present. Accommodation establishment records
holiday tourism. Hence, for example, although exclude day-trippers and tourists staying with
the level of economic activity in a destination is friends or relatives or in other forms of unre-
likely to influence the demand for business gistered accommodation. Sample surveys may be
tourism, this factor is not considered. Further- applied at points of exit/entry to returning
more, as virtually all (time-series data based) residents or departing non-residents, or house-
empirical research is concerned with internation- hold surveys may be carried out, but in both
al tourism rather than domestic tourism, the cases often the sample size is relatively small.
review also focuses upon the determinants of International tourist expenditure data are usually
international tourism. As a consequence, factors collected by the bank reporting method or sam-
such as exchange rates, government exchange ple surveys. The former method is based on the
450 S.F. Witt, C.A. Witt / International Journal of Forecasting I1 (1995) 447-475

~i ~ ii ~ ~-~ ~ ~

X X X

. . . . L _ __

~u

~X X

X ~X X~ X X XX X X XX

X X

<~ ~ < << <.< < < << < < < < < < < ~ <

¢.J

t2 ~
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 451
"d

~ ° ~. ~ -~ ° °

XX ~ ~ X

X X X

,u
a
452 SJ z. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

X X X

X X X X X ~4

d
~ ~ ~:~ "

x x x x

x x

x x x xx x ,,m

x x x xx x x x x

x x x x x

< < o ~ ~< < < < < < ,<

,-)

:1
 S.F, Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 453

registration by authorised banks and agencies of population, it is likely that population changes in
the buying and selling of foreign currencies by generating countries will be small over the short-
travellers. There are many problems associated medium term, and hence the model will only be
with this method of data collection such as affected marginally.
identifying a transaction as a tourism transaction, However, if population changes as a result of
non-reporting of relevant transactions and the high immigration from a particular country, then
unreliability of its use for measuring receipts this may well result in large tourist flows back to
from specific origin countries (the geographic that country during subsequent years. In such
breakdown relates to the denomination of the situations, the level of foreign tourism from the
currency and not the generating country). Sam- origin country to the particular destination will
ple surveys provide more reliable data on tourist depend not only on the population level in the
expenditures, but as with visit data the sample origin country, but also on the proportion who
size is often relatively small. are recent immigrants from the destination coun-
Expenditures/receipts and visits are both try.
popular measures of tourism demand. Where per Income: In tourism demand functions, origin
capita or share is not specified in Table 1, this country income or private consumption is gener-
indicates that the dependent variable is mea- ally included as an explanatory variable, and
sured in total terms; again both per capita and commonly enters the demand function in per
total measures are popular. Similarly, where the capita form (corresponding to the specification of
trip purpose (holiday, business, etc.) is not demand in per capita terms). If (mainly) holiday
specified, the dependent variable measures tour- visits or visits to friends and relatives are under
ism demand for all purposes. (Leisure visits consideration then the appropriate form of the
denote all non-business visits.) It can be seen variable is private consumption or personal dis-
that out of all the studies only two focus on posable income, but if attention focuses on
business tourism, whereas nine specifically ex- business visits (or they form an important part of
amine holiday or leisure tourism. Most of the 40 the total), then a more general income variable
studies considered, however, just examine tour- (such as national income) should be used.
ism demand aggregated by visit purpose. Table 1 indicates that where income is not
Population: The level of foreign tourism from included in a model, there is usually some form
a given origin is expected to depend upon the of aggregate expenditure variable in its place, or
origin population. In six of the studies consid- in the case of business travel a measure of
ered, population features as an explanatory business activity between the origin and destina-
variable, but more often the effect of population tion. If the objective is to explain share of
is accommodated by modifying the dependent tourism demand, income/aggregate expenditure
variable to become international tourism de- is often not included as an explanatory variable;
mand per capita (15 studies). It is clear, there- however, this procedure can be criticised as it is
fore, that the impact of population changes is likely that as income increases, long-haul/expen-
ignored in many cases. sive destinations will gain share at the expense of
The main justification for not having popula- short-haul/cheap destinations.
tion as a separate explanatory variable is that its Own price: The appropriate form of the price
presence may cause multicollinearity problems, variable is by no means clear. In the case of
as population tends to be highly correlated with tourism there are two elements of price: the cost
income (see next section). On the other hand, of travel to the destination; and the cost of living
the procedure adopted whereby demand is for the tourist in the destination. Although the
specified in per capita terms, in effect, constrains theoretical justification for including transport
the population elasticity to equal unity (if a cost as a demand determinant does not appear to
log-linear model is under consideration). Al- be disputed, many authors exclude this variable
though it is theoretically incorrect to exclude from the set of explanatory variables on the
454 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

grounds of potential multicollinearity problems own is not an acceptable proxy (Martin and
and lack of data availability. (In fact, multicol- Witt, 1987).
linearity need not be a problem; instead, it may Substitute prices: Economic theory suggests
be a sign of cointegration, which would suggest that the prices of substitutes may be important
the use of error correction models.) determinants of demand. For example, an in-
In certain studies where econometric forecast- crease in holiday prices to Spain may increase
ing models have been developed for internation- demand for holidays to Portugal.
al tourism demand, a specific destination tour- Mostly, those substitution possibilities allowed
ists' cost of living variable is incorporated in the for in international tourism demand studies are
models. Usually, however, the consumer price restricted to tourists' destination living costs. A
index in a country is taken to be a proxy for the common form in which substitute prices enter
cost of tourism in that country. In general, this the demand function is to specify the tourists'
procedure is adopted on the grounds of lack of cost of living variable in the form of the destina-
more suitable data, e.g. an index "defined over tion value relative to the origin value, thus
the basket of goods purchased by tourists, rather merely permitting substitution between tourist
than over the usual typical consumer basket" visits to the foreign destination under considera-
(Kliman, 1981, p.490). (In fact, such indices are tion and domestic tourism. The usual justifica-
now published for certain countries and major tion for this form of relative price index is that
towns.) Whichever destination price variable is domestic tourism is the most important substi-
used, it needs to be adjusted by the rate of tute for foreign tourism. Other studies incorpo-
exchange in order to transform it into origin rate substitute prices in a more sophisticated
country currency. manner; they allow for the impact of competing
Exchange rates are also sometimes used separ- foreign destinations by specifying the tourists'
ately to represent tourists' living costs. Although cost of living variable as destination value rela-
they usually appear in addition to either a tive to a weighted average value calculated for a
specific tourists' cost of living variable or a set of alternative destinations, or by specifying a
consumer price index proxy, they may be the separate weighted average substitute destination
sole representation of tourists' living costs. The cost variable.
usual justification for including an exchange rate Just as tourists' living costs in substitute desti-
variable in international tourism demand func- nations are likely to influence the demand for
tions is that consumers are more aware of tourism to a given destination, so travel costs to
exchange rates than destination costs of living for substitute destinations may also be expected to
tourists, and hence are driven to use exchange have an impact. Furthermore, if the data are
rate as a proxy variable (Gray, 1966; Artus, disaggregated by transport mode, then travel
1970). However, the use of exchange rate alone cost to the same destination by alternative trans-
can be misleading because even though the port mode(s) would influence tourism demand to
exchange rate in a destination may become more a particular destination by a given transport
favourable, this could be counterbalanced by a mode. However, although some theoretical at-
relatively high inflation rate. tention has been paid to the notion of substitute
Empirical results evaluating the precise form travel costs in the literature, they rarely feature
of the tourists' cost of living element of the own in tourism demand functions.
price variable which should be included in If a price variable is specified as own price
econometric forecasting models indicate that the relative to substitute prices, then the variable is
exchange-rate-adjusted consumer price index listed generally under both own price and substi-
(either alone or together with a separate ex- tute prices in Table 1. The only exceptions are
change rate variable) is a reasonable proxy for the very restrictive cases where the sole substi-
the cost of tourism, but that exchange rate on its tute destination price considered is the cost of
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 455

living for tourists in the generating country; in therefore more likely to influence tourist flows to
fact, in those models where the cost of living for the destination concerned.
tourists is measured by the consumer price Marketing has not featured often in tourism
index, this procedure is equivalent to specifying demand models, but a critical review of studies
the own price variable in real terms. which do include some form of marketing vari-
Qualitative effects: Dummy variables can be able appears in Witt and Martin (1987c) and a
included in international tourism demand func- subsequent review appears in Crouch et al.
tions to allow for the impact of 'one-off' events (1992).
(see Table 1). For example, when governments Lagged dependent variable: A lagged depen-
impose foreign currency restrictions on their dent variable, i.e. an autoregressive term, can be
residents (e.g. the £50 annual limit introduced in justified on the grounds of habit persistence.
the U K during late 1966 to late 1969), this is Once people have been on holiday to a par-
expected to reduce outward tourism. Similarly, ticular country and liked it, they tend to return
the 1973 and 1979 oil crises are expected to have to that destination. There is much less uncertain-
temporarily reduced international tourism de- ty associated with holidaying again in that coun-
mand; although the impacts of the oil crises on try compared with travelling to a previously
holiday prices and consumer incomes are in- unvisited foreign country. Furthermore, knowl-
corporated in these explanatory variables, a edge about the destination spreads as people talk
further reduction in international tourism de- about their holidays and show photographs,
mand is likely on account of the psychological thereby reducing uncertainty for potential vis-
impact of the resultant uncertainties in the world itors to that country. In fact, this 'word of
economic situation. Witt and Martin (1987b) mouth' recommendation may well play a more
discuss a range of one-off events which have important role in destination selection than does
been accommodated by dummy variables. commercial advertising. A type of learning pro-
Trend: A trend mainly represents a steady cess is in operation and as people are, in general,
change in the popularity of a destination country risk averse, the number of people choosing a
over the period considered as a result of chang- given alternative in any year depends on the
ing tastes. It also, however, captures the time- numbers who chose it in previous years.
dependent effects of all other explanatory vari- A second justification for the inclusion of a
ables not explicitly included in the equation, lagged dependent variable in tourism demand
such as changes in air service frequencies and functions comes from the supply side. Supply
demographic changes in the origins. constraints may take the form of shortages of
Marketing: National tourist organisations en- hotel accommodation, passenger transportation
gage in sales-promotion activities specifically to capacity and trained staff, and these often cannot
attempt to persuade potential tourists to visit the be increased rapidly. Time is also required to
country, and these activities may take various build up contacts among tour operators, hotels,
forms including media advertising and public airlines and travel agencies. Similarly, once the
relations. Hence, promotional expenditure is tourist industry to a country has become highly
expected to play a role in determining the level developed it is unlikely to dwindle rapidly. If a
of international tourism demand. Much tourism- partial adjustment mechanism is postulated to
related marketing activity is not, however, allow for rigidities in supply, this results in the
specific to a particular destination (for example, presence of a lagged dependent variable in the
general travel agent and tour operator advertis- tourism demand function (Gujarati, 1988, chap-
ing) and therefore is likely to have little impact ter 16).
on the demand for tourism to that destination. Functional form: Most studies utilise individ-
The promotional activities of national tourist ual disaggregated demand equations. However,
organisations are destination specific and are in a few instances a complete system of demand
456 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

equations is specified, usually the Almost Ideal an additional criterion in order to select the final
Demand System developed by Deaton and model.)
Muellbauer (1980) (see, for example, O'Hagan The most comprehensive study in terms of the
and Harrison, 1984; White, 1985). The majority variety of origins/destinations covered is by
of studies employ a log-linear functional form, Martin and Witt (1988), and the first ten models
although the linear form also appears quite are taken from this study. The four models
often. The popularity of the log-linear model selected from other studies are specifically
stems from the fact that it is convenient to think chosen because they include explanatory vari-
in terms of elasticities. Additionally, in several ables not covered in the Martin and Witt (1988)
studies both linear and log-linear models were study. Thus model 11 incorporates population as
tried initially, but the latter model was selected a demand determinant, models 12 and 13 in-
because it yielded superior empirical results in corporate marketing expenditure and model 14
terms of 'correct' coefficient signs and model fit. incorporates a lagged dependent variable and
Where market share is modelled, a semi-log travel time. (The elasticity values presented in
functional form is generally specified. model 14 are impulse estimates.) All the models
Probit and logit models also feature in certain are specified in log-linear form.
cases. Such discrete choice models can have a Population features as an explanatory variable
variety of applications in tourism demand; for only in model 11, but the estimated elasticity of
example, to explain the holiday/no holiday deci- 12.4 seems far too high to be realistic. (The
sion or the foreign holiday/domestic holiday population elasticity is expected to be fairly close
decision. A recent application of logit analysis by to unity.) Furthermore, an 'incorrect' coefficient
Witt et al. (1992a) examines conference tourism sign is estimated for income. It appears, there-
demand, specifically the conference attendance/ fore, that multicollinearity between population
non-attendance decision. The objective is to and income may well be a problem in this model,
forecast the proportion of association members which is why population does not, in general,
from each country who will attend that associa- feature as an explanatory variable in tourism
tion's annual conference, and the probability of a demand models. The results from model 11 are
member attending is modelled as a function of not included in the discussion which follows,
the following set of predictor variables: income, Income appears in each model in Table 2, but
travel cost, conference fee, origin characteristics the values of the estimated elasticities vary
and destination attractiveness. As the probability considerably, ranging from 0.4 to 6.6. However,
is bounded between 0 and 1, the logit trans- other than the 0.4 value, the estimated income
formation is utilised (McCullagh and Nelder, elasticities exceed unity, showing clearly that
1989, chapter 4). foreign tourism is generally regarded as a luxury.
Empirical results: Examples of estimated The median value of 2.4 strongly supports a
econometric models are presented in Table 2. As priori expectations about the luxury nature of
is generally the case in econometric studies of foreign tourism. Many of the differences in
tourism demand, little attention has been paid to elasticity estimates can be readily explained. For
whether the models are well specified or not. example, although the income elasticity obtained
Thus, for example, Durbin Watson (DW) tests in model 5 (0.4) is somewhat lower than ex-
for autocorrelation appear in the studies and the pected, travel from the USA to Canada is likely
Cochrane-Orcutt procedure is used where the to be regarded in a similar manner to domestic
authors consider appropriate, but diagnostics for (USA) tourism, i.e. a necessity, rather than a
outliers, parameter constancy, etc. have not luxury; whereas, overseas travel from the USA
been examined. It is therefore necessary to to the UK (model 6) is viewed as a luxury
exercise caution when interpreting the empirical (income elasticity is 2.4). By contrast, the value
results. (In the Martin and Witt (1988) study, of the income elasticity for travel from France to
outside sample forecasting accuracy was used as Switzerland (model 1) is similar to that for travel
 S J:. Witt, C.A. Witt / International Journal of Forecasting 11 (1995)447-475 457

~8

~9 979~ 9~9979~~~9 ?9 .... 99~

i ~ l l l r ii ~ i II l

~ ~ h ~¸
i i i i i

888~88 888 ~ 8 ~
E
0
=
8
458 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

from France to the UK (model 2) at 2.8. Again, the models, showing, in particular, that political
this makes sense as Switzerland is a destination events can have marked impacts on tourism
with a very high cost of living for tourists, demand. Models 12 and 13 indicate that market-
whereas visiting the UK involves a high cost of ing expenditure does have a positive impact on
travel on account of the sea border, and there- tourism demand, but that the estimated elas-
fore holidays to Switzerland and the UK are ticities are low. Model 14 demonstrates the
likely to be regarded as being at about the same negative impact of travel time on tourism de-
luxury level by the French. (However, the in- mand, and the importance of habit persistence
come elasticity for travel from France to neigh- and/or supply constraints (as represented by the
bouring cheap destinations (such as Spain) would lagged dependent variable).
be expected to be much lower-see, for example, In terms of traditional regression analysis,
the difference in income elasticities for travel reasonably good empirical results are obtained.
from Germany to Switzerland (high cost) and Most of the coefficients have 'correct' signs and
Austria (low COSt).) are of reasonable magnitude, many of the co-
As far as own price is concerned, travel cost efficients are statistically significant at the 5%
appears in 11 of the 14 models, and the elasticity level, and in all cases (apart from one) the
estimates range f r o m - 0 . 0 4 t o - 4 . 3 , with the coefficient of determination takes the value 0.9
median value being - 0.5. The three cases where or above. However, in the light of the lack of
demand is most responsive to own travel cost are testing to see whether or not the models are well
those models which incorporate statistically sig- specified and recent developments in economet-
nificant substitute travel costs. Destination cost ric methodology which focus on the specification
features in 12 models, and the elasticity esti- of the dynamic structure of time series, the
mates range f r o m - 0.05 t o - 1.5, and with the quality of the empirical results obtained is ques-
median value b e i n g - 0.7. Tourists thus seem to tionable. The traditional approach implicitly
react in a fairly similar manner to destination assumes that economic data are stationary (at
cost changes and travel cost changes. Exchange least around a deterministic time trend), but it
rate appears as an explanatory variable in only has been shown that economic data are often
five o f the 14 models and in only one case is it non-stationary, and that this can lead to serious
the sole representation of tourists' destination problems with statistics such as t and DW, and
living costs. The estimated elasticities range from measures such as R 2. The studies presented in
0.6 to 2.3, with a median value of 1.8, indicating Table 1 (including the examples examined in
that for those origin-destination pairs, where more detail in Table 2) are all traditional econo-
exchange rate is important, there is a relatively metric investigations, and hence are susceptible
high sensitivity of demand to exchange rate to the charge of possibly showing spurious re-
changes. This supports the argument that "the suits; tests for non-stationarity have not been
effect of change in exchange rate on foreign carried out. Subject to this potential limitation,
travel is not similar to the effect of differential the substantial volume of research on econo-
rates of inflation" (Artus, 1970, p.605). The metric models of tourism demand appears to
fairly high absolute values obtained for the price have resulted in a considerable gain in our
elasticities again support a priori expectations understanding of the economic and other forces
regarding the luxury nature of foreign tourism. affecting demand.
With regard to substitute prices, travel cost
appears in six of the 14 models but destination 3.1.2. Spatial models
cost in only three. The travel cost values span
the range 0.2 to 6.3, with a median value of 1.6, Gravity models are spatial models and repre-
and the destination cost elasticity values vary sent a particular class of multiple regression
from 0.1 to 3.3, with a median value of 1.1. model which has been widely used in the past to
A wide range of dummy variables features in forecast tourism demand. Gravity models are
 S.F. Win, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 459

based on the gravity law of spatial interaction, were extended tO trips made by people with
which states (in the travel context) that the differing incomes, then immediately the gravity
degree of interaction between two geographic model (1) ceases to apply and a more sophisti-
areas varies directly with the degrees of con- cated theory of travel demand is required. Gravi-
centration of persons in the two areas and ty models can, however, still provide a partial
inversely with the distance separating them. The explanation and starting point for these more
demand functions considered earlier are derived complex models.
from the theory of consumer behaviour whereas More sophisticated formulations of spatial
gravity models evolved as part of the early work models may additionally include, as explanatory
of social physicists who believed that social variables, origin income, the price level in the
phenomena could be explained by physical laws destination, substitute prices, etc. Taking
and the fundamental idea was originally derived logarithms, the final form of the model may thus
by analogy with Newton's gravitational law. The closely resemble the consumer theory-based de-
most common form of travel gravity model is: mand function considered earlier, in spite of the
different framework from which the model was
bl b2 b3
Tij = a ( P i Pj / d q ) (1) derived. These more complicated spatial models
include the intervening opportunities model
where Tq denotes the number of trips taking (Pyers, 1966), the alternative opportunities
place between node i and node j; Pi and Pj are model (Long, 1970) and models of the demand
the populations at nodes i and j, respectively; dq for abstract transport modes (Quandt and
denotes the distance between node i and node j; Baumol, 1966, 1969; Quandt and Young, 1969;
and a, bl, b 2 and b 3 are constants. Young, 1969; Quandt, 1970).
It can be seen that Newton's gravitational law Recent causal multiple regression studies of
is a special case of Eq. (1) with b I = b: = 1 and tourism demand have not directly followed the
b 3 = 2. In some applications of gravity models spatial model approach; in fact, the latest re-
Eq. (1) is generalised so that dq, instead of being search effort on the application of gravity models
just the actual distance between nodes i and j, is to tourism occurred in the mid-1970s (Gordon
considered as a general measure of impedance and Edwards, 1973; Malamud, 1973; Durden
between them and it may be better represented and Silberman, 1975). Furthermore, these were
by the travel time, travel cost or generalised cost all cross-section studies. However, the common-
(that is the travel cost plus the product of the ly accepted multiplicative functional form of the
travel time and the monetary evaluation of tourism demand function follows directly from
time). the gravity model specification. Additionally, the
The gravity law of spatial interaction has been spatial model origins of certain econometric
popular as a tool of spatial analysis because of its models explaining tourism demand can be readi-
empirical regularity, but it suffers from several ly traced (Crampon and Tan, 1973; Witt,
drawbacks. One problem is that, as an analogy, 1980a,b; Kliman, 1981).
it lacks a firm theoretical foundation and this
means, moreover, that there is little theory to 3.2, Time-series methods
help in interpreting the values of the parameters
in the model. When attempts have been made to In general, current research which applies
strengthen the theoretical foundation of the time-series forecasting methods to tourism data
gravity model, it has been shown that the model is only published in the context of forecasting
only applies under very restrictive conditions. performance comparisons and is therefore cov-
For example, Wilson (1967) demonstrates that ered in Section 5 of this paper.
gravity models can only be valid if an homoge- Early time-series studies on tourism forecast-
neous person/trip purpose category is being ing, however, often reported forecasts generated
considered. If, therefore, the area of interest by application of the forecasting technique to
460 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

tourism data and sometimes also an analysis of obtained from an extensive aircraft technology
the accuracy of the forecasts. Univariate Box- questionnaire and a much shorter air traffic
Jenkins forecasting has been quite popular in this projection questionnaire as derived in May 1974.
respect. For example, Geurts et al. (1976) used Of particular interest from the tourism angle
the technique to forecast tourist visits to Hawaii were the (inaccurate) predictions that the rate of
using monthly data and assessed the accuracy of growth in air traffic would decrease slightly on
24 one-month-ahead forecasts in terms of per- account of: (a) airport saturation; (b) noise
centage error. They concluded that "The Box- control at airports; and (c) increased real fuel
Jenkins technique produced a very accurate costs. In fact, English and Kearnan (1976) con-
forecast of tourists coming to Hawaii... the cluded that the confidence level associated with
average forecasting error is 3.50%" (p.8). (How- these predictions was satisfactory up to 1990(!),
ever, the mean absolute percentage error but that beyond then the dispersions in estimates
(MAPE) of the forecasts is in fact 8.3%!) of air travel predictions were too high to provide
Similarly, the Canadian Government Office of a reliable consensus.
Tourism (1977) used the Box-Jenkins technique A second example of Delphi forecasting in
to forecast Canadian international tourist re- tourism relates to a survey which was carried out
ceipts and expenditures (quarterly data) and during the International Symposium on Tourism
visits to Canada from the USA (monthly data). in Washington, DC, in 1979; the tourism experts
Decomposition methods of forecasting also present comprised both academics and prac-
attracted some attention during the 1970s in the titioners. The methodology utilised is described
tourism literature. This is particularly associated by Seely et al. (1980) and the main results are
with the work of BarOn (1972, 1973, 1975) who presented by Kibedi (1981). A comprehensive
used the Census X l l Method (modified to allow list of events likely to influence tourism busines-
for factors such as the effects of festival days) to ses during the 1980s was examined and Kibedi
forecast a range of time series-inward tourist reports the findings under headings such as: built
visits, outward tourist visits, international tour- and urban environment; economic environment;
ism receipts and hotel bed-nights. Seasonality is natural resources environment; social, cultural
a characteristic feature of tourism data and the and political environments; technology environ-
popularity of the Census X l l Method in tourism ment; international environment; travel environ-
applications has mainly been attributable to the ment; tourism information; tourism employment
excellence of its seasonal decomposition. and training; tourism and energy; tourism im-
pacts; and tourism co-operation and co-ordina-
tion. Examples of forecasts generated by the
4. Empirical research on qualitative forecasting study are: (a) natural resources will be rationed
for tourist activities; (b) border formalities will
The number of published studies on quali- be relaxed; and (c) tourism will play an im-
tative forecasting in the tourism field is very portant role in the economic development of
limited. Furthermore, from a methodological most countries.
point of view they are of little interest as they are More recently, a Delphi study was carried out
just standard applications. by Liu (1988), who used the technique to fore-
The Delphi method of forecasting is the quali- cast tourism to Hawaii by the year 2000. Two
tative forecasting method that has attracted the groups of individuals were questioned-destina-
most attention in the tourism literature. An early tion experts, and experts from major travel and
Delphi forecast, of interest to airlines, was tour agencies that supplied tourists to Hawaii.
concerned with predicting air travel and aircraft The forecasts range over issues such as visitor
technology to the year 2000 (English and Kear- arrivals, percentage of domestic arrivals, visitor-
nan, 1976). The study presents the findings to-resident ratio and maximum visitor accommo-
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 461

dation. The experts were supplied with data and 5. Accuracy comparisons
statistical projections regarding tourism in
Hawaii together with the questionnaire, so the There are quite a number of studies which
resulting Delphi forecast combines qualitative purport to compare the accuracy of different
and quantitative techniques. The empirical re- forecasting techniques when applied to tourism
suits show a high degree of consistency between data. However, until recently there were rela-
the forecasts made by local experts and the tively few comparative studies of forecasting
forecasts made by overseas suppliers and also accuracy which examined outside sample fore-
between managerial expectations and statistical casting performance.
projections. Studies which compare the accuracy of fore-
Scenario projection has received only scant casts of tourism demand generated by various
attention in the tourism literature, but a recent techniques are presented in Table 3. Attention is
example was published by Schwaninger (1989). restricted to those studies which focus on outside
Schwaninger discusses probable trends in tour- sample comparisons. It is clear that very little
ism between the years 2000 and 2010, with assessment of relative forecasting performance
particular reference to the industrialised coun- took place until the late 1980s, but that consider-
tries of Western Europe. Various aspects of the able progress has been made since then. This
future, as they affect tourism, are considered- contrasts with the general forecasting literature
economic, sociocultural, ecological, technologi- where substantial empirical forecasting perform-
cal and political-and projections are made under ance comparisons already appeared in the litera-
different headings within each of these ture in the 1970s (e.g. Groff, 1973; Makridakis
categories. Thus, as far as economic aspects are and Hibon, 1979), followed by publication of the
concerned, Schwaninger examines the impact of results of a major forecasting competition in the
changes in: energy costs; the informal economy; early 1980s (Makridakis et al., 1982).
the nature of tourism supply and demand; The ranking of forecasting methods included
seasonality in tourism; and sociodemographics. in the studies in terms of popularity is as follows
Sociocultural aspects include the increasing dif- (number of appearances in parentheses): econo-
ferentiation and pluralisation of tourism de- metric (13), naive 1-no change (10), naive 2-
mand, the increased desire for 'active' holidays, constant growth rate compared with corre-
and the emergence of new specialised markets sponding previous period (8), exponential
and market segments. The main ecological smoothing (8), trend curve analysis (general)
change in the future is expected to be an increas- (7), univariate Box-Jenkins (5), autoregression
ing concern for the environment among the (univariate) (4), Gompertz (specific trend curve
population, including holidaymakers and desti- to allow for product life cycle concept) (4),
nation hosts. As a result, tourists will tend to multivariate Box-Jenkins (1), Delphi (1), classi-
avoid holiday spots which have suffered major cal decomposition (1), Census X l l (1), har-
environmental damage and stricter planning con- monic smoothing (1) and generalised adaptive
trois will be enforced by destinations. Tech- filtering (1).
nological developments will have a marked im- The 13 studies presented in Table 3 which
pact on the way the various components of include econometric forecasts give rise to 48
holidays are sold, leading to considerable substi- cases of tourism forecasting in which the accura-
tution of existing sales channels. Finally, the cy of econometric models is compared with that
political aspects include reducing the statutory of other methods (as opposed to just five cases in
number of working hours and also increased which only non-econometric forecasting tech-
planning and protectionist intervention, particu- niques are compared). Econometric forecasting
larly with regard to efforts to curtail the ecologi- proves to be the most accurate method in just
cally damaging effects of tourism. 29% of the 48 cases, followed by naive 1 (23%).
462 S~F. W~tt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

~ o

X X

;~XX X X

XXXX X X

X X

0 X

X X

XX~4 X X XX

~ x X X XX

< x~

XX

oil! I ~'~-~
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 463

<

o
< ~ X ~ X X ~ X ~ X X ~ X X X ~ X ~ X ~

~ X ~ X ~ X ~ X X X ~ X X X ~ ~ X X

a.

#..8

t~

~o
464 SaY. Wilt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

~~eoo~~
~a,l=

X~XX XX ~ X ~ ~x~

~g

r~

X X ~

~XXX

X X ~ ~ X x E

~d

.8
cD

~I~ ~ ~'~ ~i ~I~o~ !

 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 465

The lack of consistent superior performance of The various forecasting models were estimated
the (relatively complex) econometric models is over the period 1965-80 and used to forecast
disturbing, particularly as the (extremely simple) tourist flows in 1981 and 1982. They were then
no change model performs nearly as well. (The re-estimated using data for the period 1965-81
good performance of the latter model is indica- and used to forecast tourist flows in 1982 and
tive of the presence of inertia and non-stationari- 1983. For the UK and the USA, the estimation
ty in tourism demand.) Furthermore, the econo- period was then extended to 1982 and forecasts
metric forecasts are ex post, which gives them an generated for 1983 and 1984 and, finally, for the
unfair advantage over the other methods. UK the procedure was repeated one year for-
If attention now focuses on those studies ward. Autoregression was selected as a forecast-
which do not include econometric models in the ing technique rather than full A R I M A modelling
set of forecasting techniques considered, it can on account of the limited number of observations
be seen that there are two groups of studies: available for estimating the model.
those that compare two non-naive techniques; The figures in parentheses are rankings (1
and those that compare non-naive with naive denotes most accurate, ..., 7 denotes least accur-
techniques. The comparisons in the former ate). Naive 1 has the lowest MAPE of all the
group show that more complicated methods do forecasting methods for each of the origin coun-
not outperform simpler methods: univariate tries when one-year-ahead forecasts are consid-
Box-Jenkins does not outperform exponential ered. On moving to two-years-ahead forecasts,
smoothing; and multivariate Box-Jenkins does however, autoregression performs best in each
not outperform univariate Box-Jenkins. (How- case. International tourism demand does change
ever, details regarding model selection criteria over time and while a no change forecast may
are not provided in the studies.) On the other prove adequate over a one-year time horizon,
hand, the comparisons in the second group show improved forecasting can generally be achieved
that more complex methods do outperform sim- over a two-year time horizon by taking the
ple naive methods: trend curve analysis gener- dynamics of the time series into account. (Kolb
ates more accurate forecasts than naive 1; and and Steckler (1993) have suggested a test to
exponential smoothing generates more accurate determine whether predictions are statistically
forecasts than both naive 1 and naive 2. significantly better than naive 1 forecasts, but
The most recent comprehensive studies of this has not yet been applied in a tourism
tourism forecasting accuracy comparisons are context.)
those by Martin and Witt (1989) and Witt and Econometric forecasts are not ranked particu-
Witt (1989a, 1991); these are now examined in larly high overall in terms of accuracy, but the
detail. position achieved relative to the other forecast-
ing methods varies considerably according to
5.1. Error magnitude accuracy origin country. This lack of superior perform-
ance occurs in spite of using ex post forecasts for
The forecasting performance of the seven the econometric method.
techniques considered in the Martin and Witt
(1989) study in terms of MAPE is presented in 5.2. Directional change accuracy
Table 4, broken down by origin country and time
horizon. The econometric models used to gener- It is often argued that it is more important to
ate the forecasts are discussed in Martin and be able to forecast the direction of change of a
Witt (1988); the hypothesised models are shown data series rather than its actual level. The
in Table 1 and examples of estimated models are challenge is to examine if it is possible to say,
presented in Table 2. For each origin country, with a reasonable level of certainty, whether
the MAPE was calculated from data on six there will be more or fewer tourist visits next
destinations over 2-5 years. year than this year so that operators in the
466 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

Table 4
Forecasting performance - MAPE - by forecasting method, origin country and forecasting horizon
Forecasting Forecasting Origin country
horizon method
(years) France Germany UK USA
NaNe 1 9.06 (1) 5.80 (1) 12.73 (1) 11.94(1)
Naive 2 12.76 (5) 10.55 (6) 15.73 (2) 14.10 (4)
Autoregression 13.45 (7) 5.96 (2) 15.81 (3) 15.03 (5)
Exponential
smoothing 9.38 (2) 7.29 (3) 16.72 (4) 12.95 (3)
Trend curve
analysis 12.54 (4) 8.33 (4) 22.12 (6) 20.93 (7)
Gompertz 13.41 (6) 9.07 (5) 24.12 (7) 16.48 (6)
Econometrics 10.98 (3) 11.28 (7) 20.91 (5) 12.35 (2)

Naive 1 10.08 (2) 7.59 (2) 18.24 (2) 19.43 (4)

Naive 2 22.13 (7) 20.77 (7) 29.42 (5) 17.34 (3)
Autoregression 10.04 (1) 6.10 (1) 16.75 (1) 10.73 (1)
Exponential
smoothing 10.22 (3) 12.17 (4) 23.24 (3) 20.48 (5)
Trend curve
analysis 17.41 (6) 10.61 (3) 30.04 (6) 29.56 (7)
Gompertz 15.46 (5) 13.03 (5) 30.45 (7) 21.19 (6)
Econometrics 14.07 (4) 15.35 (6) 26.00 (4) 11.27 (2)
Notes: 1 Figures in parentheses denote rankings
2 Cell numbers used in MAPE calculations:
France, Germany one year ahead 18
France, Germany two years ahead 12
USA one year ahead 24
USA two years ahead 18
UK one year ahead 30
UK two years ahead 24
Source: Martin and Witt (1989).

tourism industry at least receive an indication as change. (Pesaran and Timmermann (1992) have
to whether to plan for an increase or decrease in developed a statistical test of forecast accuracy
demand, even if the magnitude of the expected when the focus of the analysis is on the correct
increase or decrease is somewhat uncertain. The prediction of direction of change, but this has
measure of direction of change error used is the not yet been applied in a tourism context.)
percentage of directions of movement forecast The forecasting performances of the seven
correctly. techniques considered in Table 4 are assessed in
The empirical work on directional change terms of directional change accuracy by Witt and
accuracy in tourism just examines a one-year- Witt (1989a) for the same group of origin coun-
ahead time horizon. As the naive 1 model tries. The empirical results are presented in
forecasts no change, it will, in general, generate Table 5. The relative forecasting accuracy of a
neither a correct nor an incorrect forecast of particular method varies considerably according
direction of change. In order to outperform this to origin country. Overall, however, economet-
model, a forecasting method must, therefore, rics yields the most accurate forecasts of direc-
forecast over 50% of directions of movement tional changes, followed by exponential smooth-
correctly (assuming a symmetric loss function); ing, autoregression and the no change model.
clearly if a technique yields only 40% accuracy Furthermore, econometrics does not underper-
on this basis, it would be better to assume no form no change for any origin, and forecasts the
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 467

Table 5
Forecasting performance - percentage of directions of movement forecast correctly - by forecasting method and origin country
(one year forecasting horizon)
Forecasting Origin country All
method origins
France Germany UK USA
Naive 1 50 (3 = ) 50 (4 = 50(3 = ) 50(2=) 200(4)
Naive 2 39 (7) 39 (7) 67(2) 50(2=) 195(5)
Autoregression 60 (2) 67 (2) 49(5) 44(4) 220(3)
Exponential smoothing 67 (I) 50 (4 = 70(1) 43(5) 230(2)
Trend curve analysis 50 (3 = ) 50 (4 = 47(6) 33 (7) 180(7)
Gompertz 50 (3 = ) 61 (3) 33 (7) 42 (6) 186(6)
Econometrics 50 (3 = ) 72 (1) 50(3=) 75 (1) 247(1)
Notes: 1 Figures in parentheses denote rankings
2 Cell numbers used in direction of change error calculations:
France, Germany 18
USA 24
UK 30
Source: Witt and Witt (1989a)

direction of movement of tourism demand cor- ance on moving from size of error to direction of
rectly in over 70% of cases for two origins. It change error, as a measure of forecasting accura-
appears, therefore, that for forecasting direction cy, occurs for econometrics. This could be due to
of change of tourist flows over a one-year time the inertia present in tourism demand (as evi-
horizon, econometric forecasting models, in par- denced by the relatively good forecasting per-
ticular, but also exponential smoothing models formance of naive 1 in terms of error mag-
provide valuable information. It should, how- nitude), which the econometric model fails to
ever, be borne in mind that the econometric accommodate adequately. Econometrics per-
forecasts are ex post and hence have a somewhat forms well in terms of forecasting direction of
unfair advantage. change, but overforecasts the size of the change
As discussed in Section 5.1, naive 1 outper- to such an extent that assuming no change is
forms all other forecasting methods for all origin more accurate.
countries over a one-year time horizon in terms
of error magnitude. Some of the rankings corre- 5.3. Turning point accuracy
spond to some extent across the direction of
change error/size of error criteria but, in gener- Rather than just looking generally at direc-
al, the rankings do not correspond. These results tional accuracy, it may be desirable to examine
are not surprising, since direction of change turning point accuracy. (This could refer to long
error and error magnitude are two completely term trend changes or business cycle turning
different dimensions of forecast error and the points.) Turning points represent a particular
discussions by Armstrong and Collopy (1992) subset of directional changes and it may be of
and Fildes (1992), relating to theoretical and most interest to the various sectors of the tour-
empirical considerations when selecting a fore- ism industry to know whether there is likely to
cast error measure, suggest that different error be a change in the trend of demand for their
measures are likely to give different results. (On products. Thus, for example, if tourism demand
the other hand, it could be argued that a well- has been rising steadily for several years, is this
specified model should do well on both accuracy likely to continue or when is a downturn likely to
criteria.) occur? A tourism firm may wish to know
Overall, the greatest improvement in perform- whether it is appropriate to recapitalise or to
468 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

Table 6
Forecasting performance - percentage of trend changes forecast correctly - by forecasting method and origin country (one year
forecasting horizon)
Forecasting Origin country All
method origins
France Germany UK USA
Naive 1 50 (4 = ) 50 (5) 50 (4) 50 (4) 200 (4)
Naive 2 25 (7) 29 (7) 76 (1) 60 (3) 190 (6)
Autoregression 57 (3) 57 (3 = ) 55 (3) 38 (7) 207 (3)
Exponential smoothing 88 (1) 43 (6) 71 (2) 70 (2) 272 (1)
Trend curve analysis 75 (2) 57 (3 = ) 24 (7) 40 (5 = ) 196 (5)
Gompertz 35 (6) 71 (1 = ) 43 (5) 40 (5 = ) 189 (7)
Econometrics 50 (4 = ) 71 (1 = ) 38 (6) 80 (1) 239 (2)
Notes: 1 Figures in parentheses denote rankings.
2 Cell numbers used in trend change error calculations:
France, Germany 18
USA 24
UK 30
Source: Witt and Witt (1991)

spend m o r e on maintenance with a view to Yn-2 ~>Yn-1

writing off plant and equipment. A good trend
forecast will be of great interest in such a
>y~ and Z > y n - Upturn (UT)
situation. If the business has reliable information
Z <~y~ = No upturn ( N U T )
that a downturn is likely to occur, it m a y well be
o p p o r t u n e not to recapitalise until one sees (3)
where the trend leads.
Turning point error occurs when either the where Yl, Y2, ..., Y~ denote given past m e a s u r e -
forecasting system incorrectly fails to predict a ments of a variable, and Z=--yn+ 1 is the first
change in the trend of a data series (a missed future value of the series.
trend change) or incorrectly predicts a trend As in the case of direction of change error, the
change (a false signal). T r e n d changes can be empirical work on turning point accuracy in
split into downturns and upturns and varying tourism just examines a o n e - y e a r - a h e a d time
n u m b e r s of observations can be employed in the horizon. F u r t h e r m o r e , the naive 1 m o d e l will
definitions (Zellner et al., 1990). The measure of neither generate a correct nor an incorrect fore-
trend change error used is the percentage of cast of trend change. H e n c e (as was the case in
turning points forecast correctly (that is the the previous section) in order to o u t p e r f o r m the
percentage of downturns, no downturns, upturns naive 1 model, a forecasting m e t h o d must
and no upturns forecast correctly). achieve over 50% accuracy in terms of turning
point error.
Yn-2 < Y n - 1 Witt and Witt (1991) examine the forecasting
performance of the seven forecasting techniques
considered in Tables 4 and 5 across the same set
<yn and Z<y.~- D o w n t u r n (DT)
of origin countries in terms of turning point
Z ~>y~ =- No downturn ( N D T )
accuracy; the results are summarised in T a b l e 6.
(2) As was the case for directional change accuracy,
the relative forecasting p e r f o r m a n c e of a par-
and ticular method varies considerably according to
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 469

origin country. Overall, however, exponential By examining the major econometric studies
smoothing yields the most accurate forecasts of which have taken place over the last 30 years, it
trend changes followed by econometrics, au- has been possible to trace improvements in
toregression and naive 1. Each of the first three model building. For example, the inclusion of
methods underperforms naive 1 for one origin, substitute prices has become much more com-
so none of the techniques is uniformly superior monplace. In addition, a large number of elas-
to the assumption of no change when forecasting ticity estimates are now available. One of the
trend changes for this limited sample. main findings with respect to econometric
It appears that for forecasting turning points in models is that it is not possible to build a single
tourist flows over a one-year time horizon, model which is appropriate for all origin-desti-
exponential smoothing models in particular, but nation pairs. Certain explanatory variables in-
also econometric models, provide valuable in- fluence tourism demand for some origin-destina-
formation. Although autoregression also outper- tion pairs but not others, and the estimated
forms naive 1 for more origins than it under- coefficients also vary considerably across tourist
performs, the performance never exceeds 57% flows. However, it is clear that virtually no
of trend changes forecast accurately. Further- attention has been paid to improving model
more, when the rankings of the seven forecasting building techniques. In particular, the consider-
methods are averaged over the four origin coun- able advances in econometric methodology dur-
tries, the following rank order is obtained for ing recent years have largely been ignored. It is
those methods which achieve an average ranking essential for future econometric studies of tour-
higher than naive 1, starting with the most ism demand to take on board these develop-
accurate: exponential smoothing (2.75), econo- ments, in particular in the areas of diagnostic
metrics (3.00) and autoregression (4.00), where checking, error correction models and co-inte-
the figures in parentheses denote average rank- gration. The lack of diagnostic checking in the
ings. (The average ranking for naive 1 is 4.25.) econometric studies considered clearly limits the
This provides further support for the conclusion usefulness of the empirical results.
that exponential smoothing and econometric As far as forecasting technique comparisons
forecasts of changes in trends of tourist flows are concerned, it has been shown that the no
should be of practical value to planners in the change model and autoregression consistently
international tourism industry. Again, it should, outperform the constant growth rate model,
however, be borne in mind that the econometric exponential smoothing, trend curve analysis,
forecasts are ex post. Gompertz and econometrics when accuracy is
There are marked similarities between the measured in terms of error magnitude. (The
rankings in terms of the direction of change and results regarding the superiority of no change
turning point error criteria. However, a com- over econometrics are in agreement with findings
parison of the empirical results obtained for the obtained by Witt (1992a) when analysing the
trend change error criterion with those obtained accuracy of commercially produced econometric
for the MAPE criterion shows little similarity of forecasts of tourism demand.) The implications
rankings. of this result for the practitioner seem to be that
he/she should ignore the more complicated,
costly and time-consuming forecasting methods
6. Summary and conclusions and just use last year's demand as the forecast
for this year, i.e. assume no change for one-year-
Although in theory many techniques are avail- ahead forecasts. When looking two years ahead,
able for forecasting tourism demand, in practice the practitioner may be well advised to expend
attention has focused on just a few of these just a little more effort and try to build an
techniques in the empirical literature on tourism autoregressive model to use for forecasting, as
demand forecasting. this method consistently performed relatively
470 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

better over the two-year horizon than any other adopted favours econometric forecasting (see
method. However, it should be noted that these Fildes, 1985). On the other hand, the result that
results are based on data relating to tourist flows naive 1 outperforms more complex univariate
between established generators and destinations. methods supports the M2-Competition findings
Clearly, the conclusions regarding relative per- (Makridakis et al., 1993) that more sophisticated
formance may be different if attention focuses forecasting methods rarely outperform simple
on, say, the rapidly developing destination coun- ones with highly volatile series (as is the case
tries in Asia, where the no change model may generally for international tourism).
well be relatively less accurate. Furthermore, the Danaher and Brodie (1992) have examined
reservations already discussed with regard to the the predictive accuracy of simple versus complex
econometric models used to generate the fore- market share models from both theoretical and
casts should be borne in mind. It may well be empirical standpoints, and conclude that "The
that econometric forecasts, using the most up-to- number of estimation points and number of
date methodological developments, would be parameters are very important ... complex
more accurate. In particular, the superior per- model[s] should always be used ... when the
formance of the autoregressive model over a number of estimation points is reasonably large,
two-year-ahead time horizon suggests dynamic since the drop in prediction accuracy due to
mis-specification. parameter estimation is not significant" (p.622).
The relative accuracy of naive 1 is in some In the case of the tourism demand forecasting
ways very perplexing, as demand clearly does models under consideration, between 16 and 19
change from year to year. (Similar results have, observations were used to estimate the models
however, been obtained in other areas; for and as many as nine parameters are present in
example, Brodie and de Kluyver (1987) showed the econometric models. Clearly, therefore, the
that naive forecasting will often yield better low number of degrees of freedom could be a
results than econometric models when predicting major cause of the relatively poor forecasting
market share.) International tourism demand is performance of the econometric models.
highly volatile and is influenced by many factors, The failure of econometric forecasting to out-
such as economic growth, exchange rates and air perform simple extrapolative models may be a
fares, and econometric models used for tourism result of any of a number of other factors.
forecasting explicitly take into account the im- Although considerable effort has gone into im-
pact on demand of changes in these causal proving the specification of the explanatory vari-
variables. Both the superiority of univariate ables over the last 30 years, there has not been a
time-series forecasting methods over economet- corresponding improvement in model specifica-
ric forecasting and the superiority of naive 1 over tion techniques. There has been a failure on the
the more complex univariate methods seem part of tourism forecasters to adopt the main-
surprising at first. Indeed, the former result stream developments in econometric methodolo-
contrasts with the findings of Fildes (1985) who gy which have occurred over the period; seem-
discovered, when examining studies which com- ingly 'good' model estimation results may have
pared short-term econometric and extrapolative been spurious and, hence, poor selection of an
forecasts, that in a large majority of cases the appropriate forecasting model may have re-
econometric forecasts were more accurate. The suited. Other possible explanations for the fail-
result appears to be even more surprising given ure of econometric forecasting include the unre-
that ex post forecasts are used for the econo- liability of data on tourism demand itself as well
metric method which tend to give it an unfair as on several of the explanatory variables, and
advantage over the other methods; in practice it the fact that holiday decision making is rather
is unlikely that econometric forecasts would have complex and so many factors influence the level
the same degree of accuracy. However, it is not of tourism demand that any econometric model
necessarily the case that the ex post methodology must be a very great simplification.
 S.F. Witt, C.A. Witt I International Journal of Forecasting 11 (1995) 447-475 471

A major advantage of econometric models accuracy chosen. The rankings according to

over univariate time-series models is that a percentage of directions of movement forecast
company in the international tourism industry correctly and percentage of turning points fore-
may be able to use an econometric forecasting cast correctly are similar. However, a compari-
system to explore the consequences of alter- son of MAPE rankings with direction of change
native future policies on tourism demand ('what error/turning point error rankings indicates little
if' forecasting), which is not possible with non- similarity between the two groups. These results
causal models. As Makridakis (1986) notes, "the are in agreement with those obtained by Arm-
purpose of econometric models is not purely strong and Collopy (1992), and Fildes (1992)
forecasting. Instead, they attempt to explain argues for the adoption of a variety of error
economic or business phenomena and increase measures when presenting forecasting accuracy
our understanding of relationships between and results. As assessment of the relative perform-
among variables. To this direction econometric ance of different tourism forecasting techniques
models provide unique information not available is highly dependent upon the choice of accuracy
by time series methods" (p.17). Clearly, though, measure, the tourism forecasting requirement
if an econometric model is mis-specified it will must be considered carefully before deciding on
offer poor advice. a forecasting method. As Makridakis (1993)
When accuracy is measured in terms of direc- notes, "the 'best' method will have to be related
tion of change error, it is possible to outperform to the purpose of forecasting-its value for im-
the no change model for one-year-ahead fore- proving decision making and the needs and
casts; i.e. it is possible to forecast the direction concerns of the person or situation using the
of change of tourism demand with some degree forecasts" (p.527).
of accuracy. (The results regarding the superiori- Makridakis (1986) points out that "The at-
ty of econometrics over no change are in agree- titude that a single method can forecast well
ment with findings obtained by Witt (1992a) for across all situations is not supported by the
commercially produced forecasts.) Firms in the empirical evidence" (p.22). Furthermore, Fildes
tourism industry can, therefore, benefit by know- and Lusk (1984) argue that "no reasonable
ing whether to plan for an increase or decrease forecaster can identify the 'best' method from
in demand. It has also been shown that it is the various forecasting competitions and adopt
possible to outperform the no change model that method for his/her specific forecasting prob-
when forecasting accuracy is defined in terms of lem .... Typically, the forecaster should consider
turning point error, i.e. it is possible to forecast a range of methods, and analyse their compara-
when there will be changes in trends of tourism tive performance over a random sample of those
demand with a reasonable level of accuracy. series of interest" (p.429). This review has
Once a trend in tourist visits has become estab- drawn on the empirical tourism literature in
lished, firms can benefit by knowing when this is order to provide appropriate guidelines for the
likely to alter, i.e. when a turning point is likely tourism demand forecasting situation.
to occur.
Exponential smoothing and econometric fore-
casts are valuable in terms of indicating the
direction of change of tourism demand and References
changes in its trends over a one-year-ahead time
horizon. These forecasts should be of practical Arbel, A. and S.A. Ravid, 1983, An industry energy price
value to those involved in planning in the inter- impact model: The case of the hotel industry, Applied
national tourism industry. Economics 15(6), 705-714.
Archer, B.H., 1980, Forecasting demand: Quantitative and
The results regarding the relative forecasting intuitive techniques, International Journal of Tourism
accuracy of the various techniques differ con- Management 1(1), 5-12.
siderably according to the measure of forecasting Archer, B.H., 1987, Demand forecasting and estimation, in:
472 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

J.R.B. Ritchie and C.R. Goeldner, eds., Travel, tourism international tourism to Australia: A regression analysis,
and hospitality research (Wiley, New York) 77-85. Tourism Management 13, 196-208.
Armstrong, J.S. and F. Collopy, 1992, Error measures for Danaher, P.J. and R.J. Brodie, 1992, Predictive accuracy of
generalizing about forecasting methods: Empirical com- simple versus complex econometric market share models:
parisons, International Journal of Forecasting 8, 69-80. Theoretical and empirical results, International Journal of
Artus, J.R., 1970, The effect of revaluation on the foreign Forecasting 8, 613-626.
travel balance of Germany, IMF Staff Papers 17,602-617. Deaton, A. and J. Muellbauer, 1980, An almost ideal
Artus, J.R., 1972, An econometric analysis of international demand system, American Economic Review 70, 312-326.
travel, IMF Staff Papers 19, 579-614. Durden, G.C. and J. Silberman, 1975, The determinants of
Ashley, R., 1983, On the usefulness of macroeconomic Florida tourist flows: A gravity model approach, Review of
forecasts as inputs to forecasting models, Journal of Regional Studies 5(3), 31-41.
Forecasting 2, 211-223. Edwards, A., 1985, International tourism forecasts to 1995:
EIU special report no. 188 (Economist Publications, Lon-
BarOn, R.R.V., 1972, Seasonality in tourism-Part 1, Interna-
don).
tional Tourism Quarterly, Special Article No.6.
English, J.M. and G.L. Kearnan, 1976, The prediction of air
BarOn, R.R.V., 1973, Seasonality in tourism-Part 2, Interna- travel and aircraft technology to the year 2000 using the
tional Tourism Quarterly, Special Article No.6. Delphi method, Transportation Research 10, 1-8.
BarOn, R.R.V., 1975, Seasonality in tourism, Technical Fildes, R., 1985, Quantitative forecasting-the state of the
Series No.2 (Economist Intelligence Unit, London). art: Econometric models, Journal of the Operational
BarOn, R.R.V., 1984, Forecasting tourism and travel series, Research Society 36(7), 549-580.
Problems of Tourism 3, 24-39. Fildes, R., 1992, The evaluation of extrapolative forecasting
Barry, K. and J. O'Hagan, 1972, An econometric study of methods, International Journal of Forecasting 8, 81-98.
British tourist expenditure in Ireland, Economic and Social Fildes, R. and E.J. Lusk, 1984, The choice of a forecasting
Review 3(2), 143-161. model, OMEGA 12, 427-435.
Blackwell, J., 1970, Tourist traffic and the demand for Fritz, R.G., C. Brandon and J. Xander, 1984, Combining
accommodation: Some projections, Economic and Social time-series and econometric forecast of tourism activity,
Review 1(3), 323-343. Annals of Tourism Research 11(2), 219-229.
Bond, M.E. and J.R. Ladman, 1972, International tourism Fujii, E. and J. Mak, 1981, Forecasting tourism demand:
and economic development: A special case for Latin Some methodological issues, Annals of Regional Science
America, Mississippi Valley Journal of Business and Econ- 15, 72-83.
omics 8(1), 43-55. Gapinski, J.H. and H.P. Tuckman, 1976, Travel demand
Brodie, R.J. and C.A. de Kluyver, 1987, A comparison of functions for Florida bound tourists, Transportation Re-
the short term forecasting accuracy of econometric and search 10, 267-274.
naive extrapolation models of market share, International Geurts, M., 1982, Forecasting the Hawaiian tourist market,
Journal of Forecasting 3, 423-437. Journal of Travel Research 21(1), 18-21.
Brooke, M.Z., P.J. Buckley and S.F. Witt, 1985, The Geurts, M.D. and I.B. Ibrahim, 1975, Comparing the Box-
international travel and tourism forecast (Industry Fore- Jenkins approach with the exponentially smoothed fore-
casts, London). casting model: Application to Hawaii tourists, Journal of
Calantone, R.J., C.A. Di Benedetto and D. Bojanic, 1987, Marketing Research 12, 182-188.
A comprehensive review of the tourism forecasting litera- Geurts, M.D., T.A. Buchman and I.B. Ibrahim, 1976, Use
ture, Journal of Travel Research 26(2), 28-39. of the Box-Jenkins approach to forecast tourist arrivals,
Canadian Government Office of Tourism, 1977, Methodolo- Journal of Travel Research 14(4), 5-8.
gy for short-term forecasts of tourism flows: Research Gordon, I.R. and S.L. Edwards, 1973, Holiday trip genera-
report no. 4 (Canadian Government, Office of Tourism, tion, Journal of Transport Economics and Policy 7(2),
Ottawa, Canada). 153-168.
Chadee, D. and Z. Mieczkowski, 1987, An empirical analysis Gray, H.P., 1966, The demand for international travel by the
of the effects of the exchange rate on Canadian tourism, United States and Canada, International Economic Review
Journal of Travel Research 26(1), 13-17. 7(1), 83-92.
Choy, D.J.L., 1984, Forecasting tourism revisited, Tourism Groff, G.K., 1973, Empirical comparisons of models for
Management 5(3), 171-176. short-range forecasting, Management Science 20, 22-31.
Cline, R.S., 1975, Measuring travel volumes and itineraries Gujarati, D.N., 1988, Basic econometrics, 2nd edn.
and forecasting future travel growth to individual Pacific (McGraw-Hill, New York).
destinations, in: S.P. Ladany, ed., Management science Gunadhi, H. and C.K. Boey, 1986, Demand elasticities of
applications to leisure-time operations (North-Holland, tourism in Singapore, Tourism Management 7(4), 239-
New York) 134-145. 253.
Crampon, L.J. and K.T. Tan, 1973, A model of tourism flow Jud, G.D., 1974, Tourism and economic growth in Mexico
into the Pacific, Tourist Review 28(3), 98-104. since 1950, Inter-American Economic Affairs 28(1), 19-
Crouch, G.I., L. Schultz and P. Valerio, 1992, Marketing 43.
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 473

Jud, G.D. and H. Joseph, 1974, International demand for Martin, C.A. and S.F. Witt, 1989, Forecasting tourism
Latin American tourism, Growth and Change (Jan.) 24- demand: A comparison of the accuracy of several quantita-
31. tive methods, International Journal of Forecasting 5, 1-13.
Kibedi, G., 1981, Future trends in international tourism, Means, G. and R. Avila, 1986, Econometric analysis and
Tourist Review 36(1), 3-6. forecasts of US international travel: Using the new TRAM
Kliman, M.L., 1981, A quantitative analysis of Canadian model, World Travel Overview 1986/87 (American Ex-
overseas tourism, Transportation Research 15A(6), 487- press Publishing, New York) 90-107.
497. Means, G. and R. Avila, 1987, An econometric analysis and
Kolb, R.A. and H.O. Steckler, 1993, Are economic forecasts forecast of US travel and the 1987 TRAM model update,
significantly better than nai've predictions? An appropriate World Travel Overview 1987/88 (American Express Pub-
test, International Journal of Forecasting 9, 117-120. lishing, New York) 102-123.
Means, G. and R. Avila, 1988, Globalization redirects the
Kwack, S.Y., 1972, Effects of income and prices on travel
travel industry: The 1988 TRAM analysis, World Travel
spending abroad 1960 111-1967 IV, International Economic
Overview 1988/89 (American Express Publishing, New
Review 13(2), 245-256.
York) 86-100.
Latour, B., 1974, Forecasting by a major airline, in; The O'Hagan, J.W. and M.J. Harrison, 1984, Market shares of
measurement of tourism (IUOTO Seminar on Forecasting,
US tourist expenditure in Europe: An econometric analy-
Haslemere) 42-47. sis, Applied Economics 16, 919-931.
Little, J.S., 1980, International travel in the US balance of Oliver, F.R., 1971, The effectiveness of the UK travel
payments, New England Economic Review (May/June) allowance, Applied Economics 3, 219-226.
42-55. Papadopoulos, S.I. and S.F. Witt, 1985, A marketing analysis
Liu, J.C., 1988, Hawaii tourism to the year 2000, Tourism of foreign tourism in Greece, in: S. Shaw, L. Sparks and
Management 9(4), 279-290. E. Kaynak, eds., Marketing in the 1990s and beyond:
Loeb, P., 1982, International travel to the United States: An Proceedings of Second World Marketing Congress (Uni-
economic evaluation, Annals of Tourism Research 9(1), versity of Stirling, Stirling) 682-693.
5-20. Paraskevopoulos, G., 1977, An econometric analysis of
Long, W.H., 1970, The economics of air travel gravity international tourism, Lecture series 31 (Center of Plan-
models, Journal of Regional Science 10(3), 353-363, ning and Economic Research, Athens).
McCuUagh, P. and J.A. Nelder, 1989, Generalized linear Pesaran, M.H. and A. Timmermann, 1992, A simple non-
models, 2nd edn. (Chapman and Hall, London). parametric test of predictive performance, Journal of
Makridakis, S., 1986, The art and science of forecasting-an Business and Economic Statistics 10, 461-465.
assessment and future directions, International Journal of Peterson, G.L., D.J. Stynes and J.R. Arnold, 1985, The
Forecasting 2, 15-39. stability of a recreation demand model over time, Journal
Makridakis, S., 1993, Accuracy measures: Theoretical and of Leisure Research 17(2), 121-132.
practical concerns, International Journal of Forecasting 9, Pyers, C.E., 1966, Evaluation of intervening opportunities
527-529. trip distribution model, Highway Research Record 114,
Makridakis, S. and M. Hibon, 1979, Accuracy of forecasting: 71-88.
An empirical investigation (with discussion), Journal of the Quandt, R.E., 1970, ed., The demand for travel: Theory and
Royal Statistical Society A 142, 97-145. measurement (Lexington, MA).
Makridakis, S., A. Andersen, R. Carbone, R. Fildes, M. Quandt, R.E. and W.J. Baumol, 1966, The demand for
Hibon, R. Lewandowski, J. Newton, E. Parzen and R. abstract transport modes: Theory and measurement, Jour-
Winkler, 1982, The accuracy of extrapolation (time series) nal of Regional Science 6, 13-26.
methods: Results of a forecasting competition, Journal of Quandt, R.E. and W.J. Baumol, 1969, The demand for
Forecasting 1, 111-154. abstract transport modes: Some hopes, Journal of Regional
Makridakis, S., C. Chatfield, M. Hibon, M. Lawrence, T. Science 9, 159-162.
Mills, K. Ord and L.F. Simmons, 1993, The M2-Competi- Quandt, R.E. and K.H. Young, 1969, Cross-sectional travel
tion: A real time judgmentally based forecasting study, demand models: Estimation and tests, Journal of Regional
International Journal of Forecasting 9, 5-22. Science 9, 201-214.
Malamud, B., 1973, Gravity model calibration of tourist Quayson, J. and T. Var, 1982, A tourism demand function for
travel to Las Vegas, Journal of Leisure Research 5(4), the Okanagan, BC, Tourism Management 3(2), 108-115.
23-33. Rosensweig, J.A., 1988, Elasticities of substitution in Carib-
Martin, C.A. and S.F. Witt, 1987, Tourism demand forecast- bean tourism, Journal of Development Economics 29(1),
ing models: Choice of appropriate variable to represent 89-100.
tourists' cost of living, Tourism Management 8(3), 233- Schwaninger, M., 1989, Trends in leisure and tourism 2000-
246. 2010: Scenario with consequences for planners, in: S.F.
Martin, C.A. and S.F. Witt, 1988, Substitute prices in models Witt and L. Moutinho, eds., Tourism marketing and
of tourism demand, Annals of Tourism Research 15(2), management handbook (Prentice-Hall, Hemel Hempstead,
255-268. UK) 599-605.
474 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475

Seely, R.L., H. Iglars and D. Edgell, 1980, Utilizing the Witt, S.F., 1991a, How successful are commercial tourism
Delphi technique at international conferences: A method forecasting models? in: F. Bradley, ed., Marketing thought
for forecasting international tourism conditions, Travel around the world: Proceedings of European Marketing
Research Journal 1, 30-35. Academy 20th Annual Conference (University College
Sheldon, P.J. and T. Var, 1985, Tourism forecasting: A Dublin) 1573-1588.
review of empirical research, Journal of Forecasting 4(2), Witt, S.F., 1991b, Assessing the accuracy of published
183-195. econometric forecasts of international tourism demand,
Smeral, E., S.F. Witt and C.A. Witt, 1992, Econometric Tourism-building credibility for a credible industry:
forecasts: Tourism trends to 2000, Annals of Tourism Proceedings of the Travel and Tourism Research Associa-
Research 19(3), 450-466. tion 22nd Annual Conference (University of Utah, Salt
Smith, A.B. and J.N. Toms, 1967, Factors affecting demand Lake City, UT) 473-483.
for international travel to and from Australia, Occ. paper Witt, S.F., 1992a, The track records of tourism forecasting
11 (Bureau of Transport Economics, Canberra). services, in: P. Johnson and B. Thomas, eds., Choice and
Summary, R., 1987, Estimation of tourism demand by demand in tourism (Manseli, London) 209-220.
multivariable regression analysis: Evidence from Kenya, Witt, S.F., 1992b, Tourism forecasting: How well do private
Tourism Management 8(4), 317-322. and public sector organizations perform? Tourism Manage-
Tie-Sheng, W. and G. Li-Cheng, 1985, Domestic tourism ment 13(1), 79-84.
development in China: A regression analysis, Journal of Witt, S.F. and C.A. Martin, 1987a, Econometric models for
Travel Research 24(2), 13-16. forecasting international tourism demand, Journal of
Uysal, M. and J.L. Crompton, 1984, Determinants of de- Travel Research 15(3), 23-30.
mand for international tourist flows to Turkey, Tourism Witt, S.F. and C.A. Martin, 1987b, Measuring the impacts of
Management 5(4), 288-297. mega-events on tourism flows, The role and impact of
Uysal, M. and J.L. Crompton, 1985, An overview of ap- mega-events and attractions on regional and national
proaches used to forecast tourism demand, Journal of tourism development: Proceedings of International As-
Travel Research 23(4), 7-15. sociation of Scientific Experts in Tourism (AIEST) 37th
Van Doorn, J.W.M., 1982, Can futures research contribute to Congress (AIEST, St. Gallen, Switzerland) 213-221.
tourism policy? Tourism Management 3(3), 149-166. Witt, S.F. and C.A. Martin, 1987c, International tourism
Van Doom, J.W.M., 1984, Tourism forecasting and the demand models-inclusion of marketing variables, Tourism
policymaker: Criteria of usefulness, Tourism Management Management 8(1), 33-40.
5(1), 24-39. Witt, S.F. and C.A. Martin, 1989, Demand forecasting in
Vanhove, N., 1980, Forecasting in tourism, Tourist Review tourism and recreation, in: C. Cooper, ed., Progress in
35(3), 2-7. tourism, recreation, and hospitality management, Vol. 1
Wandner, S.A. and J.D. Van Erden, 1980, Estimating the (Belhaven Press, London) 4-32.
demand for international tourism using time series analy- Witt, S.F. and R.A.C. Rice, 1981, An empirical comparison
sis, in: D.E. Hawkins, E.L. Shafer and J.M. Rovelstad, of alternative forecasting methods as applied to the UK
eds., Tourism planning and development issues (George foreign holiday market, in: J. Sparkes and S.F. Witt, eds.,
Washington University, Washington DC) 381-392. Business forecasting for financial management (Barmarick,
Western European Airports Association, 1976, Air Traffic Hull, UK) 16-20.
Forecasts for Western Europe (WEEA, London). Witt, C.A. and S.F. Witt, 1989a, Measures of forecasting
White, K.J., 1985, An international travel demand model: accuracy: Turning point error v. size of error, Tourism
US travel to Western Europe, Annals of Tourism Research Management 10(3), 255-260.
12(4), 529-546. Witt, S.F. and C.A. Witt, 1989b, An empirical assessment of
Wilson, A.G., 1967, A statistical theory of spatial distribu- relative forecasting performance based on directional ac-
tion models, Transportation Research 1(3), 253-269. curacy, Tourism research-globalization, the Pacific Rim
Witt, S.F., 1980a, An abstract mode-abstract (destination) and beyond: Proceedings of Travel and Tourism Research
node model of foreign holiday demand, Applied Econ- Association 20th Annual Conference (University of Utah,
omics 12(2), 163-180. Salt Lake City, UT) 401-407.
Witt, S.F., 1980b, An econometric comparison of UK and Witt, S.F. and C.A. Witt, 1991, Tourism forecasting: Error
German foreign holiday behaviour, Managerial and Deci- magnitude, direction of change error and trend change
sion Economics 1(3), 123-131. error, Journal of Travel Research 30(2), 26-33.
Witt, S.F., 1983, A binary choice model of foreign holiday Witt S.F., M. Dartus and A.M. Sykes, 1992a, Modeling
demand, Journal of Economic Studies 10(1), 46-59. conference tourism, Tourism partnerships and strategies-
Witt, S.F., 1990, Cash flow forecasting in the international merging vision with new realities: Proceedings of Travel
tourism industry, in: R. Aggarwal and C.F. Lee, eds., and Tourism Research Association 23rd Annual Confer-
Advances in financial planning and forecasting 4, Part B: ence (TI'RA, Wheat Ridge, CO) 116-124.
International dimensions of financial management (JAI Witt, S.F., G.D. Newbould and A.J. Watkins, 1992b, Fore-
Press, Greenwich, USA) 229-244. casting domestic tourism demand: Application to Las
 S.F. Witt, C.A. Witt / International Journal of Forecasting 11 (1995) 447-475 475

Vegas arrivals data, Journal of Travel Research 31(1), University of Wales, Swansea. He is co-author of Por~olio
36-41. Theory and Investment Management, Practical Financial
Young, K.H., 1969, An abstract mode approach to the Management, Practical Business Forecasting, The Manage-
demand for travel, Transportation Research 3, 443-461. ment of International Tourism and Modeling and Forecasting
Demand in Tourism, and co-editor of the Tourism Marketing
Zellner, A., C. Hong and G.M. Gulati, 1990, Turning points and Management Handbook. He has published widely in
in economic time series, loss structures and Bayesian journals on tourism forecasting, and is on the editorial boards
forecasting, in: S.J. Press and A. Zellner, eds., Bayesian of Tourism Management, the Journal of Travel Research, the
and likelihood methods in statistics and econometrics: Journal of International Consumer Marketing and the Journal
Essays in honor of George A. Barnard (North-Holland, of Euromarketing.
Amsterdam) 371-393.
Christine WIqT is Lecturer in Operations Management at the
University of Bradford Management Centre. She has pub-
lished numerous journal articles on tourism forecasting and is
co-author of Modeling and Forecasting Demand in Tourism.
Biographies: Stephen WIqT is Lewis Professor of Tourism Other research interests include the application of operations
Studies in the European Business Management School at the management techniques to service industries.