Gaussian processes for daily demand prediction in tourism planning

• DOI: http://doi.org/10.1002/for.2644
• Published date: 01 April 2020
• Author: Dries F. Benoit,Wai Kit Tsang
• Date: 01 April 2020

Received: 25 March 2019 Revised: 20 September 2019 Accepted: 2 December 2019

DOI: 10.1002/for.2644

RESEARCH ARTICLE

Gaussian processes for daily demand prediction in tourism

planning

Wai Kit Tsang Dries F. Benoit

Faculty of Economics and Business

Administration, Ghent University,Ghent,

Belgium

Correspondence

Wai Kit Tsang,Faculty of Economics

and Business Administration, Ghent

University, Tweekerkenstraat 2, 9000

Ghent, Belgium.

Email: waikit.tsang@ugent.be

Funding information

Bijzonder Onderzoeksfonds

Abstract

This study proposes Gaussian processes to forecast daily hotel occupancy at

a city level. Unlike other studies in the tourism demand prediction literature,

the hotel occupancy rate is predicted on a daily basis and 45 days ahead of

time using online hotel room price data. A predictive framework is introduced

that highlights feature extraction and selection of the independent variables.

This approach shows that the dependence on internal hotel occupancy data can

be removed by making use of a proxy measure for hotel occupancy rate at a

city level. Six forecasting methods are investigated, including linear regression,

autoregressive integrated moving average and recent machine learning meth-

ods. The results indicate that Gaussian processes offer the best tradeoff between

accuracy and interpretation by providing prediction intervals in addition to

point forecasts. It is shown how the proposed framework improves managerial

decision making in tourism planning.

KEYWORDS

Gaussian processes, tourism demand forecast, feature extraction, daily data, prediction interval

1INTRODUCTION

Due to the perishable nature of many tourismproducts and

services, accurately forecasting demand is important for

the tourism industry (Archer et al., 1987; Wandner et al.,

1980). Hotel managers assign employees and plan work-

ing schedules depending on the expected number of hotel

bookings. Tour guides and other tourism service providers

plan operations in anticipation of the expected demand.

Accurate tourism demand prediction is crucial for air-

line companies and car rental services to set their prices

optimally (Talluri & VanRyzin, 2006).

Previous tourism demand studies primarily use data

from surveys or governmental statistics, which have a

low temporal frequency (Peng et al., 2014; Song et al.,

2010). Due to the significant time lag between registration

and publication of the data, the practical deployability of

models relying on such data is limited. In addition, such

statistics, often gathered offline, are highly aggregated,

which smooths out irregularities in the temporal pattern.

Tourism demand forecasting studies often use autoregres-

sive integrated moving average (ARIMA) and seasonal

ARIMA (SARIMA) models when there are no significant

structural breaks in the time series (Chu, 2008; Goh & Law,

2002).

For yearly and quarterly data a variety of tourism

demand models are used in the literature. Time series

models, such as the Black–Scholes–Merton (BSM) model

by Greenidge (2001), are used to extract cyclical and

seasonal patterns (Wooldridge, 2015). Static econometric

models such as the gravity model, used by Wei (2007)

to predict inbound tourism demand to China, are ideal

for analyzing elasticities and explanatory variables, but

do not take short-run dynamics into consideration (Song

et al., 2008). Dynamic econometric models, such as the

time varying parameter (TVP) model, perform well in

1-year-ahead forecasting (Witt et al., 2003).

This study contributes to the existing literature by

modeling online hotel room price data to predict not

monthly or yearly, but daily hotel occupancy rates at a

wileyonlinelibrary.com/journal/for ©2019 John Wiley & Sons, Ltd.Journal of Forecasting. 2020;39:551–568. 551

TSANG AND BENOIT

city level using Gaussian processes. Hotel room price data

are publicly available on websites such as agoda.com,

expedia.com, trip.com and booking.com, the latter web-

site being the data source from which the hotel room price

data are scraped for this study. We show how this public

data can allow governments and tourism businesses a solid

approach to estimate tourism demand. In the remainder

of the study we refer to these data as online data. Instead

of using the true hotel occupancy rates, a proxy measure

for hotel occupancy rate at a city level is provided by a

Belgian hotel company in the field of online travel agency

that relied on public hotel room price data that are vali-

dated against internal hotel occupancy data from 576 hotel

accommodations in Brussels. This approach removes the

dependence on internal hotel occupancy data, which is

proprietary and not necessarily available.

In addition, we also show in this study how feature pre-

processing, feature extraction, and selection methods sig-

nificantly boost the predictive performance. This has been

overlooked by existing approaches in the demand predic-

tion literature. Gaussian processes are used to predict the

hotel occupancy rate at a city level and are benchmarked

against linear regression, ARIMA and machine learning

models. Gaussian processes produce prediction intervals,

a measure to quantify uncertainty that policymakers can

use for reliable decision making. Note that, as outlined

in Section 4, these prediction intervals are more flexi-

ble and realistic than similar intervals produced by other

approaches.

In Section 2 the literature is discussed. In Section 3 the

data are explored. Section 4 describes the methodological

framework. In Section 5 the results and findings of the

research study are presented. Finally,in Section 6 the con-

tributions are summarized, pointers for decision makers

are offered, and future research avenues are explored.

2LITERATURE REVIEW

2.1 Temporal granularity

Tourism demand predictions on a daily basis are not com-

mon in the tourism forecasting literature (Peng et al.,

2014). Predictive models in the past were primarily built

using data from offline sources, such as tourism bureaus

that provide explanatory variables aggregated over a given

time period (Song & Li, 2008). While annual and quarterly

data statistics are quite prevalent, more recent research

also uses monthly data for predictions (Gunter & Önder,

2015).

Tourism demand, however, also fluctuates on a daily

basis, as the number of mid-week bookings differs from the

weekend. These variations in daily tourism demand make

effective policy planning more difficult for tourist busi-

nesses and governments. When tourism demand exceeds

supply,decision makers should adequately react to prevent

considerable strain on tourism infrastructure and services;

or when there is a period of lower tourism demand, miss-

ing the time window to invest can lead to loss in profitabil-

ity (Karamustafa & Ulama, 2010; Pegg et al., 2012). Many

tourism products are perishable, such as hotel rooms, air-

line seats, and car rentals. The moment the deadline passes

and a seat remains unfilled, the product, and in conse-

quence its revenue, is permanently lost (Dharmaratne,

1995). Moreover, the revenue is impacted by differing

demands across weekdays (Malasevska & Haugom, 2018)

and price elasticities that change from week to weekend

(Morlotti et al., 2017).

2.2 Tourism demand forecasting

methods

Tourism forecasting models have been studied in a large

variety, from early time series models to more recentones

such as machine learning models. Time series methods,

such as the autoregressive moving average (ARMA) pro-

cess (Box et al., 2015), predict how the future trend will

behave and require no more than one historical data series

(Hamilton, 1994). Time series models are quite suitable

to deal with linear trends, such as Brown's double expo-

nential smoothing (DES) model (Brown, 2004), which has

been used to predict tourist arrivals in Hawaii (Geurts &

Ibrahim, 1975). These models rely on historical demand

only and do not directly link to explanatory variables and,

in consequence, cannot analyze tourist behavior (Song &

Li, 2008). A second group of methods, which is more useful

for decision makers, are econometric models (Song et al.,

2008). The error correction model (ECM), for example,

is capable of assessing both long-run equilibrium and

short-run disequilibrium relationships, and have already

been used to analyze Canadian and Tunisian tourism

(Ouerfelli, 2008; Veloce, 2004). At an international level,

the dynamic almost-ideal demand system (AIDS) model

has been shown to make reliable forecasts for European

tourism demand by UK citizens (De Mello & Fortuna,

2005). A third and final category in the literature is the

group of artificial intelligence (AI) methods (Wang & Hsu,

2008). As machine learning methods are rapidly devel-

oping and big data are becoming more available, these

models are more frequently used for tourism prediction

(Claveria et al., 2016). Chen and Wang (2007) have com-

bined a genetic algorithm (GA) with support vector regres-

sion (SVR) to predict tourist arrivals in China. Although

AI models are specialized in achieving a good predictive

performance (Claveria et al., 2016), trained AI models are

complex and not always interpretable in terms of uncer-

tainty (Burger et al., 2001).

While feature extraction—the conversion of raw data

into a set of useful features—and feature selection—the

552