Comparison of Near Neighbour and Neural Network in Travel Forecasting

• Author: Elena Olmedo
• Published date: 01 April 2016
• Date: 01 April 2016
• DOI: http://doi.org/10.1002/for.2370

Comparison of Near Neighbour and Neural Network

in Travel Forecasting

ELENA OLMEDO*

Departamento Economia Aplicada I, Universidad de Sevilla, Spain

ABSTRACT

In this paper we confirm the existence of nonlinear dynamics in a time series of airport arrivals. We subsequently

propose alternative non-parametric forecasting techniques to be used in a travel forecasting problem, emphasizing

the difference between the reconstruction and learning approach. We compare the results achieved in point prediction

versus sign prediction. The reconstruction approach offers better results in sign prediction and the learning approach in

point prediction. Copyright © 2015 John Wiley & Sons, Ltd.

key words forecasting; neural networks; reconstruction; travel; nonlinearity

INTRODUCTION

The travel and tourism industry is a key economic sector in Spain, so forecasting visitor arrivals is crucial for better

tourism planning and administration. Hotels, restaurants and transport lines require accurate predictions to avoid

failures in providing services. Also, policymakers rely on accurate forecasts to set national, regional and local tourism

strategies.

Tourism demand studies tend to use linear parametric time series forecasting models. The most popular are the

naïve method (Burger et al., 2001; Chu, 2004), the exponential smoothing model (Law, 2000b; Burger et al.,

2001; Kim and Ngo, 2001) and autoregressive integrated moving average (ARIMA) models (Qu and Zhang, 1996;

Law, 2000a, 2004; Goh and Law, 2002).

However, if the time series is the result of a nonlinear dynamic system, linear forecasting will fail. The pre-

diction is improved if we use nonlinear models such as the piecewise linear model (Chu, 2011) or regime

switching model (Huarng et al., 2011). The main problems in these cases are proper model selection and data

dependency.

Many researchers have turned to complex nonlinear methods such as support vector machines (Pai and Hong,

2005; Pai et al., 2006; Chen and Wang, 2007), fuzzy logic (Wang, 2004), genetic programming (Álvarez-Diaz

et al., 2009) and neural networks (Law and Au, 1999; Burger et al., 2001; Cho, 2003; Kon and Turner, 2005;

Palmer et al., 2006), but very few authors have investigated the assumption of linearity to assess whether the use

of nonlinear models is justified.

In this paper, we confirm the existence of nonlinear dynamics using different nonlinear tests. After that, we

compare the results of two different nonlinear forecasting techniques. The comprehension approach or phase-space

reconstruction (PSR), by means of reconstruction theorem, is suitable when the unknown underlying system has a

low dimension. The learning approach, by virtue of artificial neural networks (ANN), is more appropriate when

the system is not simple enough to be reconstructed.

Both approaches use nonlinear functions, but the reconstruction approach is a local method whereas the learning

approach is a global method that uses the whole domain to make forecasts.

This paper is structured in three sections. In the first section, we verify the presence of nonlinearity. In the second

section, we briefly explain the forecasting methods, while in the last section we apply the methods to the forecasting

of airport arrivals at Palma de Mallorca airport (PMI).

DETECTION OF NONLINEARITY

Before forecasting the time series using nonlinear techniques, we should confirm the presence of nonlinearity in data

(Barnett et al., 1995, 1996, 1997; Ashley and Patterson, 2000). We propose the use of the BDS test (Brock et al.,

1996) and Kaplan test (Kaplan, 1994, 1995) because of their generality. These tests come from chaos theory but

can be generally applied.

*Correspondence to: Elena Olmedo, Departamento Economia Aplicada I, Universidad de Sevilla, Avda. Ramon y Cajal 1, 41018 Sevilla, Spain.

E-mail: olmedo@us.es

Journal of Forecasting,J. Forecast. 35, 217–223 (2016)

Published online 11 November 2015 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/for.2370

Copyright © 2015 John Wiley & Sons, Ltd.