Incorporating Explanatory Effects of Neighbour Airports in

Forecasting Models for Airline Passenger Volumes

Nilgun Ferhatosmanoglu

1

and Betul Macit

2

1

Department of Industrial and Systems Engineering, University of Turkish Aeronautical Association, Ankara, Turkey

2

Department of Industrial Engineering, Gazi University, Ankara, Turkey

Keywords: Forecasting, Airport Networks, TBATS, Regression with ARIMA Errors, Airline Passenger Volumes,

Neighbour Effects in Modelling.

Abstract: Forecasting airline passenger volumes can be helpful for flight and airport capacity planning. While there

are many parameters affecting the passenger volume, to our knowledge no work has directly studied the

effect of neighbour airports in modelling of passenger volumes. We develop an integrated model for

forecasting the number of passengers arriving/departing an airport, considering the airport’s interactions

with its neighbour airports. In particular, we analyse the time series of the flights arriving to and departing

from two largest airports in Turkey, namely Ankara Esenboga and Istanbul Ataturk Airports, and explore

the interactions between these airports by using them as regressors for each other. We also apply

independent models based on TBATS which was previously proposed in the literature to handle multiple

seasonalities. In our experiments, TBATS performs better than ARIMA for independent modelling, and

TBATS with multiple seasonal periods outperforms TBATS with single seasonality in majority of the cases.

In several cases, the forecasting accuracy increases when the neighbour airports’ traffic data is used in

modeling the passenger volumes.

1 INTRODUCTION

Civil aviation authorities and airline companies need

short and long term forecasts for effective flight and

capacity planning. A wide range of forecasting

models are developed including econometric

modelling, univariate time series modelling, time

series decomposition, non-linear regression models

and gravity models (Scarpel, 2013). While it is

intuitive that the traffic of an airport is not

independent of its neighbour airports, to our

knowledge this is not directly taken into

consideration in modelling and forecasting airport

traffic. Research is needed to investigate how the

traffic model of an airport can incorporate its

‘neighbours’ traffic as they affect each other

possibly with a small time shift. In this paper, we

consider interactions between neighbour airports in

developing time series forecasting models for air

traffic volumes. We handle two neighbour airports

as a dyad in an airport network and compare

independent and neighbour-dependent models to

forecast the number of passengers for particular

routes.

As a case study, we analyse time series patterns

of domestic and international flights arriving to and

departing from Ataturk and Esenboga International

Airports in Turkey over the course of a year. We

propose a neighbour-dependent approach using

regression with ARIMA errors and explore

explanatory relations by regressor time series. For

independent modelling, we consider ARIMA and

TBATS models for developing independent

forecasting models. TBATS (Trigonometric, Box-

Cox transform, ARMA errors, Trend, and Seasonal

components) model was proposed to deal over

parameterization and handle both non-integer period

and dual-calendar effects (De Livera et al., 2011).

ARIMA models enable fitting the patterns in data

with smallest number of estimated parameters.

TBATS handles multiple seasonality which we

observe in Turkish flight data. We elaborate

accuracy performance of TBATS method and

ARIMA models in independent modelling as well.

By comparing the accuracy of independent and

dependent models we are able to explore

contribution of neighbour relations on forecasting

performance.

178

Ferhatosmanoglu, N. and Macit, B.

Incorporating Explanatory Effects of Neighbour Airports in Forecasting Models for Airline Passenger Volumes.

DOI: 10.5220/0005752801780185

In Proceedings of 5th the International Conference on Operations Research and Enterprise Systems (ICORES 2016), pages 178-185

ISBN: 978-989-758-171-7

Copyright

c

2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

In our experiments, TBATS performs better than

ARIMA for independent modelling. TBATS models

with multiple seasonal periods perform better than

models with single seasonality. These results verify

that TBATS model performs well in airline data

with multiple seasonality. Using explanatory time

series of neighbours’ air passenger volumes is found

to be useful in several cases.

The remainder of the paper contains the

following sections. First, we present a literature

overview on air passenger flow problem and the

forecasting methods used in this study. We then

present the proposed methodology and empirical

results. Finally, we conclude with future research

directions.

2 RELATED WORK

Air transportation has achieved a remarkable growth

both worldwide and in Turkey, e.g., the total number

of passengers in Turkey has risen 14.3% in the last

decade (TOBB, 2013). According to

EUROCONTROL forecasts, Turkey will be the

arrival or departure point for the greatest number of

extra flights in the future European airspace by 2035

(EUROCONTROL, 2013). In our study, we generate

forecasting models for the number of air travel

passengers for different routes between Ankara

Esenboga and Istanbul Ataturk Airports.

Several methods have been proposed for

forecasting the number of air travel passengers in the

context of air travel demand, pax growth and air

travel flow. Traditional methods, such as neural

networks, exponential smoothing, Box-Jenkins, and

Holt-Winters, are commonly applied in this context.

Nam et al. use neural networks for predicting

international air passenger volume between US and

Mexico and compare them with regression and

exponential smoothing forecasting models (Nam et

al., 1997). Neural network models are also compared

with well-known Box-Jenkins and Holt-Winters

Methods (Faraway et al., 1998). In an application

study, neural networks are shown to outperform the

traditional econometric approach for forecasting

Brazilian air transport passenger traffic (Alekseev et

al., 2009).

Samagaio and Wolters propose ARIMA and

exponential smoothing models for forecasting the

number of passengers for 2008-2020 to help

decision making for establishing a new airport

(Samagaio et al., 2010). An application of fuzzy

regression model is developed to forecast the

demand of Rhodes airport (Profillidis, 2000).

Grosche et al. propose two gravity models using

geo-economic factors as independent variables for

estimation of airline passenger volume between city

pairs (Grosche et al., 2007). Fildes et al. explore the

relations between air traffic flows to different

countries by pooled ADL model (Flides et al., 2011).

They enhance their models with “world trade”

variable and conclude that pooled ADL model with

“world trade” variable outperformed the alternatives.

Benitez et al. propose a modified Grey Model for

airlines routes pax growth for long lead-time

(Benitez et al., 2013). ARIMAX and SARIMA

based models are recently used to forecast Hong

Kong airport's passenger throughput till 2015 (Tsui

et al., 2014). Time series involved in our analysis

involve multiple seasonal patterns. Hence we use a

recent proposal, TBATS, which handles multiple

seasonality (De Livera et al., 2011).

3 METHODOLOGY

In this section, we highlight the methods for

handling the seasonality from the literature and

introduce the details of our methodology to

incorporate the interactions of two airports into their

forecasting models. In particular, we study Ataturk

and Esenboga Airports in Turkey and investigate if

they affect each other while forecasting their

passenger volumes. The data set of the number of

passengers for Ataturk and Esenboga Airports in

2011 is obtained from the General Directorate of

State Airports Authority of Turkey. Eight time series

of the number of passengers in international and

domestic incoming & outgoing flights of these

airports are generated. We build models both

independently and neighbour-dependently and

compare their performance. For independent models,

which do not consider the neighbour effects, we

investigate the conventional ARIMA and the

recently proposed TBATS approach on forecasting

airline passenger volumes. We also study a

neighbour dependent approach where we use the

traditional regression with ARIMA errors approach

and incorporate the neighbour effects as a regressor

time series.

We now summarize the methods we applied in

our independent modelling and present our

neighbour dependent modelling approach.

Incorporating Explanatory Effects of Neighbour Airports in Forecasting Models for Airline Passenger Volumes

179

3.1 Independent Modelling with

ARIMA and TBATS

We apply ARIMA and TBATS methods for

independent modelling and analysis. ARIMA (Auto-

Regressive Integrated Moving Average) is a basic

approach for analysis and forecasting of equally

spaced univariate time series data. Box and Jenkins

proposed an entire family of ARIMA models and an

analysis to find the smallest number of estimated

parameters needed to fit the patterns in the data.

Box-Jenkins methodology involves three steps;

identification, estimation and diagnostic checking

(Pankratz, 1983). As a baseline for comparison, we

use automated ARIMA model fitting function in

forecast package of R programming.

Our preliminary data analysis reveals that our

passenger volume time series involve multiple

seasonality. Most commonly used methods for

modelling seasonal time series, such as Holt-

Winters, exponential smoothing approach, ARIMA

models, suffer dealing with double seasonality.

Recently, exponentially weighted methods for

multiple seasonal time series are proposed (De

Livera, 2010). To deal with double seasonality, an

extension of Holt-Winter is proposed (Taylor, 2003).

In another study, a multiple seasonal method is

developed that allows the seasonal cycle to be

updated more than once during the period of the

cycle (Gould et al, 2007). Also time series may have

complex seasonal patterns such as patterns with a

non-integer period, have high frequency multiple

seasonal patterns or may have dual calendar effects.

De Livera et al. propose a new innovations state

space model based approach that is capable of

dealing over parameterization and tackling with both

non-integer period and dual-calendar effects (De

Livera et al., 2011). They improve the traditional

single seasonal exponential smoothing methods, and

introduce two algorithms. They propose TBATS,

which stands for Trigonometric, Box-Cox transform,

ARMA errors, Trend, and Seasonal components

Model.

The Box-Cox transformation, ARMA errors,

Trend and Seasonal components (BATS) are defined

by;

()

=

(

−1)/ℎ ≠ 0

ℎ = 0

,

()

=

+

+

∑

+

(3.1)

=

+

+

(3.2)

=

(

1−

)

+

+

(3.3)

()

=

()

+

Υ

(3.4)

where ∈ is the Box-Cox transformation

parameter,

……

denote the constant periods

of the n seasonal components, is the long run

trend,

is an (,) process with Gaussian

white noise innovations having zero mean and

constant variance, and for =1,…., ,

is the

local stochastic level,

is the short term trend and

()

is the stochastic level of the −ℎ seasonal

component.

De Livera et al. proposed a new trigonometric

representation of seasonal components based on

Fourier series.

()

=

∑

,

()

(3.5)

,

()

=

,

()

()

+

,

∗()

()

+

Υ

()

(3.6)

,

∗()

=−

,

(

)

+

,

∗

(

)

(

)

+

Υ

()

(3.7)

Where Υ

()

and Υ

()

are smoothing parameters,

()

=2/

.

,

()

describe the stochastic level of

the i

th

seasonal component, and the stochastic

growth in the level of the i

th

component that is

needed to describe the change in the seasonal

component over time is described by

,

∗()

. The

number of harmonics required for the i

th

seasonal

component is denoted by

(De Livera et al.,

2011). The performance of these approaches on

forecasting passenger volumes is presented in the

experimental section.

3.2 Neighbour Dependent Modelling

In neighbour-dependent analysis, we incorporate the

explanatory effects of regressor time series of

neighbour airports using regression with ARIMA

errors method.

=

+

,

+⋯+

,

+

(3.8)

One of the key assumptions of multiple

regression is that

is an uncorrelated series. For

regression with ARIMA, it is considered that

contains autocorrelations. The resulting model is

now a regression model with ARIMA errors.

Equation 3.8 still holds but

is modeled as an

ARIMA process. We follow the notations of

Makridakis et al. (1998) for stating regression with

ARIMA model. For example, if

is an ARIMA (1,

1, 1) model, 3.8 can be written

=

+

,

+⋯+

,

+

(3.9)

ICORES 2016 - 5th International Conference on Operations Research and Enterprise Systems

180

where

(

1−

)(

1−

)

=

(

1−

)

and

is a white-noise series.

For identification of regressor time series, we

analyse the cross correlation between time series and

consider highly correlated and weakly correlated

series for building significant explanatory relations.

Initially we build regression with ARIMA models in

line with the correlation between Ataturk and

Esenboga series and we add regressor variables into

the model individually. Then we check the internal

cross correlation in Ataturk and Esenboga series in

order to identify regressor pairs. In this analysis it is

revealed that all series in Ataturk dataset are highly

correlated, hence we do not use any pairs while

building explanatory models for Esenboga series

with Ataturk data. In Esenboga dataset, we find out

that Esenboga International Incoming Passengers

data is weakly correlated with the rest of the series

and this result enables us to use pairs as regressors

while building explanatory models for Ataturk data.

4 EMPIRICAL ANALYSIS

To evaluate the performance of the presented

methods for forecasting the number of passengers

for an airport, we collected a list of real time series

as presented in Table 1. We adjusted our dataset by

forming equal time intervals for all time series.

Volumes of passengers are quarterly aggregated by

six-hour time periods for each day that helps to

detect seasonal patterns inherent in data. Following

the data adjustment, we divided the available one-

year data into training and test sets. We built models

with a training set involving 1092 data points and

tested the models with 124 data points.

Table 1: List of analysed time series.

Ataturk A. Domestic Flights Incoming Number

of Passengers (ADI)

Ataturk A. Domestic Flights Outgoing Number

of Passengers (ADO)

Ataturk A. International Flights Incoming

Number of Passengers (AII)

Ataturk A. International Flights Outgoing

Number of Passengers (AIO)

Esenboga A. Domestic Flights Incoming

Number of Passengers (EDI)

Esenboga A. Domestic Flights Outgoing

Number of Passengers (EDO)

Esenboga A. International Flights Incoming

Number of Passengers (EII)

Esenboga A. International Flights Outgoing

Number of Passengers (EIO)

All series have multiple seasonal patterns and

high frequency seasonality. For each set of data, four

independent models are fit: (I) ARIMA model with

frequency=4,(II) ARIMA model with frequency=28,

(III) TBATS Model with frequency=28 and (IV)

TBATS Model with double seasonality. For

neighbour-dependent analysis, we build regression

with ARIMA models with convenient regressor time

series. MAPE (Mean Absolute Percentage Error) is

the preferred forecasting accuracy measure for

simplicity when all data are positive and much

greater than zero (Hyndman and Koehler, 2006). We

also report MAE (Mean Absolute Percentage Error)

and MASE (Mean Absolute Scaled Error) based

results of our experiments (Hyndman and Koehler,

2006).

4.1 Independent Analysis with ARIMA

and TBATS Models

We first analyse the time-series, ACF

(Autocorrelation Function) and PACF (Partial

Autocorrelation Function) plots for all data sets. For

brevity, we present the models for one representative

data set (i.e., ADI) in detail and for the rest of the

series we present the model results.

The ACF and PACF plots illustrate two seasonal

periods (Figure 1). The first seasonality arises from

aggregating daily data quarterly by 6 hour time

periods and the seasonal period is four. The second

seasonality is observed with the help of PACF plot,

28 periods indicates the weekly seasonality in the

training data set.

Figure 1: ACF and PACF plots of ADI data.

Based on these results, we fit four independent

models. (I) ARIMA model with frequency=4, (II)

Incorporating Explanatory Effects of Neighbour Airports in Forecasting Models for Airline Passenger Volumes

181

ARIMA model with frequency=28, (III) TBATS

Model with frequency=28 and (IV) TBATS Model

with double seasonality for each dataset. We

developed these four independent models for each of

the eight time series and according to MAPE (Mean

Absolute Percentage Error) measure we present the

best independent models for each of the time series

in Table 3.

Table 2: Forecasting Accuracy of Independent Models for

ADI.

Model MAE MAPE MASE

ARIMA(0,1,2)(2,0,0)[4] 727.15 19.81 1.24

ARIMA(4,0,0)(0,1,1)[28] 513.41 12.45 1.05

TBATS(0.71, {2,1},

0.809, {<28,7>})

455.22 11.69 0.93

TBATS(0.684, {2,1},

0.861, {<4,1>, <28,6>})

441.30 11.07 0.90

Table 3: Best Independent Models for All Time Series.

Data The Best Model MAPE

ADO TBATS(1, {2,1}, -, {<28,8>}) 12.03

AII ARIMA(4,0,0)(0,1,1)[28] 9.24

AIO

TBATS(0.998, {5,4}, -, {<4,1>,

<28,5>})

9.70

EDI

TBATS(0.999, {4,5}, -,

{<28,8>})

11.94

EDO ARIMA(4,0,0)(0,1,1)[28] 7.85

EII

TBATS(0.327, {4,4}, -,

{<28,7>})

52.02

EIO

TBATS(0.095, {0,0}, -, {<4,1>,

<28,5>})

56.58

In half of the independent models, TBATS

model with multiple seasonality outperformed other

models according to MAPE measure. In the second

half of the independent models, TBATS and

ARIMA models with weekly seasonality outperform

other models. These results show that TBATS

successfully handles multiple seasonality in our

airline passenger volume time series, and it yields

better forecasting accuracies than the traditional

ARIMA approach.

4.2 Neighbour-dependent Analysis with

Regression with ARIMA Errors

We generated a cross correlation matrix for time

series in the preliminary analysis phase. Table 4

depicts the cross correlation in all series. According

to correlation values in this matrix, we establish the

explanatory relations between time series and we

build regression with ARIMA models with regressor

time series. While building regression with ARIMA

models for Ataturk Airport time series data, initially

we analysed cross correlation with Esenboga Airport

data. For example, ADI data can be paired with data

of EIO, EDI and EDO data as explanatory Regressor

variables.

Table 4: Cross Correlation between All Series.

ADI ADO AII AIO EDI EDO EII EIO

ADI 1.00 0.53 0.44 0.88 0.72 0.87 0.16 0.57

ADO 0.53 1.00 0.83 0.50 0.53 0.55 0.51 0.40

AII 0.44 0.83 1.00 0.46 0.42 0.39 0.62 0.31

AIO 0.88 0.50 0.46 1.00 0.64 0.79 0.21 0.59

EDI 0.72 0.53 0.42 0.64 1.00 0.74 0.26 0.34

EDO 0.87 0.55 0.39 0.79 0.74 1.00 0.20 0.53

EII 0.16 0.51 0.62 0.21 0.26 0.20 1.00 0.27

EIO 0.57 0.40 0.31 0.59 0.34 0.53 0.27 1.00

Figure 2: Explanatory Relations between Time Series.

Taking into account all possible routes and

connections we consider the explanatory relations in

Figure 2. In this figure, black undirected edges

represent the mutual interactions, the gray directed

edges represent the single sided influence.

According to these relations, it is clear that the

volume of passengers in international incoming

flights cannot be explained by neighbour effects.

This observation also makes sense in practice.

Hence we do not build regression with ARIMA

models for international incoming passengers data.

But we consider these series as explanatory variables

in other models.

When we consider all possible routes and

connections that may affect ADI data, we may

expect ADI data to be correlated with all series in

Esenboga data. We build seven neighbour-

dependent models for ADI dataset and compare

contribution of regression variables in forecasting

performance. Regression with ARIMA models for

ADI data is presented in Table 5. We find out that

ICORES 2016 - 5th International Conference on Operations Research and Enterprise Systems

182

the best model in neighbour-dependent models is

regression with ARIMA model including EDO data

as regressor variable, and it performs better than an

independent ARIMA model (seasonal period: 4). We

compare these neighbour dependent model results

with the independent model results in Table 6 by

MAPE accuracy measure. For this data set, the best

model out of all models is the independent TBATS

model with multiple seasonality.

Table 5: Regression with ARIMA models for ADI Data

(Neighbour dependent models).

Model MAE MAPE MASE

ARIMA(0,1,2)(2,0,0)[4]

REG. with EDI

721.32 18.80 1.23

ARIMA(0,1,3)(2,0,0)[4]

REG. with EDO

632.39 16.36 1.08

ARIMA(1,1,1)(2,0,0)[4]

REG. with EII

1063.30 32.13 1.82

ARIMA(3,1,1)(0,0,1)[4]

REG. with EIO

778.20 21.99 1.33

ARIMA(0,1,3)(2,0,0)[4]

REG. with EII and EIO

1020.09 30.04 1.74

ARIMA(1,1,1)(2,0,0)[4]

REG. with EII and EDI

690.17 17.71 1.18

ARIMA(0,1,3)(2,0,0)[4]

REG. with EII and EDO

619.53 15.88 1.06

Table 6: Forecasting Accuracy of Independent Models for

ADI.

Model MAE MAPE MASE

ARIMA(0,1,2)(2,0,0)[4] 727.15 19.81 1.24

ARIMA(4,0,0)(0,1,1)[28] 513.41 12.45 1.05

TBATS(0.71, {2,1},

0.809, {<28,7>})

455.22 11.69 0.93

TBATS(0.684, {2,1},

0.861, {<4,1>, <28,6>})

441.30 11.07 0.90

Ataturk Airport Domestic and International

Outgoing Passengers data are correlated with all

series in Esenboga data. For this reason, we add all

series as regressor variables individually and we also

consider the weakly interrelated Esenboga time

series pairs as regressor variables. Regression with

ARIMA models with the best test results are

presented in Table 7.

The best model for ADO data in the neighbour-

dependent approach is when EDI is used as a

regressor variable. It outperforms the worst model,

ARIMA (seasonal period: 4) in independent models.

For ADO data, the best model out of all models is

the independent TBATS model with single

seasonality. The comparison of the best performing

dependent and independent models is presented in

Table 8.

For AIO data, the best neighbour-dependent

model is the regression with ARIMA model

including EDO data. The best model out of all

models is the independent TBATS model with

multiple seasonality. The comparison of the best

performing dependent and independent models is

presented in Table 9.

Table 7: The Best Regression with ARIMA models for

ADO and AIO Data.

Data The Best Neighbour

Dependent Models

MAPE

ADO

ARIMA(3,1,0)(0,0,1)[4]

REG. with EDI

16.59

AIO

ARIMA(0,1,1)(2,0,1)[4

REG. with EDI

11.48

Table 8: Comparison of Models for ADO Data.

Data Models MAPE

ADO

ARIMA(3,1,0)(0,0,1)[4]

REG. with EDI

16.59

TBATS(1, {2,1}, -, {<28,8>}) 12.03

In regression with ARIMA models for Esenboga

Airport, due to the strong cross correlation in

Ataturk Airport time series data, each regression

with ARIMA model built for Esenboga Airport

routes involves one of the Ataturk Airport time

series. The accuracy performance of regression with

ARIMA models is demonstrated in Tables 10-11.

Table 9: Comparison of Neighbour Dependent and

Independent Models for AIO Data.

Data Models MAPE

AIO

ARIMA(0,1,1)(2,0,1)[4] REG.

with EDI

11.48

TBATS(0.998, {5,4}, -,

{

<4,1>, <28,5>

}

)

9.70

Table 10: Regression with ARIMA models for EDI Data

(Neighbour Dependent Models).

Model MAE MAPE MASE

ARIMA(3,1,1)(0,0,1)[4]

REG.with ADI

257.95 12.12 0.70

ARIMA(3,1,0)(2,0,0)[4]

REG.with ADO

372.63 15.94 1.02

ARIMA(0,1,3)(1,0,1)[4]

with drift REG. with AII

254.69 12.93 0.70

ARIMA(0,1,3)(1,0,1)[4]

with drift REG. with

AIO

277.25 14.20 0.76

Incorporating Explanatory Effects of Neighbour Airports in Forecasting Models for Airline Passenger Volumes

183

For EDI data, the best model out of all models is

the independent TBATS model with single

seasonality. The second best model is the neighbour-

dependent model, i.e., regression with ARIMA

model involving ADI data as regressor variable. The

comparison is presented in Table 11.

Regression with ARIMA models with best

forecasting accuracy for Esenboga Outgoing

(Domestic and International) Passengers data are

presented in Table 12.

Table 11: Comparison of Neighbour Dependent and

Independent Model for EDI Data.

Data Models MAPE

EDI

TBATS(0.999,{4,5},-,

{

<28,8>

}

)

11.94

ARIMA(3,1,1)(0,0,1)[4]

REG. with ADI

12.12

Table 12: The Best Regression with ARIMA models for

EDO and EIO Data.

Data The Best Neighbour Dependent

Models

MAPE

EDO

ARIMA(3,1,0)(2,0,0)[4] REG.

with ADI

20.99

EIO

ARIMA(3,1,0)(0,0,1)[4] REG.

with AII

64.33

For EDO data, the best model out of all models

is the independent ARIMA model with seasonal

period 28. For EIO data, the best model out of all

models is the independent TBATS model with

multiple seasonalities. The comparisons are

presented in Table 13 and Table 14.

Table 13: Comparison of Neighbour Dependent and

Independent Model for EDO Data.

Data Models MAPE

EDO

ARIMA(3,1,0)(2,0,0)[4]

REG. with ADI

20.99

ARIMA(4,0,0)(0,1,1)[28] 7.85

Table 14: Comparison of Neighbour Dependent and

Independent Model for EIO Data.

Data Models MAPE

EIO

ARIMA(3,1,0)(0,0,1)[4]

REG. with AII

64.33

TBATS(0.095, {0,0}, -, {<4,1>,

<28,5>

}

)

56.58

Figure 3: Resulting Explanatory Relations.

The experimental results illustrate some

improvements using the explanatory regression with

ARIMA models. Figure 3 summarizes the observed

explanatory relationships between the data sets that

showed considerable improvements in accuracies of

the forecasting models. The head of the arrow shows

the dependent time series while the tail shows the

regressor time series.

5 CONCLUSIONS

In this paper, we investigate incorporation of the

data of the neighbour airports in forecasting the

traffic volume of an airport. To analyse the

contribution of neighbour effects, we report on

forecasting accuracies of independent and

neighbour-dependent models for a variety of real

time series data sets. In several cases, we observe

improvement on forecasting performance when

neighbour-dependent models are used. We also

compared the performance of independent models

based on TBATS vs. ARIMA. In half of the series,

TBATS model with multiple seasonality

outperforms the other models. For the rest, TBATS

with single seasonality and ARIMA models provide

comparable results. For future work, we are planning

to observe the performance of TBATS model on

long term airline passenger data involving dual

calendar based seasonality. Dual calendar effects

were studied for demand cash (Du Toit, 2011) and

European tourist arrivals (Hassani et al., 2015) in the

literature.

ACKNOWLEDGEMENTS

Data used in this project was gathered in a

collaborative effort with Ankara Development

ICORES 2016 - 5th International Conference on Operations Research and Enterprise Systems

184

Agency. This study was supported in part by the

TUBITAK Career Award Project (112M950) of

Nilgun Ferhatosmanoglu.

REFERENCES

Alekseev, K.P.G., Seixas, J.M., 2009. A multivariate

neural forecasting modeling for air transport–

Preprocessed by decomposition. Journal of Air

Transport Management, 15, 212–216.

Benitez, R.B.C., Parades, R.B.C., Lodewijks, G., Nabais,

J.L., 2013. Damp trend Grey Model forecasting

method for airline industry. Expert Systems with

Applications, 40, 4915–4921.

De Livera, A.M., 2010a. Exponentially weighted methods

for multiple seasonal time series, International

Journal of Forecasting, 26, 655–657.

De Livera, A. M., 2010b. Automatic forecasting with a

modiﬁed exponential smoothing state space

framework. Department of Econometrics & Business

Statistics, Monash University.

De Livera, A.M., Hyndman, R.J., Snyder, R.D., 2011.

Forecasting Time Series with Complex Seasonal

Patterns Using Exponential Smoothing. Journal of the

American Statistical Association, 106:496, 1513-1527.

Du Toit, D.J., 2011. ATM Cash Management for a South

African Retail Bank, Master of Science Thesis,

Stellenbosch University.

European Organisation for the Safety of Air Navigation

(EUROCONTROL), 2013. The Challenges of Growth

2013, Task: European Air Traffic in 2035.

EUROCONTROL Statistics and Forecast Service, 14-

16.

Faraway, J., Chatfield, C., 1998. Time Series Forecasting

with Neural Networks: A Comparative Study Using

the Airline Data. Applied Statistics, 47, Part 2, pp.

231-250.

Fildes, R., Wei, Y., Ismail, S., 2011. Evaluating the

forecasting performance of econometric models of air

passenger traffic flows using multiple error measures,

International Journal of Forecasting, 27, 902–922.

Gould, P.G., Koehler, A.B., Ord, J.K., Synder, R.D.,

Hyndman, R.J., Vahid-Araghi, F., 2008. Forecasting

Time-Series with Multiple Seasonal Patterns.

European Journal of Operational Research, 207–222.

Grosche, T., Rothlauf, F., Heinzl, A., 2007. Gravity

models for airline passenger volume estimation.

Journal of Air Transport Management, 13, 175–183.

Hassani, H., Silva, E.M., Antonakakis, N., Filis, G.,

Gupta, R., 2015. Forecasting Accuracy Evaluation of

Tourist Arrivals: Evidence from Parametric and Non-

Parametric Techniques, Working Paper, University of

Pretoria.

Hyndman, R.J. , Koehler, A.B., 2006. Another look at

measures of forecast accuracy. International Journal

of Forecasting, 22, 679-688.

Makridakis, S., Wheelwright, S.C., Hyndman, R.J., 1998.

Forecasting Methods and Applications, Third Edition,

John Wiley&Sons.

Nam, Kyungdoo; Yi, Junsub; Prybutok, Victor R., 1997.

Predicting Airline Passenger Volume. The Journal of

Business Forecasting Methods & Systems, Vol. 16,

No. 1.

Pankratz, A., 1983. Forecasting with Univariate Box-

Jenkins Models, Concepts and Cases, John Wiley and

Sons.

Profillidis, V.A., 2000. Econometric and fuzzy models for

the forecast of demand in the airport of Rhodes,

Journal of Air Transport Management, 6, 95-100.

Samagaio, A., Wolters, M., 2010. Comparative analysis of

government forecasts for the Lisbon Airport, Journal

of Air Transport Management, 16, 213–217.

Scarpel, R.A., 2013. Forecasting air passengers at São

Paulo International Airport using a mixture of local

experts model, Journal of Air Transport Management,

26, 35-39.

Schultz, R.L., 1972. Studies of Airline Passenger Demand:

A Review, Transportation Journal, Vol. 11, No. 4, pp.

48-62.

Taylor, J.W., 2003. Short-term electricity demand

forecasting using double seasonal exponential

smoothing. Journal of the Operational Reseach

Society, 54, 799-805.

Tsui, W.H.K., Balli, H.O., Gilbey, A., Gow H., 2014.

Forecasting of Hong Kong airport's passenger

throughput, Tourism Management, Vol. 42, pp. 62–76.

TOBB, Turkish Civil Aviation Assembly Sector Report,

May 2013.

185