Tourism Forecasting in South Africa – some perspectives - Skool vir ...

runmidgeΤεχνίτη Νοημοσύνη και Ρομποτική

20 Οκτ 2013 (πριν από 3 χρόνια και 8 μήνες)

79 εμφανίσεις

Page
1

Tourism Forecasting in South
Africa


Some Perspectives

Andrea
Saayman

North
-
West University,
Potchefstroom Campus

Page
2

Agenda


Some facts about tourism to South Africa


Review of academic studies


Neural networks


Pure time series forecasts


ARDL forecasts


VEC and TVP forecasts


Seasonality in tourist arrivals


Current challenges

Page
3

Some facts about tourism to SA


The sanction years:


Domestic tourism focus


International tourism stagnation


Stagnant years:


Total tourist arrivals 1980


702 794


Total tourist arrivals 1990


1 029 094


Average growth in arrivals 4.3% per year


Change started in 1991/2:


Tourist arrivals in 1992


2 891 721


Peak arrivals in 2008


9 728 860

Page
4

Time series of tourist arrivals



0
2000000
4000000
6000000
8000000
10000000
12000000
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
Page
5

African versus intercontinental
tourists



0
0.2
0.4
0.6
0.8
1
1.2
Page
6

South Africa’s top 15





Africa
Europe
North America
Asia
Australasia
Central and South
America
Middle East
Indian Ocean
islands
0
1,000,000
2,000,000
Zimbabwe
Lesotho
Mozambique
Swaziland
Botswana
UK
USA
Germany
Namibia
Zambia
Nigeria
Malawi
Netherlands
France
Australia
Page
7

Tourism’s growing importance in
the economy

1994/5

2011

International tourists

4 684 064

8 339 351

Visitor exports

R10 billion

R75 billion

Contribution to

GDP

3%

8.6%

Contribution

to e
mployment

550,000

1,188,000

Page
8

Review of academic studies


Only a handful of academic papers on
forecasting tourist arrivals


Focusing only on intercontinental tourist
arrivals


2001


Burger et al. using neural networks


2010


Saayman

&
Saayman

comparing pure
time series forecast accuracy


2012


Louw

&
Saayman

using ARDL models


2012


Botha &
Saayman

comparing TVP and
VEC forecasts


Page
9

Neural networks


A practitioner’s guide


Case study of US tourist demand for city
of Durban (1992
-
1998)


Compared time
-
series methods with
neural networks


Back
-
propagation algorithm with momentum
used to train process

Page
10

Neural networks

MAPE

Naïve

11.24

Moving

average

10.89

Exponential

smoothing

10.04

ARIMA

11.30

Multiple regression

7.20

Neural network

5.07

Neural network (12 month)

11.00

Page
11

Univariate

forecasts


Compared accuracy of
univariate

time
-
series forecasts for tourist arrivals from top
5 intercontinental markets


Monthly arrivals from 1994 to 2006


Ex post forecasts for 2007


Page
12

Comparison based on MAPE

Model

Germany

France

UK

Netherlands

USA

12
-
month forecast

Naïve 1

3.239 (5)

2.142 (3)

3.859 (5)

3.828 (4)

1.579 (3)

Naïve 2

2.256 (3)

4.587 (5)

2.313 (4)

4.748 (5)

1.824 (5)

Holt Winters

0.708 (2)

1.059 (2)

0.582 (2)

1.158 (2)

0.809 (2)

ARIMA

2.864 (4)

2.149 (4)

1.951 (3)

1.653 (3)

1.612 (4)

SARIMA

0.610 (1)

0.954 (1)

0.395 (1)

0.828 (1)

0.613 (1)

6
-
month forecast

Naïve 1

3.534 (5)

2.208 (3)

4.195 (5)

6.000 (5)

1.879 (4)

Naïve 2

1.794 (3)

2.450 (5)

2.166 (3)

2.653 (4)

2.071 (5)

Holt Winters

0.812 (2)

1.093 (2)

0.515 (2)

1.036 (2)

1.221 (2)

ARIMA

3.119 (4)

2.239 (4)

2.203 (4)

1.735 (3)

1.804 (3)

SARIMA

0.787 (1)

0.991 (1)

0.452 (1)

0.724 (1)

0.754 (1)

Page
13

SARIMA forecasts



Page
14

Univariate

forecasts


More accurate forecasts of overseas
arrivals in SA with techniques that account
for seasonality


SARIMA forecasts outperform others,
including Holt
-
Winters


Non
-
seasonal ARIMA
-
models perform
poorly in this context


Policy application remains limited

Page
15

ARDL forecasts


Forecasted arrivals from Asia, Europe,
South America, North America,
Australasia and UK


Ex post forecasts


1 to 3 year horizon


Quarterly data from 1994 to 2004


ARDL model with ECM


Included income, travel cost, price,
infrastructure variables

Page
16

Results


forecasting

Forecast
length

RMSPE

MAPE

Asia

3 years ahead

12.043

9.661

2 years ahead

8.419

6.653

1 year ahead

4.823

3.917

United
Kingdom

3 years ahead

21.239

9.215

2 years ahead

18.571

7.516

1 year ahead

19.483

1.828

South America

3 years ahead

30.976

18.514

2 years ahead

28.192

26.028

1 year ahead

16.948

16.398

Page
17

Results


forecasting

Forecast
length

RMSPE

MAPE

Europe

3 years ahead

31.456

18.406

2 years ahead

32.957

18.626

1 year ahead

39.096

26.488

Australasia

3 years ahead

40.933

35.538

2 years ahead

32.014

26.034

1 year ahead

15.047

11.031

North
America

3 years ahead

36.652

32.017

2 years ahead

26.933

22.773

1 year ahead

14.801

11.772

Page
18

ARDL forecasts


Long run


real GDP per capita, real price
and infrastructure significant


Demand is income elastic over both short and
long run


Infrastructure only creates long run benefit


Demand is relative price inelastic over both short
and long run


Transport cost has relatively small effect


Forecast accuracy:


Accuracy good for 1
-
year horizon


UK and Asia models presented best results


Page
19


Forecast models for arrivals from
continents


Quarterly data from 1994 to 2009


Ex ante forecasts for 1 year (over FIFA
WC)


VECM form benchmark model


Compare TVP
-
LRM and TVP
-
ECM
specification


Used AR form of transition equation


ECM and TVP forecasts

Page
20

Forecasting
accuracy



North
America

UK

Europe

South
America

Asia

Australia

VECM













MAPE

0.072

0.076



0.286

0.175

0.122

MAD/MEAN

0.069

0.077

0.115

0.408

0.167

0.119

TVP













MAPE

0.095

0.101

0.101

0.314

0.102

0.108

MAD/MEAN

0.085

0.096

0.097

0.734

0.088

0.119

TVP
-
EC













MAPE

0.123

0.264

0.251

0.215

0.113

0.103

MAD/MEAN

0.115

0.268

0.252

0.338

0.112

0.103

Page
21

Demand
Elasticities:

North America



Page
22


Intercontinental tourist arrivals to South
Africa


Income elastic, but price inelastic destination


Comparing methods:


VECM superior in more stable environment


TVP
-
LRM superior when gradual adjustment or shock
long ago


TVP
-
ECM superior in short
-
term shock situations


Demand elasticities is becoming more consistent

ECM and TVP forecasts

Page
23

Seasonality in intercontinental
arrivals


SARIMA models outperform other non
-
seasonal models


2009 paper by
Shen
, Li & Song


Deterministic seasonal dummies


Stochastic treatment


of seasonality



0
20,000
40,000
60,000
80,000
100,000
120,000
140,000
160,000
180,000
1994Q1
1995Q1
1996Q1
1997Q1
1998Q1
1999Q1
2000Q1
2001Q1
2002Q1
2003Q1
2004Q1
2005Q1
2006Q1
2007Q1
2008Q1
2009Q1
Page
24

Current challenges


Forecasts for SA done by WTTC


Econex

forecasted on ad
-
hoc basis using
ARDL


A need for:


More continuous forecasts


More inclusive forecasts


Focusing on more than arrivals

“Travel and Tourism research and forecasting in
South Africa needs significant improvement, both
in terms of quantity and quality”

Page
25

Current challenges


There is a need for:


A dedicated tourism forecasting unit


Austrian WIFO


Australian forecasting committee


Wider scope to serve a variety of industry
needs


Skills development in forecasting


Better co
-
operation between all parties


Page
26

Thank you