Modelling and Forecasting Australian Domestic Tourism

Transcription

1 Modelling and Forecasting Australian Domestic Tourism George Athanasopoulos & Rob Hyndman

2 Outline Background 1 Background 2 Data 3 Regression models 4 Exponential smoothing via innovations state space models 5 Innovations state space models with exogenous variables 6 Forecasts 7 Conclusions and future research

3 Background Australian Tourism Industry: 1 International Arrivals 2 Outbound 3 Domestic Tourism

4 Background Australian Tourism Industry: 1 International Arrivals 2 Outbound 3 Domestic Tourism $55 billion - more than 3 times international arrivals (TFC 2005)

5 Background Australian Tourism Industry: 1 International Arrivals 2 Outbound 3 Domestic Tourism $55 billion - more than 3 times international arrivals (TFC 2005) Infrastructure maintenance

6 Background Australian Tourism Industry: 1 International Arrivals 2 Outbound 3 Domestic Tourism $55 billion - more than 3 times international arrivals (TFC 2005) Infrastructure maintenance

7 Background Australian Tourism Industry: 1 International Arrivals 2 Outbound 3 Domestic Tourism $55 billion - more than 3 times international arrivals (TFC 2005) Infrastructure maintenance My research - Research Fellow

8 Background Australian Tourism Industry: 1 International Arrivals 2 Outbound 3 Domestic Tourism $55 billion - more than 3 times international arrivals (TFC 2005) Infrastructure maintenance My research - Research Fellow Tourism Australia STCRC Monash University

9 Background Outline of presentation: 1 Data 2 Regression framework 3 Exponential smoothing 4 Exp smoothing + Exogenous variables 5 Forecasts 6 Conclusions and Further research

10 Outline Data 1 Background 2 Data 3 Regression models 4 Exponential smoothing via innovations state space models 5 Innovations state space models with exogenous variables 6 Forecasts 7 Conclusions and future research

11 Data National Visitor Survey - Visitor Nights (1998Q1-2005:Q2)

12 Data National Visitor Survey - Visitor Nights (1998Q1-2005:Q2) Holiday VFR Business Other

13 Data Aggregate Data & TFC Forecasts: VN Sample TFC forecasts

14 Outline Regression models 1 Background 2 Data 3 Regression models 4 Exponential smoothing via innovations state space models 5 Innovations state space models with exogenous variables 6 Forecasts 7 Conclusions and future research

15 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t )

16 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i)

17 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend

18 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita

19 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index

20 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index GDP t - Growth rate of real GDP per capita

21 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index GDP t - Growth rate of real GDP per capita BALI t - 1 for 2002:Q4 and beyond, 0 otherwise

22 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index GDP t - Growth rate of real GDP per capita BALI t - 1 for 2002:Q4 and beyond, 0 otherwise OLYMP t - 1 for 2000:Q4, 0 otherwise

23 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index GDP t - Growth rate of real GDP per capita BALI t - 1 for 2002:Q4 and beyond, 0 otherwise OLYMP t - 1 for 2000:Q4, 0 otherwise MAR t, JUN t, SEP t - Seasonal dummies

24 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index GDP t - Growth rate of real GDP per capita BALI t - 1 for 2002:Q4 and beyond, 0 otherwise OLYMP t - 1 for 2000:Q4, 0 otherwise MAR t, JUN t, SEP t - Seasonal dummies ε t - random error term

25 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index GDP t - Growth rate of real GDP per capita BALI t - 1 for 2002:Q4 and beyond, 0 otherwise OLYMP t - 1 for 2000:Q4, 0 otherwise MAR t, JUN t, SEP t - Seasonal dummies ε t - random error term

26 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index GDP t - Growth rate of real GDP per capita BALI t - 1 for 2002:Q4 and beyond, 0 otherwise OLYMP t - 1 for 2000:Q4, 0 otherwise MAR t, JUN t, SEP t - Seasonal dummies ε t - random error term Step 1: Run OLS and test for upto 1 lag of each variable.

27 Regression models Tourism demand function: VN i t = f (t, DEBT t, DPI t, GDP t, BALI t, OLYMP t, MAR t, JUN t, SEP t, ε t ) VN i t - ln(visitor nights per capita travelling for purpose i) t - exponential trend DEBT t - Growth rate of real personal debt per capita DPI t - Growth rate of domestic price index GDP t - Growth rate of real GDP per capita BALI t - 1 for 2002:Q4 and beyond, 0 otherwise OLYMP t - 1 for 2000:Q4, 0 otherwise MAR t, JUN t, SEP t - Seasonal dummies ε t - random error term Step 1: Run OLS and test for upto 1 lag of each variable. Step 2: Sequentially drop insignificant parameters and estimate efficiently using SUR.

28 Regression models Estimated demand system: Regressor Holiday VFR Business Other Intercept (13.33) (21.03) (22.84) (47.28) t (0.50) (0.88) D t (1.23) (2.00) P t (1.64) 7.58 (2.89) Y t (8.14) BALI t (17.75) OLYMP t (51.26) MAR t (13.06) (26.87) (24.28) (64.74) JUN t (12.40) (26.87) (24.51) (64.74) SEP t (14.01) (27.84) (25.57) (66.86) R R Significant at the 5% level.

29 Outline Exponential smoothing via innovations state space models 1 Background 2 Data 3 Regression models 4 Exponential smoothing via innovations state space models 5 Innovations state space models with exogenous variables 6 Forecasts 7 Conclusions and future research

30 Exponential smoothing via innovations state space models Innovation state space models - ETS(A,-,A):

31 Exponential smoothing via innovations state space models Innovation state space models - ETS(A,-,A): No trend Additive trend Damped trend y t = l t 1 + s t m + ε t y t = l t 1 + b t 1 + s t m + ε t y t = l t 1 + b t 1 + s t m + ε t l t = l t 1 + αε t l t = l t 1 + b t 1 + αε t l t = l t 1 + b t 1 + αε t s t = s t m + γε t b t = b t 1 + βε t b t = φb t 1 + βε t s t = s t m + γε t s t = s t m + γε t ŷ n+h = l n + s n+h m ŷ n+h = l n + hb n + s n+h m ŷ n+h = l n + (1 + φ + + φ h 1 )b n +s n+h m where: 0 < α < 1, 0 < β < α, 0 < γ < 1, 0 < φ < 0.98.

32 Exponential smoothing via innovations state space models Innovation state space models - ETS(A,-,A): No trend Additive trend Damped trend y t = l t 1 + s t m + ε t y t = l t 1 + b t 1 + s t m + ε t y t = l t 1 + b t 1 + s t m + ε t l t = l t 1 + αε t l t = l t 1 + b t 1 + αε t l t = l t 1 + b t 1 + αε t s t = s t m + γε t b t = b t 1 + βε t b t = φb t 1 + βε t s t = s t m + γε t s t = s t m + γε t ŷ n+h = l n + s n+h m ŷ n+h = l n + hb n + s n+h m ŷ n+h = l n + (1 + φ + + φ h 1 )b n +s n+h m where: 0 < α < 1, 0 < β < α, 0 < γ < 1, 0 < φ < 0.98.

33 Outline Innovations state space models with exogenous variables 1 Background 2 Data 3 Regression models 4 Exponential smoothing via innovations state space models 5 Innovations state space models with exogenous variables 6 Forecasts 7 Conclusions and future research

34 Innovations state space models with exogenous variables Innovation state space model including exogenous variables - ETSX(A,A D,N,X):

35 Innovations state space models with exogenous variables Innovation state space model including exogenous variables - ETSX(A,A D,N,X): Damped trend y t = l t 1 + b t 1 + x tδ + ε t

36 Innovations state space models with exogenous variables Innovation state space model including exogenous variables - ETSX(A,A D,N,X): Damped trend y t = l t 1 + b t 1 + x tδ + ε t l t = l t 1 + b t 1 + αε t b t = φb t 1 + βε t ŷ n+h = l n + (1 + φ + + φ h 1 )b n + ˆx n+hˆδ where: 0 < α < 1, 0 < β < α, 0 < γ < 1, 0 < φ < X = (DEBT, DPI, GDP, BALI, OLYMP, MAR, JUN, SEP)

37 Innovations state space models with exogenous variables Estimates of the ETSX models: Parameter Holiday VFR Business Other α β φ Variable D t P t Y t BALI t OLYMP t MAR t JUN t SEP t

38 Outline Forecasts 1 Background 2 Data 3 Regression models 4 Exponential smoothing via innovations state space models 5 Innovations state space models with exogenous variables 6 Forecasts 7 Conclusions and future research

39 Forecasts Forecast comparisons:

40 Forecasts Forecast comparisons: Holdout sample: (2004:Q3-2005:Q2)

41 Forecasts Forecast comparisons: Holdout sample: (2004:Q3-2005:Q2) MAPE Regr ETS ETSX TFC Holiday VFR Business Other Total Average

42 Forecasts Forecast comparisons: Holdout sample: (2004:Q3-2005:Q2) MAPE Regr ETS ETSX TFC Holiday VFR Business Other Total Average

43 Forecasts Forecast comparisons: Holdout sample: (2004:Q3-2005:Q2) MAPE Regr ETS ETSX TFC Holiday VFR Business Other Total Average

44 Forecasts Forecast comparisons: Holdout sample: (2004:Q3-2005:Q2) MAPE Regr ETS ETSX TFC Holiday VFR Business Other Total Average

45 Forecasts Long term annual forecasts: Regr ETS ETSX TFC

46 Forecasts Long term annual forecasts for each component: PANEL A: Holiday PANEL B: VFR PANEL C: Business PANEL D: Other Regr ETS ETSX TFC

47 Outline Conclusions and future research 1 Background 2 Data 3 Regression models 4 Exponential smoothing via innovations state space models 5 Innovations state space models with exogenous variables 6 Forecasts 7 Conclusions and future research

48 Conclusions and future research Domestic tourism and existing forecasts:

49 Conclusions and future research Domestic tourism and existing forecasts: Statistical models outperform TFC forecasts

50 Conclusions and future research Domestic tourism and existing forecasts: Statistical models outperform TFC forecasts Existing long term forecasts over-optimistic

51 Conclusions and future research Domestic tourism and existing forecasts: Statistical models outperform TFC forecasts Existing long term forecasts over-optimistic Identified important economic relationships

52 Conclusions and future research Domestic tourism and existing forecasts: Statistical models outperform TFC forecasts Existing long term forecasts over-optimistic Identified important economic relationships Future research:

53 Conclusions and future research Domestic tourism and existing forecasts: Statistical models outperform TFC forecasts Existing long term forecasts over-optimistic Identified important economic relationships Future research: Further development of ETSX

54 Conclusions and future research Domestic tourism and existing forecasts: Statistical models outperform TFC forecasts Existing long term forecasts over-optimistic Identified important economic relationships Future research: Further development of ETSX Comprehensive Monte Carlo examining the proposed two step procedure Construction of prediction intervals via theory or simulation Application to other data e.g. international arrivals

55 Conclusions and future research Domestic tourism and existing forecasts: Statistical models outperform TFC forecasts Existing long term forecasts over-optimistic Identified important economic relationships Future research: Further development of ETSX Comprehensive Monte Carlo examining the proposed two step procedure Construction of prediction intervals via theory or simulation Application to other data e.g. international arrivals Hierarchical forecasting - Australia, States, Regional

56 Conclusions and future research Thank you!!!