Forecasting tourist arrivals using origin country macroeconomics

ABSTRACT

This study utilizes both disaggregated data and macroeconomic indicators in order to examine the importance of the macroeconomic environment of origin countries for analysing destinations’ tourist arrivals. In particular, it is the first study to present strong empirical evidence that both of these features in tandem provide statistically significant information of tourist arrivals in Greece. The forecasting exercises presented in our analysis show that macroeconomic indicators conducive to better forecasts are mainly origin country-specific, thus highlighting the importance of considering the apparent sharp national contrasts among origin countries when investigating domestic tourist arrivals. Given the extent of the dependency of the Greek economy on tourism income and also the perishable nature of the tourist product itself, results have important implications for policymakers in Greece.

KEYWORDS:

• Tourist arrivals forecasting
• seasonal ARIMA
• Diebold–Mariano test
• disaggregated data
• macroeconomic indicators

JEL CLASSIFICATION:

• C22
• C53
• F19
• O10

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

¹ See, inter alia, Kulendran and Wilson (2000); Cho (2001); Lim and McAleer (2001a); Goh and Law (2002); Cho (2003); Kulendran and Witt (2003); Chen (2005); Vu and Turner, (2005); Kim and Moosa (2005); Vu and Turner (2006); Wong et al. (2007); Coshall (2009); Santos (2009); Brida and Risso (2011); Gounopoulos, Petmezas, and Santamaria (2012); Zheng et al. (2012); Wan, Wang, and Woo (2013).

² See for instance, Kulendran and Wilson (2000); Song and Witt (2000); Kulendran and Witt (2003); Song and Witt (2006); Wong, Song, and Chon (2006); Wong et al. (2007).

³ See indicatively, Song, Witt, and Jensen (2003); Wong et al. (2007).

⁴ See, for example, Lim and McAleer (2001b); Chen (2005); Vu and Turner, (2005); Yu, Schwartz, and Humphreys (2007); Zheng et al. (2012).

⁵ More specifically, the first estimation period of the models is $\hat{T}$ = 37 months, that is, from January 2003 until January 2006. The remaining $h$ = 89 months of our sample size are used for the evaluation period of the out-of-sample forecasts. In order to proceed to the first out-of-sample forecast (i.e. $t+1$ forecast or month 38), we estimate the models using the initial 37 months. For each subsequent out-of-sample forecast, we add to the estimation period an additional month. For example, for the $t+2$ forecast, we use $\hat{T}+1$ monthly observations. The total number of observations is $T=\hat{T}+h$. The out-of-sample phase has been selected in order to capture the period before, during and after the global financial crisis (as well as the Greek debt crisis).

⁶ We have also used the Hylleberg–Engle–Granger–Yoo (HEGY) unit root test (Hylleberg et al. 1990), and we find that the seasonal component of our series is stationary. Furthermore, we extract the seasonal component from our series using the TRAMO/SEATS procedure and we run unit root tests on the de-seasonalized series and the seasonal factor. The tests confirm that the series is nonstationary with a stationary seasonal component. Results are available upon request.

⁷ The results are available upon request.

⁸ The k and l denote the number of lags for the autoregressive and the moving average polynomials, respectively. The k* and l* denote the number of lags for the seasonal autoregressive and the seasonal moving average polynomials, respectively.

⁹ The certain parameter ranges of autoregressive (AR) and moving average (MA) orders for both seasonal and nonseasonal components rely on methods of orders’ selection (Schwarz’s, 1978; Akaike’s, 1974; information criteria) and correlogram diagnostics. We use the information criteria in order to set the upper bounds for the ranges of the dynamics included in the models.

¹⁰ Originally the model is denoted as SARIMA $\left(k,1,l\right)$$\left(k^{•},0,l^{•}\right)$, but for simplicity, we have removed the zero term. The proposed framework can be defined either as a SARMA specification for the dependent variable $\left(1-L\right)\left(log{y}_t\right)$ or as a SARIMA specification for the dependent variable $log{y}_t$. In the latter case, the models can be stated as SARIMA with integrated order I(1), or SARIMAX $\left(k,1,l\right)$$\left(k^{•},l^{•}\right)$.The log transformation stabilizes the variance of $y_t$, thus, it is preferred (see Lütkepohl and Xu 2012). The models’ integration order is set to 1. For higher order of integration, the forecasting accuracy deteriorates significantly.

¹¹ For the case the evaluation criterion $\mathrm{\Psi }$ is the RMSE, then ${\mathrm{\Psi }}_t^{\hspace{0.17em}}={\left(y_{t+1}-y_{t+1|t}\right)}^2$.

¹² Typical in forecasting studies is the comparison of candidate models with simple benchmark models, that is, with/without drift random walk model, first-order autoregressive model, etc. In our study, the forecasting ability of the naive benchmark models is statistically inferior. The results are available upon request.