Univariate versus multivariate time series forecasting: an application to international tourism demand
Tourist numbers from several origin countries to a particular destination country form a vector series. In the presence of a 'rich' cross-correlation structure, that is if after allowing for autocorrelation the sample cross-correlation function exhibits meaningful and statistically significant correlations, the accuracy when forecasting a particular origin-destination tourist flow is likely to be improved by utilising information from the other tourist flows. Multivariate time series models may be expected to generate more accurate forecasts than univariate models in this setting. However, in the absence of these conditions, univariate forecasting models may well outperform multivariate models. An empirical investigation of tourism demand from four European countries to the Seychelles shows an absence of such a 'rich' structure and that ARIMA exhibits better forecasting performance than univariate and multivariate state space modelling. One implication that an absence of a 'rich' cross-correlation structure holds for econometric modelling is that explanatory variables which are strongly correlated with the tourist flow series are likely to be uncorrelated across origin countries.