A novel suggestion on how to adequately treat stochastic non-stationary seasonality in tourism export forecasting

Abstract

Using a quarterly data sample of 28 EU countries ranging from 2004Q1 to 2016Q4, real tourism exports are modeled and estimated as a function of relative tourism export prices and aggregate real GDP. For almost all countries, HEGY tests applied to first-differenced variables cannot reject the null hypothesis of the presence of seasonal unit roots. Seasonal differencing would be necessary to remove the seasonal patterns from the data, yet employing seasonally differenced data results in information loss, distorted elasticity estimates, and mediocre forecast accuracy. Our suggestion is therefore to perform estimation and forecasting for each quarter separately. This approach avoids the use of seasonal differencing, results in plausible, seasonally varying estimates of income elasticities, and yields statistically significantly superior forecast accuracy, also in comparison to two data-driven benchmarks for seasonal data (i.e., SARIMA and ETS).