Forecasting Accuracy Evaluation of Tourist Arrivals: Evidence from Parametric and Non-Parametric Techniques

This paper evaluates the use of several parametric and nonparametric forecasting techniques for predicting tourism demand in selected European countries. ARIMA, Exponential Smoothing (ETS), Neural Networks (NN), Trigonometric Box-Cox ARMA Trend Seasonal (TBATS), Fractionalized ARIMA (ARFIMA) and both Singular Spectrum Analysis algorithms, i.e. recurrent SSA (SSA-R) and vector SSA (SSA-V), are adopted to forecast tourist arrivals in Germany, Greece, Spain, Cyprus, Netherlands, Austria, Portugal, Sweden and United Kingdom. This paper not only marks the introductory application of the TBATS model for tourism demand forecasting, but also marks the first instance in which the SSA-R model is effectively utilized for forecasting tourist arrivals. The data is tested rigorously for normality, seasonal unit roots and break points whilst the out-of-sample forecasts are tested for statistical significance. Our findings show that no single model can provide the best forecasts for any of the countries considered here in the short-, medium- and long-run. Moreover, forecasts from NN and ARFIMA models provide the least accurate predictions for European tourist arrivals, yet interestingly ARFIMA forecasts are better than the powerful NN model. SSA-R, SSA-V, ARIMA and TBATS are found to be viable options for modelling European tourist arrivals based on the most number of times a given model outperforms the competing models in the above order. The results enable forecasters to choose the most suitable model (from those evaluated here) based on the country and horizon for forecasting tourism demand. Should a single model be of interest, then, across all selected countries and horizons the SSA-R model is found to be the most efficient based on lowest overall forecasting error.