A comparison of singular spectrum analysis forecasting methods to forecast South African tourism arrivals data

Abstract

Singular spectrum analysis (SSA) has been shown to be a powerful non-parametric time series method by which a time series is unfolded into a Hankel structured matrix. Time series structures are then extracted by direct application of singular value decomposition to the Hankel matrix and forecasts can then be produced. SSA is especially powerful in case time series exhibit seasonality combined with trends (linear, exponential or curvilinear). The method has been shown to outperform forecasts produced by SARIMA processes when observed time series contains seasonality and trend patterns. The application of SSA for forecasting tourism time series is very recent and this paper compares the recurrent-, vector- and joint-horizon-forecasting methods using Monte Carlo Simulations and a South African tourism arrivals datase