Forecasting time series with trend and seasonal patterns based on SSA

This paper presents a combination of deterministic function and neural network for modeling time series with trend and seasonal patterns. The deterministic function is implemented to model the components obtained by singular spectrum analysis (SSA). A properly SSA decomposition would yield a trend of the original series, oscillatory and irregular component series. These trend and oscillatory series are then fitted by deterministic function and the residuals will be added to the irregular component, namely, superposed residual. Neural network seems to be the most appropriate model to handle potential nonlinearities in the superposed residual. The proposed method combines the deterministic model, i.e. exponential or polynomial, and sinusoidal function with the neural network model. Based on the empirical applications, it can be shown that the proposed method produces smaller root mean square forecast error (RMSFE) compared with the existing methods proposed in the literature.