Abstract

The window length, L, is the first parameter that must be specified in Singular Spectrum Analysis (SSA) for time series analysis. A large window length has a potential to produce a good model fit, but it is unlikely to produce a parsimonious forecasting model. In this paper, we propose a new parsimonious vector forecasting model which uses an optimal m () coefficients for forecasting, instead of the L - 1 coefficients used in the standard vector forecasting method. This model enables SSA users to consider two different values for the window length: one for reconstruction and another for forecasting. The proposed and standard methods are compared methodologically and also implemented and tested on daily observations of six stocks: AAPL, AMZN, EBAY, IBM, INTC, MSFT, between Jan 1 2000 and Dec 31 2015, each including 4025 observations. It was found that the method proposed in this paper provides major improvements regarding forecast accuracy.