Prediction intervals for the vector singular spectrum analysis forecasting algorithm in a median-based singular spectrum analysis

Abstract

In recent years singular spectrum analysis (SSA) has been used as a powerful technique to analyze time series, including theoretical developments and application to many practical problems. However, no inclusive theoretical approach has been discussed regarding the construction of confidence intervals for forecasts. Due to the prominent role of prediction intervals in evaluating the accuracy of forecasts in time series analysis, in this paper, we consider the topic of constructing prediction intervals for SSA. Namely, we revise the existing approaches for the vector SSA forecasting method and propose a new median-based alternative to this algorithm where the mean in the diagonal averaging step of the SSA algorithm is replaced by the median. The results from the existing and proposed approaches are compared by considering Monte Carlo simulations and real data applications.