Stability of singular spectrum analysis and causality in time series

The concept of causality has been widely studied in econometrics and statistics since 1969, when C. J. Granger published his paper "Investigating causal relations by econometric models and cross-spectral methods". The intuitive basis for his definition of causality is the following: time series Y is causing time series X if the use of the additional information provided by Y improves the forecast of series X. In the present thesis we focus on combining Granger's causality concept with the Singular Spectrum Analysis (SSA) technique. SSA is founded on the idea of transforming the time series into a multidimensional trajectory form (Hankel matrices), Singular Value Decomposition with subsequent projection to a lower-dimensional subspace and diagonal averaging. The main aim of the present thesis is to study the causality concept through SSA prism in details and suggest a novel causality measure, which can be used outside the stationary autoregressive class, which is the framework for Granger's original causality concept. We first apply standard statistical tests directly to simulated data to assess the improvement of forecast quality of bivariate multidimensional SSA (MSSA) of time series X and Y compared with SSA of time series X only. Although the results of performance of these tests are reasonably conclusive, the simulation method is time consuming and, thus, more theoretical understanding is desirable. We solve a fundamental scaling problem of the MSSA approach by introducing so-called linearized MSSA. The linearized MSSA approach shows a way towards a causality measure, calculated from the forecast linear recurrence formula (LRF) coefficients. We finally analyze SSA and (non-linear) bivariate MSSA approach in terms of first order stability analysis under perturbations leading to the construction of a valid suitable measure of causality. The construction of the measure requires some simplifying assumptions in the stability analysis whose validity we verify for both simulated and real data