The authors discuss the cointegration in non-stationary time series and their statistical analysis using vector autoregression. Restricting on a reduced rank situation the analysis can be performed using maximum likelihood methods that require the solution of a generalized eigenproblem. The paper shows that the standard methods used so far to solve this eigenproblem are numerically undesirable. Numerically more stable algorithms are proposed using two alternative QR decompositions. Moreover the singular case is treated. Finally some remarks about operation counts and an illustrative example are given.
Reviewer: Hans-Peter Altenburg (Heidelberg)