State Space Modelling of Cointegrated Systems using Subspace Algorithms

The use of subspace algorithms for the identification of non-stationary cointegrated stochastic systems is a promising technique that is currently under discussion. A revision of the literature provides two distinct algorithms: State Space Aoki Time Series (SSATS) identification algorithm (Aoki and Havenner 1991) and the Adapted Canonical Correlations Analysis (ACCA) of Bauer and Wagner (2002). Aoki's method is intuitively appealing, but lacks statistical foundation. In contrast, ACCA has a sound statistical basis, though intuition is somewhat lost. Both algorithms are revisited and commented. The study of the underlying ideas and properties of both previous algorithms leads us to propose a new method for subspace identification of non-stationary cointegrated stochastic systems, trying to combine the best features of each one. This new method provides a state space trend-cycle representation of a cointegrated system. Some preliminary simulation results are summarised, comparing these subspace methods with Johansen's maximum likelihood approach.