Comovements in large systems

Abstract

In this paper we study various methods for detecting the co integrating rank as the number of variables gets large. We show that the use of standard tools will always lead to misleading inferences in such settings due to excessive size distortions. Particularly the LR test tends to produce too much cointegration. We introduce a new test statistic that displays excellent size properties in both small and large systems. In addition we propose a model selection procedure for selecting the co integrating rank. A new criterion outperforms the standard informationtheoretic criteria (AIC, BIC).