Estimating large rational expectations models by FIML-some experiments using a new algorithm with bootstrap confidence limits

We set out an algorithm, based on a guided grid search procedure, for estimating large rational expectations models by FIML. Confidence limits are established by bootstrapping. We give results for a small four-equation version of the Liverpool Model of the UK and for the full Model; these suggest that there is also some bias to be adjusted for. The required bootstrap number for the small model is less than 200 for convergence of the estimates. Also should there be unit roots in the errors, these have the interpretation, given a fully specified theoretical structure, of omitted variables; in a Montecarlo study of the small model when its errors are unit roots it is found that the main parameter estimates are rather robust, with the error processes capturing these roots.