MSVARlib: a new Gauss library to estimate multivariate Hidden Markov Models

This paper introduces a new open source Gauss library to estimate Multivariate Hidden Markov Models (HMM) in their simpler specification. These new programs are based upon the works of Hamilton (1994) and Krolzig (1998) and allow assessment of models with 2, 3 or 4 states through classical optimization of the maximum likelihood method. The modular architecture of the program is presented in a first part. It has been designed to allow new improvements (generalized non linear MS models or enhancement to a Bayesian framework). A second part, gives some illustration through a three state model based on the American Industrial production and a new stochastic coincident indicator of a recession for the US economy, following the papers of Ferrara (2003), Bellone and Saint-Martin (2003) and Bellone (2004).