Learning in DSGE macroeconomics

• Ano de defesa: 2013
• Autor(a) principal: Velecico, Igor
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: DSGE; Expectativas racionais; Macroeconomia; Macroeconomics; Rational expectations
• Link de acesso: http://www.teses.usp.br/teses/disponiveis/12/12138/tde-20012014-154530/

In this thesis we analyze learning mechanisms applied to a variety of macroeconomic models. In the first chapter, we present and discuss the advantages and limitations of estimating Dynamic Stochastic General Equilibrium (DSGE) models added with learning, thus suppressing the central assumption of rational expectations. First, we introduce the reader on how learning can be inserted in those models, starting from the discussion of where and how the rational expectations operator is substituted by the learning mechanism. We then present several additional learning setups related to the information set available to agents considered by the literature, which affect directly the dynamics of the final model. Last, we estimate three different models to assess the advantages of learning in our artificially generated data and real data for Brazil. In the second chapter, we algebraically show the limitations of learning and propose two flexible methods to deal with the parameter instability in data. The first of these methods is closely related to the DSGE-VAR methodology, which we call Learning DSGE-VAR, and the second, which departs even further from the DSGE model, which we call Learning Minimum State Variable, or LMSV. Finally, in the third chapter we provide evidences that the supposedly moderate improvements found in the previous chapters have more to do with the nature of the model at hand than to the learning method itself. To do so, we simulate problems using a time-varying structure similar to the one presented in chapter 1 and evaluate the likelihood improvements with different learning mechanisms. We then provide empirical evidences of learning in reduced form models to forecast inflation, interest rates and output gap for the Brazilian economy, using ad-hoc reduced form models commonly used by practitioners.