Three Essays On The Us Business Cycle Expectations Formation And Model Comparison PDF Download

This thesis contributes to the vast literature on understanding the disturbances that cause recessions, testing the importance of the assumption of rational expectations in macroeconomic models, and assessing model selection criteria. The main objective is to assess structural instabilities in macroeconomic models and to develop a new econometric methodology to compare different assumptions regarding expectations formation. Chapter 2 examines the role of oil price, demand, supply and monetary policy shocks during the 2001 US slowdown and Great Recession. It replicates the structural vector autoregression (VAR) of Peersman (2005) and extends it with time-varying parameters and stochastic volatility. Significant time variation is found in some impulse responses, with evidence that the constant coefficients VAR is erroneously representing structural instabilities as shocks. All models find that a combination of shocks caused the 2001 slowdown and Great Recession, but the role of individual shocks differs across models. Chapter 3 assesses the assumption of rational expectations versus adaptive learning in a dynamic stochastic general equilibrium (DSGE) model for the US economy. Using the framework in Smets and Wouters (2007) and Slobodyan and Wouters (2012), it finds that expectations implied by the rational expectations model are comparable to the adaptive learning models for actual and survey data on consumption and inflation. This chapter also formally assesses the overall fit of the model with different assumptions regarding expectations formation using the deviance information criterion (DIC), which is not commonly used to compare DSGE models. It finds that the rational expectations model is comparable to the adaptive learning models according to this criterion. Chapter 4 proposes fast algorithms for computing the DIC based on the integrated likelihood for a variety of high-dimensional latent variable models. The DIC has been a widely used Bayesian model comparison criterion since Spiegelhalter et al. (2002) introduced the concept and Celeux et al. (2006) introduced a number of alternative definitions for latent variable models. However, recent studies have cautioned against the use of some of these variants. While the DIC computed using the integrated likelihood seems to perform well, it is rarely used in practice due to computational burden. This chapter shows that the DICs based on the integrated likelihoods have much smaller numerical standard errors compared to the other DICs.