Abstract
This thesis studies empirical Behavioural New Keynesian (NK) DSGE models that incorporate
heterogeneous expectations and reinforcement learning (Brock and Hommes,
1997) to endogenise the distribution of agent types, with a specific focus on the use
of higher-order estimation methods to capture the highly non-linear features of the
learning mechanism. The thesis is composed of three chapters.
Chapter 1 investigates the properties of efficient local Gaussian-based filters for
Bayesian posterior inference on the parameters of non-linear DSGE models. A variety
of filters is assessed estimating artificial data generated from the simulation of the
fifth-order Taylor expansion of a Real Business Cycle (RBC) model: the second-order
Extended Kalman filter, a linear Kalman filter applied on the stochastic steady state
resulting from a higher-order approximation of the model, and sigma-point filters
- such as the Unscented Kalman filter and the Cubature Kalman filter. Results
show these filtering techniques represent a valid alternative to the particle filter for
problems requiring highly computational efforts.
Chapter 2 develops a Behavioural NK model enriched with portfolio adjustment
costs to study long-term asset purchases. Adjustment costs on the composition of the
households’ financial portfolio allow for bond-market segmentation by introducing a
wedge on the yields paid by bonds with different duration. Reinforcement learning
combined with bounded-rational agents introduces state-dependent asset-purchases
multipliers, by linking policy measures to the sentiment prevailing in the economy.
In this framework, policy experiments support the role of asset purchase programs as
counter-cyclical measures and emphasize the importance of Central Bank credibility
for monetary policy transmission.
Chapter 3 estimates a small Behavioural NK model with trend inflation. A formal
test for parameters identification shows that core reinforcement learning parameters
can only be jointly identified using higher-order approximations of the model while
expanding the information set with a measure of the share of agents adopting a
specific expectations mechanism. Thus, a proxy for the share of näive agents based
on the Survey of Consumer Expectations by the University of Michigan is exploited
for estimating the intensity of choice and the memory parameters with the Bayesian
non-linear filtering technique selected in Chapter 1 - the second-order Extended
Kalman filter. Model estimates outperform a rational expectation counterpart in
matching higher-order empirical moments.