Discussion of "Explaining Business Cycle: News Versus Data

Download Report

Transcript Discussion of "Explaining Business Cycle: News Versus Data

“Explaining business cycles: News versus data
revisions”
Levine, Pearlman and Yang
Discussion
Frank Smets
European Central Bank
MONFISPOL Final Conference
Frankfurt am Main, 19-20 September 2011
Overview
• This paper is part of a larger research project in
which Levine and Pearlman with co-authors analyse
the empirical relevance of imperfect information for
business cycle analysis.
• This builds on the work of Pearlman et al (1986),
Svensson and Woodford (2000) and others: Rational
expectations; perfect knowledge of the structure of
the economy; possibly imperfect knowledge of the
nature of the shock processes.
• Different from other models of imperfect
information:
– Rational inattention (Sims, Mackowiak and Wiederholt, …);
– Bayesian learning of structural parameters (Wieland, …);
Overview
• In this paper, Levine et al compare a simple model
with and without news shocks using revised final
data and a model with and without imperfect
information using real-time data.
– The model is the basic New Keynesian model without
habit formation or inflation indexation
– Only real GDP and inflation are used as observable
variables. What about the short-term interest rate?
• The paper mixes three elements: news shocks,
imperfect information and real-time data, but does
not systematically explore all.
• The paper is incomplete. Need better description
of what is done and more validation.
Overview
Overview
Overview: Results
•
Under perfect information (AI), the specification
of the model with an AR(1) process is marginally
more likely than the one which also includes a
news process, ARMA(1,4).
•
Using the real-time data set including two
revisions, the imperfect information case without
news outperforms the other specifications.
However, it is not clear what it means to use the
real-time data set under the assumption of AI.
Questions
•
How well are the various shocks and parameters
identified?
–
E.g. Is it really possible to distinguish the AR(1) and
news shocks and test their relative importance? It
would be good to show the variance decomposition and
the impulse responses to analyse the differences
between both shocks.
–
Compare the prior and posterior likelihood function
around the estimated parameters.
–
Apply the method of Pesaran et al (2011) to test
identification.
–
How does one decide on the form of the news process?
Questions
•
Which information set should one use? Which
set of shocks should this be compared with?
–
E.g. the model also contains hours worked. Observing
this variable would be very helpful in pinning down the
productivity process.
–
How robust are the results to alternative informational
assumptions?
–
One drawback of the Bayesian approach is that it is not
that easy to compare results using different data sets.
The marginal likelihood is conditional on the observed
data.
Questions
•
The most original part of the paper is the use of
real-time data to identify expected future
innovations.
–
Question: Can the inference of all shocks change as
measurement error hits?
–
It would be good to show how much of the actual
movements in output and inflation are due to
measurement error;
–
What are the impulse responses to measurement error
shocks? How do the impulse responses differ across the
AI and II case?
–
Show how some key unobservable concepts like the
output gap change.
Questions
•
The most original part of the paper is the use of
real-time data to identify expected future
innovations.
–
How do you deal with the later revisions?
–
Describe the size of the measurement errors in the
data section.
–
The paper need to better explain why with the realtime data set the II case does so much better.
–
Does the same hold true if only the revised data set is
used? Why wasn’t the II case considered in the case
without measurement error.