This action might not be possible to undo. Are you sure you want to continue?
User Guide
Tommaso Mancini Griffoli, 20072008
An introduction to
the solution & estimation of DSGE models
Dynare v4  User Guide
Public beta version
Tommaso Mancini Griﬀoli
tommaso.mancini@stanfordalumni.org
This draft: June 2010
iii
Copyright c 20072010 Tommaso Mancini Griﬀoli
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License, Version 1.3 or any
later version published by the Free Software Foundation; with no Invariant
Sections, no FrontCover Texts, and no BackCover Texts.
A copy of the license can be found at: http://www.gnu.org/licenses/
fdl.txt
Contents
Contents iv
List of Figures vii
1 Introduction 1
1.1 About this Guide  approach and structure . . . . . . . . . . . 1
1.2 What is Dynare? . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Additional sources of help . . . . . . . . . . . . . . . . . . . . . 4
1.4 Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 v4, what’s new and backward compatibility . . . . . . . . . . . 5
2 Installing Dynare 7
2.1 Dynare versions . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 System requirements . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Installing Dynare . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 MATLAB particularities . . . . . . . . . . . . . . . . . . . . . . 8
3 Solving DSGE models  basics 9
3.1 A fundamental distinction . . . . . . . . . . . . . . . . . . . . . 9
3.1.1 NOTE! Deterministic vs stochastic models . . . . . . . . 10
3.2 Introducing an example . . . . . . . . . . . . . . . . . . . . . . 11
3.3 Dynare .mod ﬁle structure . . . . . . . . . . . . . . . . . . . . . 15
3.4 Filling out the preamble . . . . . . . . . . . . . . . . . . . . . . 15
3.4.1 The deterministic case . . . . . . . . . . . . . . . . . . . 16
3.4.2 The stochastic case . . . . . . . . . . . . . . . . . . . . . 16
3.4.3 Comments on your ﬁrst lines of Dynare code . . . . . . 17
3.5 Specifying the model . . . . . . . . . . . . . . . . . . . . . . . . 18
3.5.1 Model in Dynare notation . . . . . . . . . . . . . . . . . 18
3.5.2 General conventions . . . . . . . . . . . . . . . . . . . . 19
3.5.3 Notational conventions . . . . . . . . . . . . . . . . . . . 19
3.5.4 Timing conventions . . . . . . . . . . . . . . . . . . . . 19
3.5.5 Conventions specifying nonpredetermined variables . . 20
3.5.6 Linear and loglinearized models . . . . . . . . . . . . . 20
3.6 Specifying steady states and/or initial values . . . . . . . . . . 21
iv
CONTENTS v
3.6.1 Stochastic models and steady states . . . . . . . . . . . 21
3.6.2 Deterministic models and initial values . . . . . . . . . . 23
3.6.3 Finding a steady state . . . . . . . . . . . . . . . . . . . 23
3.6.4 Checking system stability . . . . . . . . . . . . . . . . . 24
3.7 Adding shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.7.1 Deterministic models  temporary shocks . . . . . . . . 25
3.7.2 Deterministic models  permanent shocks . . . . . . . . 25
3.7.3 Stochastic models . . . . . . . . . . . . . . . . . . . . . 27
3.8 Selecting a computation . . . . . . . . . . . . . . . . . . . . . . 27
3.8.1 For deterministic models . . . . . . . . . . . . . . . . . . 28
3.8.2 For stochastic models . . . . . . . . . . . . . . . . . . . 28
3.9 The complete .mod ﬁle . . . . . . . . . . . . . . . . . . . . . . . 31
3.9.1 The stochastic model . . . . . . . . . . . . . . . . . . . . 31
3.9.2 The deterministic model (case of temporary shock) . . . 32
3.10 File execution and results . . . . . . . . . . . . . . . . . . . . . 33
3.10.1 Results  stochastic models . . . . . . . . . . . . . . . . 33
3.10.2 Results  deterministic models . . . . . . . . . . . . . . 34
4 Solving DSGE models  advanced topics 37
4.1 Dynare features and functionality . . . . . . . . . . . . . . . . . 37
4.1.1 Other examples . . . . . . . . . . . . . . . . . . . . . . . 37
4.1.2 Alternative, complete example . . . . . . . . . . . . . . 38
4.1.3 Finding, saving and viewing your output . . . . . . . . . 41
4.1.4 Referring to external ﬁles . . . . . . . . . . . . . . . . . 42
4.1.5 Inﬁnite eigenvalues . . . . . . . . . . . . . . . . . . . . . 43
4.2 Files created by Dynare . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Modeling tips . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.1 Stationarizing your model . . . . . . . . . . . . . . . . . 44
4.3.2 Expectations taken in the past . . . . . . . . . . . . . . 44
4.3.3 Inﬁnite sums . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.4 Inﬁnite sums with changing timing of expectations . . . 46
5 Estimating DSGE models  basics 47
5.1 Introducing an example . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Declaring variables and parameters . . . . . . . . . . . . . . . . 48
5.3 Declaring the model . . . . . . . . . . . . . . . . . . . . . . . . 48
5.4 Declaring observable variables . . . . . . . . . . . . . . . . . . . 49
5.5 Specifying the steady state . . . . . . . . . . . . . . . . . . . . 49
5.6 Declaring priors . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.7 Launching the estimation . . . . . . . . . . . . . . . . . . . . . 52
5.8 The complete .mod ﬁle . . . . . . . . . . . . . . . . . . . . . . . 55
5.9 Interpreting output . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.9.1 Tabular results . . . . . . . . . . . . . . . . . . . . . . . 57
5.9.2 Graphical results . . . . . . . . . . . . . . . . . . . . . . 57
6 Estimating DSGE models  advanced topics 61
6.1 Alternative and nonstationary example . . . . . . . . . . . . . 61
6.1.1 Introducing the example . . . . . . . . . . . . . . . . . . 61
6.1.2 Declaring variables and parameters . . . . . . . . . . . . 66
6.1.3 The origin of nonstationarity . . . . . . . . . . . . . . . 66
6.1.4 Stationarizing variables . . . . . . . . . . . . . . . . . . 67
6.1.5 Linking stationary variables to the data . . . . . . . . . 68
6.1.6 The resulting model block of the .mod ﬁle . . . . . . . . 68
6.1.7 Declaring observable variables . . . . . . . . . . . . . . . 69
6.1.8 Declaring trends in observable variables . . . . . . . . . 69
6.1.9 Declaring unit roots in observable variables . . . . . . . 70
6.1.10 Specifying the steady state . . . . . . . . . . . . . . . . 71
6.1.11 Declaring priors . . . . . . . . . . . . . . . . . . . . . . 71
6.1.12 Launching the estimation . . . . . . . . . . . . . . . . . 71
6.1.13 The complete .mod ﬁle . . . . . . . . . . . . . . . . . . . 72
6.1.14 Summing it up . . . . . . . . . . . . . . . . . . . . . . . 74
6.2 Comparing models based on their posterior distributions . . . . 74
6.3 Where is your output stored? . . . . . . . . . . . . . . . . . . . 75
7 Solving DSGE models  Behind the scenes of Dynare 77
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
7.2 What is the advantage of a second order approximation? . . . . 77
7.3 How does dynare solve stochastic DSGE models? . . . . . . . . 78
8 Estimating DSGE models  Behind the scenes of Dynare 81
8.1 Advantages of Bayesian estimation . . . . . . . . . . . . . . . . 81
8.2 The basic mechanics of Bayesian estimation . . . . . . . . . . . 82
8.2.1 Bayesian estimation: somewhere between calibration and
maximum likelihood estimation  an example . . . . . . 84
8.3 DSGE models and Bayesian estimation . . . . . . . . . . . . . . 85
8.3.1 Rewriting the solution to the DSGE model . . . . . . . 85
8.3.2 Estimating the likelihood function of the DSGE model . 86
8.3.3 Finding the mode of the posterior distribution . . . . . 87
8.3.4 Estimating the posterior distribution . . . . . . . . . . . 87
8.4 Comparing models using posterior distributions . . . . . . . . . 90
9 Optimal policy under commitment 93
10 Troubleshooting 95
vi
List of Figures vii
Bibliography 97
List of Figures
1.1 Dynare, a bird’s eyeview . . . . . . . . . . . . . . . . . . . . . . . . 3
3.1 Structure of the .mod ﬁle . . . . . . . . . . . . . . . . . . . . . . . 16
6.1 CIA model illustration . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.2 Steps of model estimation . . . . . . . . . . . . . . . . . . . . . . . 74
8.1 Illustration of the MetropolisHastings algorithm . . . . . . . . . . 89
Work in Progress!
This is the second version of the Dynare User Guide which is still work in
progress! This means two things. First, please read this with a critical eye
and send me comments! Are some areas unclear? Is anything plain wrong?
Are some sections too wordy, are there enough examples, are these clear? On
the contrary, are there certain parts that just click particularly well? How can
others be improved? I’m very interested to get your feedback.
The second thing that a work in progress manuscript comes with is a few
internal notes. These are mostly placeholders for future work, notes to myself
or others of the Dynare development team, or at times notes to you  our read
ers  to highlight a feature not yet fully stable. Any such notes are marked
with two stars (**).
Thanks very much for your patience and good ideas. Please write either
direclty to myself: tommaso.mancini@stanfordalumni.org, or preferably on
the Dynare Documentation Forum available in the Dynare Forums.
ix
Contacts and Credits
Dynare was originally developed by Michel Juillard in Paris, France. Cur
rently, the development team of Dynare is composed of
• St´ephane Adjemian (stephane.adjemian“AT”ens.fr)
• Houtan Bastani (houtan.bastani“AT”ens.fr)
• Michel Juillard (michel.juillard“AT”mjui.fr)
• Ferhat Mihoubi (ferhat.mihoubi“AT”univevry.fr)
• George Perendia (george“AT”perendia.orangehome.co.uk)
• Marco Ratto (marco.ratto“AT”jrc.ec.europa.eu)
• S´ebastien Villemot (sebastien.villemot“AT”ens.fr)
Several parts of Dynare use or have strongly beneﬁted from publicly avail
able programs by G. Anderson, F. Collard, L. Ingber, O. Kamenik, P. Klein,
S. Sakata, F. Schorfheide, C. Sims, P. Soederlind and R. Wouters.
Finally, the development of Dynare could not have come such a long ways
withough an active community of users who continually pose questions, re
port bugs and suggest new features. The help of this community is gratefully
acknowledged.
The email addresses above are provided in case you wish to contact any
one of the authors of Dynare directly. We nonetheless encourage you to ﬁrst
use the Dynare forums to ask your questions so that other users can beneﬁt
from them as well; remember, almost no question is speciﬁc enough to interest
just one person, and yours is not the exception!
xi
Chapter 1
Introduction
Welcome to Dynare!
1.1 About this Guide  approach and structure
This User Guide aims to help you master Dynare’s main functionalities,
from getting started to implementing advanced features. To do so, this Guide
is structured around examples and oﬀers practical advice. To root this un
derstanding more deeply, though, this Guide also gives some background on
Dynare’s algorithms, methodologies and underlying theory. Thus, a secondary
function of this Guide is to serve as a basic primer on DSGE model solving
and Bayesian estimation.
This Guide will focus on the most common or useful features of the pro
gram, thus emphasizing depth over breadth. The idea is to get you to
use 90% of the program well and then tell you where else to look if you’re
interested in ﬁne tuning or advanced customization.
This Guide is written mainly for an advanced economist  like a pro
fessor, graduate student or central banker  needing a powerful and ﬂexible
program to support and facilitate his or her research activities in a variety
of ﬁelds. The sophisticated computer programmer, on the one hand, or the
specialist of computational economics, on the other, may not ﬁnd this Guide
suﬃciently detailed.
We recognize that the “advanced economist” may be either a beginning
or intermediate user of Dynare. This Guide is written to accommodate both.
If you’re new to Dynare, we recommend starting with chapters 3 and 5,
which introduce the program’s basic features to solve (including running im
pulse response functions) and estimate DSGE models, respectively. To do
1
2 CHAPTER 1. INTRODUCTION
so, these chapters lead you through a complete handson example, which we
recommend following from A to Z, in order to “learn by doing”. Once you
have read these two chapters, you will know the crux of Dynare’s functionality
and (hopefully!) feel comfortable using Dynare for your own work. At that
point, though, you will probably ﬁnd yourself coming back to the User Guide
to skim over some of the content in the advanced chapters to iron out details
and potential complications you may run into.
If you’re instead an intermediate user of Dynare, you will most likely
ﬁnd the advanced chapters, 4 and 6, more appropriate. These chapters cover
more advanced features of Dynare and more complicated usage scenarios. The
presumption is that you would skip around these chapters to focus on the top
ics most applicable to your needs and curiosity. Examples are therefore more
concise and speciﬁc to each feature; these chapters read a bit more like a ref
erence manual.
We also recognize that you probably have had repeated if not active ex
posure to programming and are likely to have a strong economic background.
Thus, a black box solution to your needs is inadequate. To hopefully address
this issue, the User Guide goes into some depth in covering the theoreti
cal underpinnings and methodologies that Dynare follows to solve
and estimate DSGE models. These are available in the “behind the scenes of
Dynare” chapters 7 and 8. These chapters can also serve as a basic primer
if you are new to the practice of DSGE model solving and Bayesian estimation.
Finally, besides breaking up content into short chapters, we’ve introduced
two diﬀerent markers throughout the Guide to help streamline your reading.
• TIP! introduces advice to help you work more eﬃciently with Dynare
or solve common problems.
• NOTE! is used to draw your attention to particularly important infor
mation you should keep in mind when using Dynare.
1.2 What is Dynare?
Before we dive into the thick of the “trees”, let’s have a look at the “forest”
from the top . . . just what is Dynare?
Dynare is a powerful and highly customizable engine, with an
intuitive frontend interface, to solve, simulate and estimate DSGE
models.
1.2. WHAT IS DYNARE? 3
Figure 1.1: The .mod ﬁle being read by the Dynare preprocessor, which then
calls the relevant Matlab routines to carry out the desired operations and
display the results.
In slightly less ﬂowery words, it is a preprocessor and a collection of Mat
lab routines that has the great advantages of reading DSGE model equations
written almost as in an academic paper. This not only facilitates the inputting
of a model, but also enables you to easily share your code as it is straightfor
ward to read by anyone.
Figure 1.2 gives you an overview of the way Dynare works. Basically, the
model and its related attributes, like a shock structure for instance, is writ
ten equation by equation in an editor of your choice. The resulting ﬁle will
be called the .mod ﬁle. That ﬁle is then called from Matlab. This initiates
the Dynare preprocessor which translates the .mod ﬁle into a suitable input
for the Matlab routines (more precisely, it creates intermediary Matlab or C
ﬁles which are then used by Matlab code) used to either solve or estimate the
model. Finally, results are presented in Matlab. Some more details on the
internal ﬁles generated by Dynare is given in section 4.2 in chapter 4.
Each of these steps will become clear as you read through the User Guide,
but for now it may be helpful to summarize what Dynare is able to do:
4 CHAPTER 1. INTRODUCTION
• compute the steady state of a model
• compute the solution of deterministic models
• compute the ﬁrst and second order approximation to solutions of stochas
tic models
• estimate parameters of DSGE models using either a maximum likelihood
or a Bayesian approach
• compute optimal policies in linearquadratic models
1.3 Additional sources of help
While this User Guide tries to be as complete and thorough as possible, you
will certainly want to browse other material for help, as you learn about new
features, struggle with adapting examples to your own work, and yearn to ask
that one question whose answer seems to exist nowhere. At your disposal,
you have the following additional sources of help:
• Reference Manual: this manual covers all Dynare commands, giving
a clear deﬁnition and explanation of usage for each. The User Guide
will often introduce you to a command in a rather loose manner (mainly
through examples); so reading corresponding command descriptions in
the Reference Manual is a good idea to cover all relevant details.
• Oﬃcial online examples: the Dynare website includes other examples
 usually well documented  of .mod ﬁles covering models and method
ologies introduced in recent papers.
• Dynare forums: this lively online discussion forum allows you to ask
your questions openly and read threads from others who might have run
into similar diﬃculties.
• Frequently Asked Questions (FAQ): this section of the Dynare site
emphasizes a few of the most popular questions in the forums.
• DSGE.net: this website, run my members of the Dynare team, is a
resource for all scholars working in the ﬁeld of DSGE modeling. Besides
allowing you to stay up to date with the most recent papers and possi
bly make new contacts, it conveniently lists conferences, workshops and
seminars that may be of interest.
1.4. NOMENCLATURE 5
1.4 Nomenclature
To end this introduction and avoid confusion in what follows, it is worthwhile
to agree on a few deﬁnitions of terms. Many of these are shared with the
Reference Manual.
• Integer indicates an integer number.
• Double indicates a double precision number. The following syntaxes
are valid: 1.1e3, 1.1E3, 1.1E3, 1.1d3, 1.1D3
• Expression indicates a mathematical expression valid in the underlying
language (e.g. Matlab).
• Variable name indicates a variable name. NOTE! These must start
with an alphabetical character and can only contain other alphabetical
characters and digits, as well as underscores ( ). All other characters,
including accents, and spaces, are forbidden.
• Parameter name indicates a parameter name which must follow the
same naming conventions as above.
• Filename indicates a ﬁle name valid in your operating system. Note
that Matlab requires that names of ﬁles or functions start with alpha
betical characters; this concerns your Dynare .mod ﬁles.
• Command is an instruction to Dynare or other program when speciﬁed.
• Options or optional arguments for a command are listed in square
brackets [ ] unless otherwise noted. If, for instance, the option must
be speciﬁed in parenthesis in Dynare, it will show up in the Guide as
[(option)].
• Typewritten text indicates text as it should appear in Dynare code.
1.5 v4, what’s new and backward compatibility
The current version of Dynare  for which this guide is written  is version
4. With respect to version 3, this new version introduces several important
features, as well as improvements, optimizations of routines and bug ﬁxes.
The major new features are the following:
• Analytical derivatives are now used everywhere (for instance, in the
Newton algorithm for deterministic models and in the linearizations nec
essary to solve stochastic models). This increases computational speed
signiﬁcantly. The drawback is that Dynare can now handle only a lim
ited set of functions, although in nearly all economic applications this
should not be a constraint.
6 CHAPTER 1. INTRODUCTION
• Variables and parameters are now kept in the order in which they are
declared whenever displayed and when used internally by Dynare. Recall
that in version 3, variables and parameters where at times in their order
of declaration and at times in alphabetical order. NOTE! This may
cause some problems of backward compatibility if you wrote programs
to run oﬀ Dynare v3 output.
• The names of many internal variables and the organization of output
variables has changed. These are enumerated in details in the relevant
chapters. The names of the ﬁles internally generated by Dynare have
also changed. (** more on this when explaining internal ﬁle structure 
TBD)
• The syntax for the external steady state ﬁle has changed. This is cov
ered in more details in chapter 3, in section 3.6.3. NOTE! You will
unfortunately have to slightly amend any old steady state ﬁles you may
have written.
• Speed. Several largescale improvements have been implemented to
speed up Dynare. This should be most noticeable when solving de
terministic models, but also apparent in other functionality.
Chapter 2
Installing Dynare
2.1 Dynare versions
The current version of Dynare (4.1) runs on both MATLAB and GNU
Octave.
There used to be versions of Dynare for Scilab and Gauss. Development
of the Scilab version stopped after Dynare version 3.02 and that for Gauss
after Dynare version 1.2.
This User Guide will exclusively focus on Dynare version 4.0 and
later.
You may also be interested by another program, Dynare++, which is
a standalone C++ program specialized in computing korder approximations
of dynamic stochastic general equilibrium models. Note that Dynare++ is
distributed along with Dynare since version 4.1. See the Dynare++ webpage
for more information.
2.2 System requirements
Dynare can run on Microsoft Windows, as well as Unixlike operating systems,
in particular GNU/Linux and Mac OS X. If you have questions about the
support of a particular platform, please ask your question on Dynare forums.
To run Dynare, it is recommended to allocate at least 256MB of RAM
to the platform running Dynare, although 512MB is preferred. Depending on
the type of computations required, like the very processor intensive Metropolis
Hastings algorithm, you may need up to 1GB of RAM to obtain acceptable
computational times.
2.3 Installing Dynare
Please refer to the section entitled “Installation and conﬁguration” in the
Dynare reference manual.
7
8 CHAPTER 2. INSTALLING DYNARE
2.4 MATLAB particularities
A question often comes up: what special MATLAB toolboxes are necessary
to run Dynare? In fact, no additional toolbox is necessary for running most of
Dynare, except maybe for optimal simple rules (see chapter 9), but even then
remedies exist (see the Dynare forums for discussions on this, or to ask your
particular question). But if you do have the ‘optimization toolbox’ installed,
you will have additional options for solving for the steady state (solve algo
option) and for searching for the posterior mode (mode compute option), both
of which are deﬁned later.
Chapter 3
Solving DSGE models  basics
This chapter covers everything that leads to, and stems from, the solution
of DSGE models; a vast terrain. That is to say that the term “solution”
in the title of the chapter is used rather broadly. You may be interested in
simply ﬁnding the solution functions to a set of ﬁrst order conditions stemming
from your model, but you may also want to go a bit further. Typically, you
may be interested in how this system behaves in response to shocks, whether
temporary or permanent. Likewise, you may want to explore how the system
comes back to its steady state or moves to a new one. This chapter covers all
these topics. But instead of skipping to the topic closest to your needs, we
recommend that you read this chapter chronologically, to learn basic Dynare
commands and the process of writing a proper .mod ﬁle  this will serve as a
base to carry out any of the above computations.
3.1 A fundamental distinction
Before speaking of Dynare, it is important to recognize a distinction in model
types. This distinction will appear throughout the chapter; in fact, it is so
fundamental, that we considered writing separate chapters altogether. But
the amount of common material  Dynare commands and syntax  is notable
and writing two chapters would have been overly repetitive. Enough suspense;
here is the important question: is your model stochastic or determinis
tic?
The distinction hinges on whether future shocks are known. In de
terministic models, the occurrence of all future shocks is known exactly at
the time of computing the model’s solution. In stochastic models, instead,
only the distribution of future shocks is known. Let’s consider a shock to a
model’s innovation only in period 1. In a deterministic context, agents will
take their decisions knowing that future values of the innovations will be zero
in all periods to come. In a stochastic context, agents will take their decisions
9
10 CHAPTER 3. SOLVING DSGE MODELS  BASICS
knowing that the future value of innovations are random but will have zero
mean. This isn’t the same thing because of Jensen’s inequality. Of course, if
you consider only a ﬁrst order linear approximation of the stochastic model,
or a linear model, the two cases become practically the same, due to certainty
equivalence. A second order approximation will instead lead to very diﬀerent
results, as the variance of shocks will matter.
The solution method for each of these model types diﬀers signiﬁcantly. In
deterministic models, a highly accurate solution can be found by numerical
methods. The solution is nothing more than a series of numbers that match
a given set of equations. Intuitively, if an agent has perfect foresight, she can
specify today  at the time of making her decision  what each of her precise
actions will be in the future. In a stochastic environment, instead, the best
the agent can do is specify a decision, policy or feedback rule for the future:
what will her optimal actions be contingent on each possible realization of
shocks. In this case, we therefore search for a function satisfying the model’s
ﬁrst order conditions. To complicate things, this function may be nonlinear
and thus needs to be approximated. In control theory, solutions to determin
istic models are usually called “closed loop” solutions, and those to stochastic
models are referred to as “open loop”.
Because this distinction will resurface again and again throughout the
chapter, but also because it has been a source of signiﬁcant confusion in the
past, the following gives some additional details.
3.1.1 NOTE! Deterministic vs stochastic models
Deterministic models have the following characteristics:
1. As the DSGE (read, “stochastic”, i.e. not deterministic!) literature
has gained attention in economics, deterministic models have become
somewhat rare. Examples include OLG models without aggregate un
certainty.
2. These models are usually introduced to study the impact of a change in
regime, as in the introduction of a new tax, for instance.
3. Models assume full information, perfect foresight and no uncertainty
around shocks.
4. Shocks can hit the economy today or at any time in the future, in which
case they would be expected with perfect foresight. They can also last
one or several periods.
5. Most often, though, models introduce a positive shock today and zero
shocks thereafter (with certainty).
3.2. INTRODUCING AN EXAMPLE 11
6. The solution does not require linearization, in fact, it doesn’t even really
need a steady state. Instead, it involves numerical simulation to ﬁnd the
exact paths of endogenous variables that meet the model’s ﬁrst order
conditions and shock structure.
7. This solution method can therefore be useful when the economy is far
away from steady state (when linearization oﬀers a poor approximation).
Stochastic models, instead, have the following characteristics:
1. These types of models tend to be more popular in the literature. Exam
ples include most RBC models, or new keynesian monetary models.
2. In these models, shocks hit today (with a surprise), but thereafter their
expected value is zero. Expected future shocks, or permanent changes
in the exogenous variables cannot be handled due to the use of Taylor
approximations around a steady state.
3. Note that when these models are linearized to the ﬁrst order, agents
behave as if future shocks where equal to zero (since their expectation is
null), which is the certainty equivalence property. This is an often
overlooked point in the literature which misleads readers in supposing
their models may be deterministic.
3.2 Introducing an example
The goal of this ﬁrst section is to introduce a simple example. Future sections
will aim to code this example into Dynare and analyze its salient features
under the inﬂuence of shocks  both in a stochastic and a deterministic envi
ronment. Note that as a general rule, the examples in the basic chapters, 3
and 5, are kept as bare as possible, with just enough features to help illustrate
Dynare commands and functionalities. More complex examples are instead
presented in the advanced chapters.
The model introduced here is a basic RBC model with monopolistic com
petition, used widely in the literature. Its particular notation adopted below
is drawn mostly from notes available on Jesus FernandezVillaverde’s very
instructive website; this is a good place to look for additional information
on any of the following model setup and discussion. Note throughout this
model description that the use of expectation signs is really only relevant
in a stochastic setting, as per the earlier discussion. We will nonetheless
illustrate both the stochastic and the deterministic settings on the basis of
this example. Thus, when thinking of the latter, you’ll have to use a bit of
imagination (on top of that needed to think you have perfect foresight!) to
12 CHAPTER 3. SOLVING DSGE MODELS  BASICS
ignore the expectation signs.
Households maximize utility over consumption, c
t
and leisure, 1−l
t
, where
l
t
is labor input, according to the following utility function
E
t
∞
t=0
β [log c
t
+ψ log(1 −l
t
)]
and subject to the following budget constraint
c
t
+k
t+1
= w
t
l
t
+r
t
k
t
+ (1 −δ)k
t
, ∀t > 0
where k
t
is capital stock, w
t
real wages, r
t
real interest rates or cost of capital
and δ the depreciation rate.
The above equation can be seen as an accounting identity, with total ex
penditures on the left hand side and revenues  including the liquidation value
of the capital stock  on the right hand side. Alternatively, with a little more
imagination, the equation can also be interpreted as a capital accumulation
equation after bringing c
t
to the right hand side and noticing that w
t
l
t
+r
t
k
t
,
total payments to factors, equals y
t
, or aggregate output, by the zero proﬁt
condition. As a consequence, if we deﬁne investment as i
t
= y
t
−c
t
, we obtain
the intuitive result that i
t
= k
t+1
− (1 − δ)k
t
, or that investment replenishes
the capital stock thereby countering the eﬀects of depreciation. In any given
period, the consumer therefore faces a tradeoﬀ between consuming and in
vesting in order to increase the capital stock and consuming more in following
periods (as we will see later, production depends on capital).
Maximization of the household problem with respect to consumption,
leisure and capital stock, yields the Euler equation in consumption, capturing
the intertemporal tradeoﬀ mentioned above, and the labor supply equation
linking labor positively to wages and negatively to consumption (the wealth
ier, the more leisure due to the decreasing marginal utility of consumption).
These equation are
1
c
t
= βE
t
_
1
c
t+1
(1 +r
t+1
−δ)
_
and
ψ
c
t
1 −l
t
= w
t
The ﬁrm side of the problem is slightly more involved, due to monopolistic
competition, but is presented below in the simplest possible terms, with a
little handwaiving involved, as the derivations are relatively standard.
3.2. INTRODUCING AN EXAMPLE 13
There are two ways to introduce monopolistic competition. We can ei
ther assume that ﬁrms sell diﬀerentiated varieties of a good to consumers who
aggregate these according to a CES index. Or we can postulate that there
is a continuum of intermediate producers with market power who each sell
a diﬀerent variety to a competitive ﬁnal goods producer whose production
function is a CES aggregate of intermediate varieties.
If we follow the second route, the ﬁnal goods producer chooses his or her
optimal demand for each variety, yielding the DixitStiglitz downward sloping
demand curve. Intermediate producers, instead, face a two pronged decision:
how much labor and capital to employ given these factors’ perfectly competi
tive prices and how to price the variety they produce.
Production of intermediate goods follows a CRS production function de
ﬁned as
y
it
= k
α
it
(e
z
t
l
it
)
1−α
where the i subscript stands for ﬁrm i of a continuum of ﬁrms between zero
and one and where α is the capital elasticity in the production function, with
0 < α < 1. Also, z
t
captures technology which evolves according to
z
t
= ρz
t−1
+e
t
where ρ is a parameter capturing the persistence of technological progress and
e
t
∼ N(0, σ).
The solution to the sourcing problem yields an optimal capital to labor
ratio, or relationship between payments to factors:
k
it
r
t
=
α
1 −α
w
t
l
it
The solution to the pricing problem, instead, yields the wellknown con
stant markup pricing condition of monopolistic competition:
p
it
=
−1
mc
t
p
t
where p
it
is ﬁrm i’s speciﬁc price, mc
t
is real marginal cost and p
t
is the aggre
gate CES price or average price. An additional step simpliﬁes this expression:
symmetric ﬁrms implies that all ﬁrms charge the same price and thus p
it
= p
t
;
we therefore have: mc
t
= ( −1)/
But what are marginal costs equal to? To ﬁnd the answer, we combine the
optimal capital to labor ratio into the production function and take advantage
of its CRS property to solve for the amount of labor or capital required to
produce one unit of output. The real cost of using this amount of any one
14 CHAPTER 3. SOLVING DSGE MODELS  BASICS
factor is given by w
t
l
it
+ r
t
k
it
where we substitute out the payments to the
other factor using again the optimal capital to labor ratio. When solving for
labor, for instance, we obtain
mc
t
=
_
1
1 −α
_
1−α
_
1
α
_
α
1
A
t
w
1−α
t
r
α
t
which does not depend on i; it is thus the same for all ﬁrms.
Interestingly, the above can be worked out, by using the optimal capital
to labor ratio, to yield w
t
[(1 − α)y
it
/l
it
]
−1
, or w
t
∂l
it
∂y
it
, which is the deﬁnition
of marginal cost: the cost in terms of labor input of producing an additional
unit of output. This should not be a surprise since the optimal capital to
labor ratio follows from the maximization of the production function (minus
real costs) with respect to its factors.
Combining this result for marginal cost, as well as its counterpart in terms
of capital, with the optimal pricing condition yields the ﬁnal two important
equations of our model
w
t
= (1 −α)
y
it
l
it
( −1)
and
r
t
= α
y
it
k
it
( −1)
To end, we aggregate the production of each individual ﬁrm to ﬁnd an
aggregate production function. On the supply side, we factor out the capital
to labor ratio, k
t
/l
t
, which is the same for all ﬁrms and thus does not depend
on i. On the other side, we have the DixitStiglitz demand for each variety. By
equating the two and integrating both side, and noting that price dispersion
is null  or that, as hinted earlier, p
it
= p
t
 we obtain aggregate production
y
t
= A
t
k
α
t
l
1−α
t
which can be shown is equal to the aggregate amount of varieties bought by
the ﬁnal good producer (according to a CES aggregation index) and, in turn,
equal to the aggregate output of ﬁnal good, itself equal to household con
sumption. Note, to close, that because the ratio of output to each factor is
the same for each intermediate ﬁrm and that ﬁrm speciﬁc as well as aggre
gate production is CRS, we can rewrite the above two equations for w
t
and r
t
without the i subscripts on the right hand side.
This ends the exposition of the example. Now, let’s roll up our sleeves and
see how we can input the model into Dynare and actually test how the model
will respond to shocks.
3.3. DYNARE .MOD FILE STRUCTURE 15
3.3 Dynare .mod ﬁle structure
Input into Dynare involves the .mod ﬁle, as mentioned loosely in the intro
duction of this Guide. The .mod ﬁle can be written in any editor, external or
internal to Matlab. It will then be read by Matlab by ﬁrst navigating within
Matlab to the directory where the .mod ﬁle is stored and then by typing in
the Matlab command line Dynare filename.mod; (although actually typing
the extension .mod is not necessary). But before we get into executing a .mod
ﬁle, let’s start by writing one!
It is convenient to think of the .mod ﬁle as containing four distinct blocks,
illustrated in ﬁgure 3.3:
• preamble: lists variables and parameters
• model: spells out the model
• steady state or initial value: gives indications to ﬁnd the steady state
of a model, or the starting point for simulations or impulse response
functions based on the model’s solution.
• shocks: deﬁnes the shocks to the system
• computation: instructs Dynare to undertake speciﬁc operations (e.g.
forecasting, estimating impulse response functions)
Our exposition below will structured according to each of these blocks.
3.4 Filling out the preamble
The preamble generally involves three commands that tell Dynare what are
the model’s variables, which are endogenous and what are the parameters.
The commands are:
• var starts the list of endogenous variables, to be separated by commas.
• varexo starts the list of exogenous variables that will be shocked.
• parameters starts the list of parameters and assigns values to each.
In the case of our example, let’s diﬀerentiate between the stochastic and de
terministic cases. First, we lay these out, then we discuss them.
16 CHAPTER 3. SOLVING DSGE MODELS  BASICS
Figure 3.1: The .mod ﬁle contains ﬁve logically distinct parts.
3.4.1 The deterministic case
The model is inherited exactly as speciﬁed in the earlier description, except
that we no longer need the e
t
variable, as we can make z
t
directly exogenous.
Thus, the preamble would look like:
var y c k i l y l w r;
varexo z;
parameters beta psi delta alpha sigma epsilon;
alpha = 0.33;
beta = 0.99;
delta = 0.023;
psi = 1.75;
sigma = (0.007/(1alpha));
epsilon = 10;
3.4.2 The stochastic case
In this case, we go back to considering the law of motion for technology, con
sisting of an exogenous shock, e
t
. With respect to the above, we therefore
3.4. FILLING OUT THE PREAMBLE 17
adjust the list of endogenous and exogenous variables, and add the parameter
ρ. Here’s what the preamble would look like:
var y c k i l y l w r z;
varexo e;
parameters beta psi delta alpha rho sigma epsilon;
alpha = 0.33;
beta = 0.99;
delta = 0.023;
psi = 1.75;
rho = 0.95;
sigma = (0.007/(1alpha));
epsilon = 10;
3.4.3 Comments on your ﬁrst lines of Dynare code
As you can tell, writing a .mod ﬁle is really quite straightforward. Two quick
comments:
NOTE! Remember that each instruction of the .mod ﬁle must be termi
nated by a semicolon (;), although a single instruction can span two lines if
you need extra space (just don’t put a semicolon at the end of the ﬁrst line).
TIP! You can also comment out any line by starting the line with two
forward slashes (//), or comment out an entire section by starting the section
with /* and ending with */. For example:
var y c k i l y l w r z;
varexo e;
parameters beta psi delta
alpha rho sigma epsilon;
// the above instruction reads over two lines
/*
the following section lists
several parameters which were
calibrated by my coauthor. Ask
her all the difficult questions!
*/
alpha = 0.33;
beta = 0.99;
delta = 0.023;
psi = 1.75;
rho = 0.95;
18 CHAPTER 3. SOLVING DSGE MODELS  BASICS
sigma = (0.007/(1alpha));
epsilon = 10;
3.5 Specifying the model
3.5.1 Model in Dynare notation
One of the beauties of Dynare is that you can input your model’s equa
tions naturally, almost as if you were writing them in an academic paper.
This greatly facilitates the sharing of your Dynare ﬁles, as your colleagues will
be able to understand your code in notime. There are just a few conventions
to follow. Let’s ﬁrst have a look at our model in Dynare notation, and
then go through the various Dynare input conventions. What you can already
try to do is glance at the model block below and see if you can recognize the
equations from the earlier example. See how easy it is to read Dynare code?
model;
(1/c) = beta*(1/c(+1))*(1+r(+1)delta);
psi*c/(1l) = w;
c+i = y;
y = (k(1)^alpha)*(exp(z)*l)^(1alpha);
w = y*((epsilon1)/epsilon)*(1alpha)/l;
r = y*((epsilon1)/epsilon)*alpha/k(1);
i = k(1delta)*k(1);
y l = y/l;
z = rho*z(1)+e;
end;
Just in case you need a hint or two to recognize these equations, here’s
a brief description: the ﬁrst equation is the Euler equation in consumption.
The second the labor supply function. The third the accounting identity. The
fourth is the production function. The ﬁfth and sixth are the marginal cost
equal to markup equations. The seventh is the investment equality. The
eighth an identity that may be useful and the last the equation of motion of
technology.
NOTE! that the above model speciﬁcation corresponds to the stochastic
case; indeed, notice that the law of motion for technology is included, as per
our discussion of the preamble. The corresponding model for the determin
istic casce would simply loose the last equation.
3.5. SPECIFYING THE MODEL 19
3.5.2 General conventions
The above example illustrates the use of a few important commands and
conventions to translate a model into a Dynarereadable .mod ﬁle.
• The ﬁrst thing to notice, is that the model block of the .mod ﬁle begins
with the command model and ends with the command end.
• Second, in between, there need to be as many equations as you declared
endogenous variables (this is actually one of the ﬁrst things that Dynare
checks; it will immediately let you know if there are any problems).
• Third, as in the preamble and everywhere along the .mod ﬁle, each line
of instruction ends with a semicolon (except when a line is too long and
you want to break it across two lines. This is unlike Matlab where if you
break a line you need to add . . . ).
• Fourth, equations are entered one after the other; no matrix representa
tion is necessary. Note that variable and parameter names used in the
model block must be the same as those declared in the preamble; TIP!
remember that variable and parameter names are case sensitive.
3.5.3 Notational conventions
• Variables entering the system with a time t subscript are written plainly.
For example, x
t
would be written x.
• Variables entering the system with a time t − n subscript are written
with (−n) following them. For example, x
t−2
would be written x(−2)
(incidentally, this would count as two backward looking variables).
• In the same way, variables entering the system with a time t+n subscript
are written with (+n) following them. For example, x
t+2
would be
written x(+2). Writing x(2) is also allowed, but this notation makes it
slightly harder to count by hand the number of forward looking variables
(a useful measure to check); more on this below . . .
3.5.4 Timing conventions
• In Dynare, the timing of each variable reﬂects when that variable is de
cided. For instance, our capital stock is not decided today, but yesterday
(recall that it is a function of yesterday’s investment and capital stock);
it is what we call in the jargon a predetermined variable. Thus, even
though in the example presented above we wrote k
t+1
= i
t
+ (1 − δ)k
t
,
as in many papers, we would translate this equation into Dynare as
k=i+(1delta)*k(1).
20 CHAPTER 3. SOLVING DSGE MODELS  BASICS
• As another example, consider that in some wage negociation models,
wages used during a period are set the period before. Thus, in the
equation for wages, you can write wage in period t (when they are set),
but in the labor demand equation, wages should appear with a one
period lag.
• A slightly more roundabout way to explain the same thing is that for
stock variables, you must use a “stock at the end of the period” concept.
It is investment during period t that sets stock at the end of period t.
Be careful, a lot of papers use the “stock at the beginning of the period”
convention, as we did (on purpose to highlight this distinction!) in the
setup of the example model above.
3.5.5 Conventions specifying nonpredetermined variables
• A (+1) next to a variable tells Dynare to count the occurrence of that
variable as a jumper or forwardlooking or nonpredetermined variable.
• BlanchardKahn conditions are met only if the number of nonpredetermined
variables equals the number of eigenvalues greater than one. If this con
dition is not met, Dynare will put up a warning.
• Note that a variable may occur both as predetermined and nonpredetermined.
For instance, consumption could appear with a lead in the Euler equa
tion, but also with a lag in a habit formation equation, if you had one.
In this case, the second order diﬀerence equation would have two eigen
values, one needing to be greater and the other smaller than one for
stability.
3.5.6 Linear and loglinearized models
There are two other variants of the system’s equations which Dynare accom
modates. First, the linear model and second, the model in explogs. In
the ﬁrst case, all that is necessary is to write the term (linear) next to the
command model. Our example, with just the equation for y
l
for illustration,
would look like:
model (linear);
yy l=yy  ll;
end;
where repeating a letter for a variable means diﬀerence from steady state.
Otherwise, you may be interested to have Dynare take Taylor series ex
pansions in logs rather than in levels; this turns out to be a very useful option
3.6. SPECIFYING STEADY STATES AND/OR INITIAL VALUES 21
when estimating models with unit roots, as we will see in chapter 5. If so,
simply rewrite your equations by taking the exponential and logarithm of each
variable. The Dynare input convention makes this very easy to do. Our ex
ample would need to be rewritten as follows (just shown for the ﬁrst two
equations)
model;
(1/exp(cc)) = beta*(1/exp(cc(+1)))*(1+exp(rr(+1))delta);
psi*exp(cc)/(1exp(ll)) = exp(ww);
end;
where, this time, repeating a letter for a variable means log of that variable,
so that the level of a variable is given by exp(repeatedvariable).
3.6 Specifying steady states and/or initial values
Material in this section has created much confusion in the past. But with
some attention to the explanations below, you should get through unscathed.
Let’s start by emphasizing the uses of this section of the .mod ﬁle. First, recall
that stochastic models need to be linearized. Thus, they need to have a steady
state. One of the functions of this section is indeed to provide these steady
state values, or approximations of values. Second, irrespective of whether
you’re working with a stochastic or deterministic model, you may be inter
ested to start your simulations or impulse response functions from either a
steady state, or another given point. This section is also useful to specify this
starting value. Let’s see in more details how all this works.
In passing, though, note that the relevant commands in this section are
initval, endval or, more rarely, histval which is covered only in the Ref
erence Manual. The ﬁrst two are instead covered in what follows.
3.6.1 Stochastic models and steady states
In a stochastic setting, your model will need to be linearized before it is solved.
To do so, Dynare needs to know your model’s steady state (more details on
ﬁnding a steady state, as well as tips to do so more eﬃciently, are provided in
section 3.6.3 below). You can either enter exact steady state values into your
.mod ﬁle, or just approximations and let Dynare ﬁnd the exact steady state
(which it will do using numerical methods based on your approximations). In
either case, these values are entered in the initval block, as in the following
fashion:
initval;
22 CHAPTER 3. SOLVING DSGE MODELS  BASICS
k = 9;
c = 0.7;
l = 0.3;
w = 2.0;
r = 0;
z = 0;
e = 0;
end;
Then, by using the command steady, you can control whether you want to
start your simulations or impulse response functions from the steady state, or
from the exact values you speciﬁed in the initval block. Adding steady just
after your initval block will instruct Dynare to consider your initial values
as mere approximations and start simulations or impulse response functions
from the exact steady state. On the contrary, if you don’t add the command
steady, your simulations or impulse response functions will start from your
initial values, even if Dynare will have calculated your model’s exact steady
state for the purpose of linearization.
For the case in which you would like simulations and impulse response
functions to begin at the steady state, the above block would be expanded to
yield:
initval;
k = 9;
c = 0.7;
l = 0.3;
w = 2.0;
r = 0;
z = 0;
e = 0;
end;
steady;
TIP! If you’re dealing with a stochastic model, remember that its lin
ear approximation is good only in the vicinity of the steady state, thus it is
strongly recommended that you start your simulations from a steady state;
this means either using the command steady or entering exact steady state
values.
3.6. SPECIFYING STEADY STATES AND/OR INITIAL VALUES 23
3.6.2 Deterministic models and initial values
Deterministic models do not need to be linearized in order to be solved. Thus,
technically, you do not need to provide a steady state for these model. But
practically, most researchers are still interested to see how a model reacts to
shocks when originally in steady state. In the deterministic case, the initval
block serves very similar functions as described above. If you wanted to shock
your model starting from a steady state value, you would enter approximate
(or exact) steady state values in the initval block, followed by the command
steady. Otherwise, if you wanted to begin your solution path from an arbi
trary point, you would enter those values in your initval block and not use
the steady command. An illustration of the initval block in the determin
istic case appears further below.
3.6.3 Finding a steady state
The diﬃculty in the above, of course, is calculating actual steady state val
ues. Doing so borders on a form of art, and luck is unfortunately part of the
equation. Yet, the following TIPS! may help.
As mentioned above, Dynare can help in ﬁnding your model’s steady state
by calling the appropriate Matlab functions. But it is usually only successful
if the initial values you entered are close to the true steady state. If you have
trouble ﬁnding the steady state of your model, you can begin by playing with
the options following the steady command. These are:
• solve algo = 0: uses Matlab Optimization Toolbox FSOLVE
• solve algo = 1: uses Dynare’s own nonlinear equation solver
• solve algo = 2: splits the model into recursive blocks and solves each
block in turn.
• solve algo = 3: uses the Sims solver. This is the default option if none
are speciﬁed.
For complicated models, ﬁnding suitable initial values for the endogenous
variables is the trickiest part of ﬁnding the equilibrium of that model. Often,
it is better to start with a smaller model and add new variables one by one.
But even for simpler models, you may still run into diﬃculties in ﬁnding
your steady state. If so, another option is to enter your model in linear
terms. In this case, variables would be expressed in percent deviations from
steady state. Thus, their initial values would all be zero. Unfortunately, if
any of your original (nonlinear) equations involve sums (a likely fact), your
24 CHAPTER 3. SOLVING DSGE MODELS  BASICS
linearized equations will include ratios of steady state values, which you would
still need to calculate. Yet, you may be left needing to calculate fewer steady
state values than in the original, nonlinear, model.
Alternatively, you could also use an external program to calculate ex
act steady state values. For instance, you could write an external Maple
ﬁle and then enter the steady state solution by hand in Dynare. But of
course, this procedure could be time consuming and bothersome, especially
if you want to alter parameter values (and thus steady states) to undertake
robustness checks.
The alternative is to write a Matlab program to ﬁnd your model’s steady
state. Doing so has the clear advantages of being able to incorporate your
Matlab program directly into your .mod ﬁle so that running loops with diﬀer
ent parameter values, for instance, becomes seamless. NOTE! When doing so,
your matlab (.m) ﬁle should have the same name as your .mod ﬁle, followed
by steadystate For instance, if your .mod ﬁle is called example.mod, your
Matlab ﬁle should be called example steadystate.m and should be saved in
the same directory as your .mod ﬁle. Dynare will automatically check the di
rectory where you’ve saved your .mod ﬁle to see if such a Matlab ﬁle exists. If
so, it will use that ﬁle to ﬁnd steady state values regardless of whether you’ve
provided initial values in your .mod ﬁle.
Because Matlab does not work with analytical expressions, though (unless
you’re working with a particular toolbox), you need to do a little work to write
your steady state program. It is not enough to simply input the equations
as you’ve written them in your .mod ﬁle and ask Matlab to solve the system.
You will instead need to write your steady state program as if you were solv
ing for the steady state by hand. That is, you need to input your expressions
sequentially, whereby each lefthand side variable is written in terms of known
parameters or variables already solved in the lines above. For example, the
steady state ﬁle corresponding to the above example, in the stochastic case,
would be: (** example ﬁle to be added shortly)
3.6.4 Checking system stability
TIP! A handy command that you can add after the initval or endval block
(following the steady command if you decide to add one) is the check com
mand. This computes and displays the eigenvalues of your system
which are used in the solution method. As mentioned earlier, a necessary con
dition for the uniqueness of a stable equilibrium in the neighborhood of the
steady state is that there are as many eigenvalues larger than one in modulus
as there are forward looking variables in the system. If this condition is not
3.7. ADDING SHOCKS 25
met, Dynare will tell you that the BlanchardKahn conditions are not satisﬁed
(whether or not you insert the check command).
3.7 Adding shocks
3.7.1 Deterministic models  temporary shocks
When working with a deterministic model, you have the choice of introducing
both temporary and permanent shocks. The distinction is that under a tem
porary shock, the model eventually comes back to steady state, while under
a permanent shock, the model reaches a new steady state. In both cases,
though, the shocks are entirely expected, as explained in our original discus
sion on stochastic and deterministic models.
To work with a temporary shock, you are free to set the duration and
level of the shock. To specify a shock that lasts 9 periods on z
t
, for instance,
you would write:
shocks;
var z;
periods 1:9;
values 0.1;
end;
Given the above instructions, Dynare would replace the value of z
t
spec
iﬁed in the initval block with the value of 0.1 entered above. If variables
were in logs, this would have corresponded to a 10% shock. Note that you
can also use the mshocks command which multiplies the initial value of an
exogenous variable by the mshocks value. Finally, note that we could have
entered future periods in the shocks block, such as periods 5:10, in order to
study the anticipatory behavior of agents in response to future shocks.
3.7.2 Deterministic models  permanent shocks
To study the eﬀects of a permanent shock hitting the economy today, such
as a structural change in your model, you would not specify actual “shocks”,
but would simply tell the system to which (steady state) values you would like
it to move and let Dynare calculate the transition path. To do so, you would
use the endval block following the usual initval block. For instance, you
may specify all values to remain common between the two blocks, except for
the value of technology which you may presume changes permanently. The
corresponding instructions would be:
26 CHAPTER 3. SOLVING DSGE MODELS  BASICS
initval;
k = 9;
c = 0.7;
l = 0.3;
w = 2.0;
r = 0;
z = 0;
end;
steady;
endval;
k = 9;
c = 0.7;
l = 0.3;
w = 2.0;
r = 0;
z = 0.1;
end;
steady;
where steady can also be added to the endval block, and serves the same
functionality as described earlier (namely, of telling Dynare to start and/ or
end at a steady state close to the values you entered. If you do not use steady
after endval, and the latter does not list exact steady state values, you may
impose on your system that it does not return to steady state. This is unusual.
In this case, your problem would become a socalled two boundary problem,
which, when solved, requires that the path of your endogenous variables pass
through the steady state closest to your endval values). In our example, we
make use of the second steady since the actual terminal steady state values
are bound to be somewhat diﬀerent from those entered above, which are noth
ing but the initial values for all variables except for technology.
In the above example, the value of technology would move to 0.1 in pe
riod 1 (tomorrow) and thereafter. But of course, the other variables  the
endogenous variables  will take longer to reach their new steady state values.
TIP! If you instead wanted to study the eﬀects of a permanent but future
shock (anticipated as usual), you would have to add a shocks block after the
endval block to “undo” the ﬁrst several periods of the permanent shock. For
instance, suppose you wanted the value of technology to move to 0.1, but only
in period 10. Then you would follow the above endval block with:
shocks;
var z;
3.8. SELECTING A COMPUTATION 27
periods 1:9;
values 0;
end;
3.7.3 Stochastic models
Recall from our earlier description of stochastic models that shocks are only
allowed to be temporary. A permanent shock cannot be accommodated due to
the need to stationarize the model around a steady state. Furthermore, shocks
can only hit the system today, as the expectation of future shocks must be
zero. With that in mind, we can however make the eﬀect of the shock propa
gate slowly throughout the economy by introducing a “latent shock variable”
such as e
t
in our example, that aﬀects the model’s true exogenous variable, z
t
in our example, which is itself an AR(1), exactly as in the model we introduced
from the outset. In that case, though, we would declare z
t
as an endogenous
variable and e
t
as an exogenous variable, as we did in the preamble of the
.mod ﬁle in section 3.4. Supposing we wanted to add a shock with variance
σ
2
, where σ is determined in the preamble block, we would write:
shocks;
var e = sigma^2;
end;
TIP! You can actually mix in deterministic shocks in stochastic models
by using the commands varexo det and listing some shocks as lasting more
than one period in the shocks block. For information on how to do so, please
see the Reference Manual. This can be particularly useful if you’re studying
the eﬀects of anticipated shocks in a stochastic model. For instance, you may
be interested in what happens to your monetary model if agents began ex
pecting higher inﬂation, or a depreciation of your currency.
3.8 Selecting a computation
So far, we have written an instructive .mod ﬁle, but what should Dynare do
with it? What are we interested in? In most cases, it will be impulse re
sponse functions (IRFs) due to the external shocks. Let’s see which are the
appropriate commands to give to Dynare. Again, we will distinguish between
deterministic and stochastic models.
28 CHAPTER 3. SOLVING DSGE MODELS  BASICS
3.8.1 For deterministic models
In the deterministic case, all you need to do is add the command simul
at the bottom of your .mod ﬁle. Note that the command takes the option
[ (periods=INTEGER) ] The command simul triggers the computation a
numerical simulation of the trajectory of the model’s solution for the number
of periods set in the option. To do so, it uses a Newton method to solve
simultaneously all the equations for every period (see Juillard (1996) for de
tails). Note that unless you use the endval command, the algorithm makes
the simplifying assumption that the system is back to equilibrium after the
speciﬁed number of periods. Thus, you must specify a large enough number
of periods such that increasing it further doesn’t change the simulation for
all practical purpose. In the case of a temporary shock, for instance, the tra
jectory will basicaly describe how the system gets back to equilibrium after
being perturbed from the shocks you entered.
3.8.2 For stochastic models
In the more common case of stochastic models, the command stoch simul is
appropriate. This command instructs Dynare to compute a Taylor approxi
mation of the decision and transition functions for the model (the equations
listing current values of the endogenous variables of the model as a func
tion of the previous state of the model and current shocks), impulse response
functions and various descriptive statistics (moments, variance decomposition,
correlation and autocorrelation coeﬃcients).
1
Impulse response functions are the expected future path of the endogenous
variables conditional on a shock in period 1 of one standard deviation.TIP!
If you linearize your model up to a ﬁrst order, impulse response functions
are simply the algebraic forward iteration of your model’s policy or decision
rule. If you instead linearize to a second order, impulse response functions
will be the result of actual Monte Carlo simulations of future shocks. This is
because in second order linear equations, you will have cross terms involving
the shocks, so that the eﬀects of the shocks depend on the state of the system
when the shocks hit. Thus, it is impossible to get algebraic average values
of all future shocks and their impact. The technique is instead to pull fu
ture shocks from their distribution and see how they impact your system, and
repeat this procedure a multitude of times in order to draw out an average
response. That said, note that future shocks will not have a signiﬁcant impact
1
For correlated shocks, the variance decomposition is computed as in the VAR literature
through a Cholesky decomposition of the covariance matrix of the exogenous variables.
When the shocks are correlated, the variance decomposition depends upon the order of the
variables in the varexo command.
3.8. SELECTING A COMPUTATION 29
on your results, since they get averaged between each Monte Carlo trial and
in the limit should sum to zero, given their mean of zero. Note that in the
case of a second order approximation, Dynare will return the actual sample
moments from the simulations. For ﬁrst order linearizations, Dynare will in
stead report theoretical moments. In both cases, the return to steady state
is asymptotic, TIP! thus you should make sure to specify suﬃcient periods
in your IRFs such that you actually see your graphs return to steady state.
Details on implementing this appear below.
If you’re interested to peer a little further into what exactly is going on
behind the scenes of Dynare’s computations, have a look at Chapter 7. Here
instead, we focus on the application of the command and reproduce below the
most common options that can be added to stoch simul. For a complete list
of options, please see the Reference Manual.
Options following the stoch simul command:
• ar = INTEGER: Order of autocorrelation coeﬃcients to compute and
to print (default = 5).
• dr algo = 0 or 1: speciﬁes the algorithm used for computing the quadratic
approximation of the decision rules: 0 uses a pure perturbation approach
as in SchmittGrohe and Uribe (2004) (default) and 1 moves the point
around which the Taylor expansion is computed toward the means of
the distribution as in Collard and Juillard (2001b).
• drop = INTEGER: number of points dropped at the beginning of sim
ulation before computing the summary statistics (default = 100).
• hp ﬁlter = INTEGER: uses HP ﬁlter with lambda = INTEGER before
computing moments (default: no ﬁlter).
• hp ngrid = INTEGER: number of points in the grid for the discreet In
verse Fast Fourier Transform used in the HP ﬁlter computation. It may
be necessary to increase it for highly autocorrelated processes (default
= 512).
• irf = INTEGER: number of periods on which to compute the IRFs
(default = 40). Setting IRF=0, suppresses the plotting of IRF’s.
• relative irf requests the computation of normalized IRFs in percentage
of the standard error of each shock.
• nocorr: doesn’t print the correlation matrix (printing is the default).
• nofunctions: doesn’t print the coeﬃcients of the approximated solution
(printing is the default).
30 CHAPTER 3. SOLVING DSGE MODELS  BASICS
• nomoments: doesn’t print moments of the endogenous variables (print
ing them is the default).
• noprint: cancel any printing; usefull for loops.
• order = 1 or 2 : order of Taylor approximation (default = 2), unless
you’re working with a linear model in which case the order is automati
cally set to 1.
• periods = INTEGER: speciﬁes the number of periods to use in simu
lations (default = 0). TIP! A simulation is similar to running impulse
response functions with a model linearized to the second order, in the
way that both sample shocks from their distribution to see how the
system reacts, but a simulation only repeats the process once, whereas
impulse response functions run a multitude of Monte Carlo trials in order
to get an average response of your system.
• qz criterium = INTEGER or DOUBLE: value used to split stable from
unstable eigenvalues in reordering the Generalized Schur decomposition
used for solving 1st order problems (default = 1.000001).
• replic = INTEGER: number of simulated series used to compute the
IRFs (default = 1 if order = 1, and 50 otherwise).
• simul seed = INTEGER or DOUBLE or (EXPRESSION): speciﬁes a
seed for the random number generator so as to obtain the same random
sample at each run of the program. Otherwise a diﬀerent sample is
used for each run (default: seed not speciﬁed). If you linearized to a
second order, Dynare will actually undertake Monte Carlo simulations
to generate moments of your variables. Because of the simulation, results
are bound to be slightly diﬀerent each time you run your program, except
if you ﬁx the seed for the random number generator. TIP! If you do
decide to ﬁx the seed, you should at least try to run your program
without using simul seed, just to check the robustness of your results.
Going back to our good old example, suppose we were interested in print
ing all the various measures of moments of our variables, want to see impulse
response functions for all variables, are basically happy with all default op
tions and want to carry out simulations over a good number of periods. We
would then end our .mod ﬁle with the following command:
stoch simul(periods=2100);
3.9. THE COMPLETE .MOD FILE 31
3.9 The complete .mod ﬁle
For completion’s sake, and for the pleasure of seeing our work bear its fruits,
here are the complete .mod ﬁles corresponding to our example for the de
terministic and stochastic case. You can ﬁnd the corresponding ﬁles in the
models folder under UserGuide in your installation of Dynare. The ﬁles are
called RBC Monop JFV.mod for stochastic models and RBC Monop Det.mod for
deterministic models.
3.9.1 The stochastic model
var y c k i l y l w r z;
varexo e;
parameters beta psi delta alpha rho gamma sigma epsilon;
alpha = 0.33;
beta = 0.99;
delta = 0.023;
psi = 1.75;
rho = 0.95;
sigma = (0.007/(1alpha));
epsilon = 10;
model;
(1/c) = beta*(1/c(+1))*(1+r(+1)delta);
psi*c/(1l) = w;
c+i = y;
y = (k(1)^alpha)*(exp(z)*l)^(1alpha);
w = y*((epsilon1)/epsilon)*(1alpha)/l;
r = y*((epsilon1)/epsilon)*alpha/k(1);
i = k(1delta)*k(1);
y l = y/l;
z = rho*z(1)+e;
end;
initval;
k = 9;
c = 0.76;
l = 0.3;
w = 2.07;
r = 0.03;
z = 0;
e = 0;
end;
32 CHAPTER 3. SOLVING DSGE MODELS  BASICS
steady;
check;
shocks;
var e = sigma^2;
end;
stoch simul(periods=2100);
3.9.2 The deterministic model (case of temporary shock)
var y c k i l y l w r ;
varexo z;
parameters beta psi delta alpha sigma epsilon;
alpha = 0.33;
beta = 0.99;
delta = 0.023;
psi = 1.75;
sigma = (0.007/(1alpha));
epsilon = 10;
model;
(1/c) = beta*(1/c(+1))*(1+r(+1)delta);
psi*c/(1l) = w;
c+i = y;
y = (k(1)^alpha)*(exp(z)*l)^(1alpha);
w = y*((epsilon1)/epsilon)*(1alpha)/l;
r = y*((epsilon1)/epsilon)*alpha/k(1);
i = k(1delta)*k(1);
y l = y/l;
end;
initval;
k = 9;
c = 0.7;
l = 0.3;
w = 2.0;
r = 0;
z = 0;
end;
steady;
3.10. FILE EXECUTION AND RESULTS 33
check;
shocks;
var z;
periods 1:9;
values 0.1;
end;
simul(periods=2100);
3.10 File execution and results
To see this all come to life, let’s run our .mod ﬁle, which is conveniently
installed by default in the Dynare “examples” directory (the .mod ﬁle cor
responding to the stochastic model is called RBC Monop JFV.mod and that
corresponding to the deterministic model is called RBC Monop Det.mod). (**
note, this may not be the case when testing the beta version of Matlab version
4)
To run a .mod ﬁle, navigate within Matlab to the directory where the
example .mod ﬁles are stored. You can do this by clicking in the “current di
rectory” window of Matlab, or typing the path directly in the top white ﬁeld
of Matlab. Once there, all you need to do is place your cursor in the Matlab
command window and type, for instance, dynare ExSolStoch; to execute
your .mod ﬁle.
Running these .mod ﬁles should take at most 30 seconds. As a result, you
should get two forms of output  tabular in the Matlab command window and
graphical in one or more popup windows. Let’s review these results.
3.10.1 Results  stochastic models
The tabular results can be summarized as follows:
1. Model summary: a count of the various variable types in your model
(endogenous, jumpers, etc...).
2. Eigenvalues should be displayed, and you should see a conﬁrmation of
the BlanchardKahn conditions if you used the command check in your
.mod ﬁle.
34 CHAPTER 3. SOLVING DSGE MODELS  BASICS
3. Matrix of covariance of exogenous shocks: this should square with
the values of the shock variances and covariances you provided in the
.mod ﬁle.
4. Policy and transition functions: Solving the rational exectation
model, E
t
[f(y
t+1
, y
t
, y
t−1
, u
t
)] = 0 , means ﬁnding an unkown function,
y
t
= g(y
t−1
, u
t
) that could be plugged into the original model and satisfy
the implied restrictions (the ﬁrst order conditions). A ﬁrst order approx
imation of this function can be written as y
t
= ¯ y + g
y
ˆ y
t−1
+ g
u
u
t
, with
ˆ y
t
= y
t
− ¯ y and ¯ y being the steadystate value of y, and where g
x
is the
partial derivative of the g function with respect to variable x. In other
words, the function g is a time recursive (approximated) representation
of the model that can generate timeseries that will approximatively sat
isfy the rational expectation hypothesis contained in the original model.
In Dynare, the table “Policy and Transition function” contains the el
ements of g
y
and g
u
. Details on the policy and transition function can
be found in Chapter 6.
5. Moments of simulated variables: up to the fourth moments.
6. Correlation of simulated variables: these are the contemporaneous
correlations, presented in a table.
7. Autocorrelation of simulated variables: up to the ﬁfth lag, as spec
iﬁed in the options of stoch simul.
The graphical results, instead, show the actual impulse response func
tions for each of the endogenous variables, given that they actually moved.
These can be especially useful in visualizing the shape of the transition func
tions and the extent to which each variable is aﬀected. TIP! If some variables
do not return to their steady state, either check that you have included enough
periods in your simulations, or make sure that your model is stationary, i.e.
that your steady state actually exists and is stable. If not, you should detrend
your variables and rewrite your model in terms of those variables.
3.10.2 Results  deterministic models
Automatically displayed results are much more scarce in the case of deter
ministic models. If you entered steady, you will get a list of your steady
state results. If you entered check, eigenvalues will also be displayed and you
should receive a statement that the rank condition has been satisﬁed, if all
goes well! Finally, you will see some intermediate output: the errors at each
iteration of the Newton solver used to estimate the solution to your model.
TIP! You should see these errors decrease upon each iteration; if not, your
model will probably not converge. If so, you may want to try to increase the
periods for the transition to the new steady state (the number of simulations
3.10. FILE EXECUTION AND RESULTS 35
periods). But more often, it may be a good idea to revise your equations. Of
course, although Dynare does not display a rich set of statistics and graphs
corresponding to the simulated output, it does not mean that you cannot cre
ate these by hand from Matlab. To do so, you should start by looking at
section 4.1.3 of chapter 4 on ﬁnding, saving and viewing your output.
Chapter 4
Solving DSGE models 
advanced topics
This chapter is a collection of topics  not all related to each other  that you
will probably ﬁnd interesting or at least understandable, if you have read,
and/ or feel comfortable with, the earlier chapter 3 on the basics of solving
DSGE models. To provide at least some consistency, this chapter is divided
into three sections. The ﬁrst section deals directly with features of Dynare,
such as dealing with correlated shocks, ﬁnding and saving your output, using
loops, referring to external ﬁles and dealing with inﬁnite eigenvalues. The
second section overviews some of the inner workings of Dynare. The goal
is to provide a brief explanation of the ﬁles that are created by Dynare to
help you in troubleshooting or provide a starting point in case you actually
want to customize the way Dynare works. The third section of the chapter
focusses on modeling tips optimized for Dynare, but possibly also helpful for
other work.
4.1 Dynare features and functionality
4.1.1 Other examples
Other examples of .mod ﬁles used to generate impulse response functions are
available on the Dynare website. In particular, Jesus FernandezVillaverde
has provided a series of RBC model variants (from the most basic to some
including variable capacity utilization, indivisible labor and investment spe
ciﬁc technological change). You can ﬁnd these, along with helpful notes and
explanations, in the Oﬃcial Examples section of the Dynare website.
Also, don’t forget to check occasionally the Dynare contributions and ex
amples forum to see if any other user has posted an example that could help
37
38 CHAPTER 4. SOLVING DSGE MODELS  ADVANCED TOPICS
you in your work; or maybe you would like to post an example there yourself?
4.1.2 Alternative, complete example
The following example aims to give you an alternative example to the one in
chapter 3, to learn the workings of Dynare. It also aims to give you exposure
to dealing with several correlated shocks. Your model may have two or
more shocks, and these may be correlated to each other. The example below
illustrates how you would introduce this into Dynare. Actually, the example
provided is somewhat more complete than strictly necessary. This is to give
you an alternative, fullblown example to the one described in chapter 3.
The model
The model is a simpliﬁed standard RBC model taken from Collard and Juil
lard (2003) which served as the original User Guide for Dynare.
The economy consists of an inﬁnitely living representative agent who values
consumption c
t
and labor services h
t
according to the following utility function
E
t
∞
τ=t
β
τ−t
_
log(c
t
) −θ
h
1+ψ
t
1 +ψ
_
where, as usual, the discount factor 0 < β < 1, the disutility of labor θ > 0
and the labor supply elasticity ψ ≥ 0.
A social planner maximizes this utility function subject to the resource
constraint
c
t
+i
t
= y
t
where i
t
is investment and y
t
output. Consumers are therefore also owners
of the ﬁrms. The economy is a real economy, where part of output can be
consumed and part invested to form physical capital. As is standard, the law
of motion of capital is given by
k
t+1
= exp(b
t
)i
t
+ (1 −δ)k
t
with 0 < δ < 1, where δ is physical depreciation and b
t
a shock aﬀecting
incorporated technological progress.
We assume output is produced according to a standard constant returns
to scale technology of the form
y
t
= exp(a
t
)k
α
t
h
1−α
t
with α being the capital elasticity in the production function, with 0 < α < 1,
and where a
t
represents a stochastic technological shock (or Solow residual).
4.1. DYNARE FEATURES AND FUNCTIONALITY 39
Finally, we specify a shock structure that allows for shocks to display
persistence across time and correlation in the current period. That is
_
a
t
b
t
_
=
_
ρ τ
τ ρ
__
a
t−1
b
t−1
_
+
_
t
ν
t
_
where ρ + τ < 1 and ρ −τ < 1 to ensure stationarity (we call ρ the coeﬃ
cient of persistence and τ that of crosspersistence). Furthermore, we assume
E
t
(
t
) = 0, E
t
(ν
t
) = 0 and that the contemporaneous variancecovariance
matrix of the innovations
t
and ν
t
is given by
_
σ
2
ψσ
σ
ν
ψσ
σ
ν
σ
2
ν
_
and where corr(
t
ν
s
) = 0, corr(
t
s
) = 0 and corr(ν
t
ν
s
) = 0 for all t = s.
This system  probably quite similar to standard RBC models you have run
into  yields the following ﬁrst order conditions (which are straightforward to
reproduce in case you have doubts. . . ) and equilibrium conditions drawn from
the description above. Note that the ﬁrst equation captures the labor supply
function and the second the intertemporal consumption Euler equation.
c
t
θh
1+ψ
t
= (1 −α)y
t
1 = βE
t
__
exp(b
t
)c
t
exp(b
t+1
)c
t+1
__
exp(b
t+1
)α
y
t+1
k
t+1
+ 1 −δ
__
y
t
= exp(a
t
)k
α
t
h
1−α
t
k
t+1
= exp(b
t
)i
t
+ (1 −δ)k
t
a
t
= ρa
t−1
+τb
t−1
+
t
b
t
= τa
t−1
+ρb
t−1
+ν
t
The .mod ﬁle
To “translate” the model into a language understandable by Dynare, we would
follow the steps outlined in chapter 3. We will assume that you’re comfort
able with these and simply present the ﬁnal .mod ﬁle below. Fist, though,
note that to introduce shocks into Dynare, we have two options (this was not
discussed in the earlier chapter). Either write:
shocks;
var e; stderr 0.009;
var u; stderr 0.009;
var e, u = phi*0.009*0.009;
end;
40 CHAPTER 4. SOLVING DSGE MODELS  ADVANCED TOPICS
where the last line speciﬁes the contemporaneous correlation between our two
exogenous variables.
Alternatively, you can also write:
shocks;
var e = 0.009^2;
var u = 0.009^2;
var e, u = phi*0.009*0.009;
end;
So that you can gain experience by manipulating the entire model, here is
the complete .mod ﬁle corresponding to the above example. You can ﬁnd the
corresponding ﬁle in the models folder under UserGuide in your installation
of Dynare. The ﬁle is called Alt Ex1.mod.
var y, c, k, a, h, b;
varexo e, u;
parameters beta, rho, alpha, delta, theta, psi, tau;
alpha = 0.36;
rho = 0.95;
tau = 0.025;
beta = 0.99;
delta = 0.025;
psi = 0;
theta = 2.95;
phi = 0.1;
model;
c*theta*h^(1+psi)=(1alpha)*y;
k = beta*(((exp(b)*c)/(exp(b(+1))*c(+1)))
*(exp(b(+1))*alpha*y(+1)+(1delta)*k));
y = exp(a)*(k(1)^alpha)*(h^(1alpha));
k = exp(b)*(yc)+(1delta)*k(1);
a = rho*a(1)+tau*b(1) + e;
b = tau*a(1)+rho*b(1) + u;
end;
initval;
y = 1.08068253095672;
c = 0.80359242014163;
4.1. DYNARE FEATURES AND FUNCTIONALITY 41
h = 0.29175631001732;
k = 5;
a = 0;
b = 0;
e = 0;
u = 0;
end;
shocks;
var e; stderr 0.009;
var u; stderr 0.009;
var e, u = phi*0.009*0.009;
end;
stoch simul(periods=2100);
4.1.3 Finding, saving and viewing your output
Where is output stored? Most of the moments of interest are stored in global
variable oo You can easily browse this global variable in Matlab by either
calling it in the command line, or using the workspace interface. In global
variable oo you will ﬁnd the following (NOTE! variables will always appear
in the order in which you declared them in the preamble block of your .mod
ﬁle):
• steady state: the steady state of your variables
• mean: the mean of your variables
• var: the variance of your variables
• autocorr: the various autocorrelation matrices of your variables. Each
row of these matrices will correspond to a variables in time t, and
columns correspond to the variables lagged 1, for the ﬁrst matrix, then
lagged 2 for the second matrix, and so on. Thus, the matrix of auto
correlations that is automatically displayed in the results after running
stoch simul has, running down each column, the diagonal elements of
each of the various autocorrelation matrices described here.
• gamma y: the matrices of autocovariances. gamma y{1} represents vari
ances, while gamma y{2} represents autocovariances where variables on
each column are lagged by one period and so on. By default, Dynare will
return autocovariances with a lag of 5. The last matrix (gamma y{7} in
42 CHAPTER 4. SOLVING DSGE MODELS  ADVANCED TOPICS
the default case) returns the variance decomposition, where each col
umn captures the independent contribution of each shock to the variance
of each variable.
Furthermore, if you decide to run impulse response functions, you will
ﬁnd a global variable oo .irfs comprising of vectors named endogenous
variable exogenous variable, like y e, reporting the values of the endoge
nous variables corresponding to the impulse response functions, as a result of
the independent impulse of each exogenous shock.
To save your simulated variables, you can add the following command at
the end of your .mod ﬁle: dynasave (FILENAME) [variable names separated
by commas] If no variable names are speciﬁed in the optional ﬁeld, Dynare will
save all endogenous variables. In Matlab, variables saved with the dynasave
command can be retrieved by using the Matlab command load mat FILENAME.
4.1.4 Referring to external ﬁles
You may ﬁnd it convenient to refer to an external ﬁle, either to compute the
steady state of your model, or when specifying shocks in an external ﬁle. The
former is described in section 3.6 of chapter 3 when discussing steady states.
The advantage of using Matlab, say, to ﬁnd your model’s steady state was
clear with respect to Dynare version 3, as the latter resorted to numerical
approximations to ﬁnd steady state values. But Dynare version 4 now uses
the same analytical methods available in Matlab. For most usage scenarios,
you should therefore do just as well to ask Dynare to compute your model’s
steady state (except, maybe, if you want to run loops, to vary your parameter
values, for instance, in which case writing a Matlab program may be more
handy).
But you may also be interested in the second possibility described above,
namely of specifying shocks in an external ﬁle, to simulate a model based on
shocks from a prior estimation, for instance. You could then retrieve the ex
ogenous shocks from the oo ﬁle by saving them in a ﬁle called dataﬁle.mat.
Finally, you could simulate a deterministic model with the shocks saved from
the estimation by specifying the source ﬁle for the shocks, using the
shocks(shocks file = datafile.mat) command. But of course, this is a bit
of a workaround, since you could also use the builtin commands in Dynare
to generate impulse response functions from estimated shocks, as described in
chapter 5.
4.2. FILES CREATED BY DYNARE 43
4.1.5 Inﬁnite eigenvalues
If you use the command check in your .mod ﬁle, Dynare will report your sys
tem’s eigenvalues and tell you if these meet the BlanchardKahn conditions.
At that point, don’t worry if you get inﬁnite eigenvalues  these are are ﬁrmly
grounded in the theory of generalized eigenvalues. They have no detrimental
inﬂuence on the solution algorithm. As far as BlanchardKahn conditions are
concerned inﬁnite eigenvalues are counted as explosive roots of modulus larger
than one.
4.2 Files created by Dynare
At times, you may get a message that there is an error in a ﬁle with a new
name, or you may want to have a closer look at how Dynare actually solves
your model  out of curiosity or maybe to do some customization of your own.
You may therefore ﬁnd it helpful to get a brief overview of the internal ﬁles
that Dynare generates and the function of each one.
The dynare preprocessors essentially does three successive tasks: (i) pars
ing of the mod ﬁle (it checks that the mod ﬁle is syntactically correct), and its
translation into internal machine representation (in particular, model equa
tions are translated into expression trees), (ii) symbolic derivation of the model
equations, up to the needed order (depending on the computing needs), (iii)
outputting of several ﬁles, which are used from matlab. If the mod ﬁle is
“ﬁlename.mod”, then the preprocessor creates the following ﬁles:
• ﬁlename.m: a matlab ﬁle containing several instructions, notably the
parameter initializations and the matlab calls corresponding to comput
ing tasks
• ﬁlename dynamic.m: a matlab ﬁle containing the model equations
and their derivatives (ﬁrst, second and maybe third). Endogenous vari
ables (resp. exogenous variables, parameters) are contained in a “y”
(resp. “x”, “params”) vector, with an index number depending on the
declaration order. The “y” vector has as many entries as their are (vari
able, lag) pairs in the declared model. The model equations residuals
are stored in a vector named “residuals”. The model jacobian is put in
“g1” matrix. Second (resp. third) derivatives are in “g2” matrix (resp.
“g3”). If the “use dll” option has been speciﬁed in the model decla
ration, the preprocessor will output a C ﬁle (with .c extension) rather
than a matlab ﬁle. It is then compiled to create a library (DLL) ﬁle. Us
ing a compiled C ﬁle is supposed to give better computing performance
in model simulation/estimation.
44 CHAPTER 4. SOLVING DSGE MODELS  ADVANCED TOPICS
• ﬁlename static.m: a matlab ﬁle containing the stationarized version of
the model (i.e. where lagged variables are replaced by current variables),
with its jacobian. Used to compute the steady state. Same notations
than the dynamic ﬁle. Replaced by a C ﬁle when “use dll” option is
speciﬁed.
4.3 Modeling tips
4.3.1 Stationarizing your model
Models in Dynare must be stationary, such that you can linearize them around
a steady state and return to steady state after a shock. Thus, you must ﬁrst
stationarize your model, then linearize it, either by hand, or by letting Dynare
do the work. You can then reconstruct expost the nonstationary simulated
variables after running impulse response functions.
For deterministic models, the trick is to use only stationary variables in
t + 1. More generally, if y
t
is I(1), you can always write y
t+1
as y
t
+ dy
t+1
,
where dy
t
= y
t
− y
t−1
. Of course, you need to know the value of dy
t
at the
ﬁnal equilibrium.
Note that in a stationary model, it is expected that variables will eventually
go back to steady state after the initial shock. If you expect to see a growing
curve for a variable, you are thinking about a growth model. Because growth
models are nonstationary, it is easier to work with the stationarized version of
such models. Again, if you know the trend, you can always add it back after
the simulation of the stationary components of the variables.
4.3.2 Expectations taken in the past
For instance, to enter the term E
t−1
y
t
, deﬁne s
t
= E
t
[y
t+1
] and then use s(−1)
in your .mod ﬁle. Note that, because of Jensen’s inequality, you cannot do
this for terms that enter your equation in a nonlinear fashion. If you do have
nonlinear terms on which you want to take expectations in the past, you
would need to apply the above manipulation to the entire equation, as if y
t
were an equation, not just a variable.
4.3.3 Inﬁnite sums
Dealing with inﬁnite sums is tricky in general, and needs particular care when
working with Dynare. The trick is to use a recursive representation of the
sum. For example, suppose your model included:
∞
j=0
β
j
x
t+j
= 0,
4.3. MODELING TIPS 45
Note that the above can also be written by using an auxiliary variable S
t
,
deﬁned as:
S
t
≡
∞
j=0
β
j
x
t+j
,
which can also be written in the following recursive manner:
S
t
≡
∞
j=0
β
j
x
t+j
= x
t
+
∞
j=1
β
j
x
t+j
= x
t
+β
∞
j=0
β
j
x
t+1+j
≡ x
t
+S
t+1
This formulation turns out to be useful in problems of the following form:
∞
j=0
β
j
x
t+j
= p
t
∞
j=0
γ
j
y
t+j
,
which can be written as a recursive system of the form:
S1
t
= x
t
+βS1
t+1
,
S2
t
= y
t
+γS2
t+1
,
S1 = p
t
S2.
This is particularly helpful, for instance, in a Calvo type setting, as
illustrated in the following brief example. The RBC model with monopolistic
competition introduced in chapter 3 involved ﬂexible prices. The extension
with sticky prices, `a la Calvo for instance, is instead typical of the new Key
nesian monetary literature, exempliﬁed by papers such as Clarida, Gali, and
Gertler (1999).
The optimal price for a ﬁrm resetting its price in period t, given that it
will be able to reset its price only with probability 1 −θ each period, is
p
∗
t
(i) = µ + (1 −βθ)
∞
k=0
(βθ)
k
E
t
[mc
n
t+k
(i)]
where µ is the markup, β is a discount factor, i represents a ﬁrm of the contin
uum between 0 and 1, and mc
t
is marginal cost as described in the example in
chapter 3. The trouble, of course, is how to input this inﬁnite sum into
Dynare?
It turns out that the Calvo price setting implies that the aggregate price
follows the equation of motion p
t
= θp
t−1
+ (1 − θ)p
∗
t
, thus implying the
following inﬂation relationship π
t
= (1 − θ)(p
∗
t
− p
t−1
). Finally, we can also
rewrite the optimal price setting equation, after some algebraic manipulations,
as
p
∗
t
−p
t−1
= (1 −βθ)
∞
k=0
(βθ)
k
E
t
[ ´ mc
t+k
] +
∞
k=0
(βθ)
k
E
t
[π
t+k
]
46 CHAPTER 4. SOLVING DSGE MODELS  ADVANCED TOPICS
where ´ mc
t+k
= mc
t+k
+µ is the deviation of the marginal cost from its natural
rate, deﬁned as the marginal cost when prices are perfectly ﬂexible.
The trick now is to note that the above can be written recursively, by
writing the right hand side as the ﬁrst term of the sum (with k = 0) plus the
remainder of the sum, which can be written as the left hand side term scrolled
forward one period and appropriately discounted. Mathematically, this yields:
p
∗
t
−p
t−1
= (1 −βθ) ´ mc
t+k
+π
t
+β θE
t
[p
∗
t+1
−p
t
]
which has gotten rid of our inﬁnite sum! That would be enough for Dynare,
but for convenience, we can go one step further and write the above as
π
t
= βE
t
[π
t+1
] +λ´ mc
t
where λ ≡
(1−θ)(1−βθ)
θ
, which is the recognizable inﬂation equation in the new
Keynesian (or new Neoclassical) monetary literature.
4.3.4 Inﬁnite sums with changing timing of expectations
When you are not able to write an inﬁnite sum recursively, as the index of
the expectations changes with each element of the sum, as in the following
example, a diﬀerent approach than the one mentioned above is necessary.
Suppose your model included the following sum:
y
t
=
∞
j=0
E
t−j
x
t
where y
t
and x
t
are endogenous variables.
In Dynare, the best way to handle this is to write out the ﬁrst k terms
explicitly and enter each one in Dynare, such as: E
t−1
x
t
+E
t−2
x
t
+. . .+E
t−k
x
t
.
Chapter 5
Estimating DSGE models 
basics
As in the chapter 3, this chapter is structured around an example. The goal
of this chapter is to lead you through the basic functionality in Dynare to
estimate models using Bayesian techniques, so that by the end of the chapter
you should have the capacity to estimate a model of your own. This chapter
is therefore very practicallyoriented and abstracts from the underlying com
putations that Dynare undertakes to estimate a model; that subject is instead
covered in some depth in chapter 8, while more advanced topics of practical
appeal are discussed in chapter 6.
5.1 Introducing an example
The example introduced in this chapter is particularly basic. This is for two
reasons. First, we did not want to introduce yet another example in this sec
tion; there’s enough new material to keep you busy. Instead, we thought it
would be easiest to simply continue working with the example introduced in
chapter 3 with which you are probably already quite familiar. Second, the
goal of the example in this chapter is really to explain features in context,
but not necessarily to duplicate a “real life scenario” you would face when
writing a paper. Once you feel comfortable with the content of this chapter,
though, you can always move on to chapter 6 where you will ﬁnd a fullﬂedged
replication of a recent academic paper, featuring a nonstationary model.
Recall from chapter 3 that we are dealing with an RBC model with mo
nopolistic competition. Suppose we had data on business cycle variations of
output. Suppose also that we thought our little RBC model did a good job of
reproducing reality. We could then use Bayesian methods to estimate the pa
rameters of the model: α, the capital share of output, β, the discount factor,
47
48 CHAPTER 5. ESTIMATING DSGE MODELS  BASICS
δ, the depreciation rate, ψ, the weight on leisure in the utility function, ρ, the
degree of persistence in productivity, and , the markup parameter. Note that
in Bayesian estimation, the condition for undertaking estimation is that there
be at least as many shocks as there are observables (a less stringent condition
than for maximum likelihood estimation). It may be that this does not allow
you to identify all your parameters  yielding posterior distributions identical
to prior distributions  but the Bayesian estimation procedure would still run
successfully. Let’s see how to go about doing this.
5.2 Declaring variables and parameters
To input the above model into Dynare for estimation purposes, we must ﬁrst
declare the model’s variables in the preamble of the .mod ﬁle. This is done
exactly as described in chapter 3 on solving DSGE models. We thus begin the
.mod ﬁle with:
var y c k i l y l w r z;
varexo e;
parameters beta psi delta alpha rho epsilon;
5.3 Declaring the model
Suppose that the equation of motion of technology is a stationary AR(1) with
an autoregressive parameter, ρ, less than one. The model’s variables would
therefore be stationary and we can proceed without complications. The al
ternative scenario with nonstationary variables is more complicated and
dealt with in chapter 6 (in the additional example). In the stationary case,
our model block would look exactly as in chater 3:
model;
(1/c) = beta*(1/c(+1))*(1+r(+1)delta);
psi*c/(1l) = w;
c+i = y;
y = (k(1)^alpha)*(exp(z)*l)^(1alpha);
w = y*((epsilon1)/epsilon)*(1alpha)/l;
r = y*((epsilon1)/epsilon)*alpha/k(1);
i = k(1delta)*k(1);
y l = y/l;
z = rho*z(1)+e;
5.4. DECLARING OBSERVABLE VARIABLES 49
end;
5.4 Declaring observable variables
This should not come as a surprise. Dynare must know which variables are
observable for the estimation procedure. NOTE! These variables must be
available in the data ﬁle, as explained in section 5.7 below. For the moment,
we write:
varobs y;
5.5 Specifying the steady state
Before Dynare estimates a model, it ﬁrst linearizes it around a steady state.
Thus, a steady state must exist for the model and although Dynare can calcu
late it, we must give it a hand by declaring approximate values for the steady
state. This is just as explained in details and according to the same syntax
outlined in chapter 3, covering the initval, steady and check commands.
In fact, as this chapter uses the same model as that outlined in chapter 3, the
steady state block will look exactly the same.
TIP! During estimation, in ﬁnding the posterior mode, Dynare recalcu
lates the steady state of the model at each iteration of the optimization rou
tine (more on this later), based on the newest round of parameters available.
Thus, by providing approximate initial values and relying solely on the built
in Dynare algorithm to ﬁnd the steady state (a numerical procedure), you will
signiﬁcantly slow down the computation of the posterior mode. Dynare will
end up spending 60 to 70% of the time recalculating steady states. It is much
more eﬃcient to write an external Matlab steady state ﬁle and let Dynare
use that ﬁle to ﬁnd the steady state of your model by algebraic procedure. For
more details on writing an external Matlab ﬁle to ﬁnd your model’s steady
state, please refer to section 3.6.3 of chapter 3.
5.6 Declaring priors
Priors play an important role in Bayesian estimation and consequently de
serve a central role in the speciﬁcation of the .mod ﬁle. Priors, in Bayesian
estimation, are declared as a distribution. The general syntax to introduce
priors in Dynare is the following:
50 CHAPTER 5. ESTIMATING DSGE MODELS  BASICS
estimated params;
PARAMETER NAME, PRIOR SHAPE, PRIOR MEAN, PRIOR STANDARD ERROR [, PRIOR
3
rd
PARAMETER] [,PRIOR 4
th
PARAMETER] ;
end;
where the following table deﬁnes each term more clearly
PRIOR SHAPE Corresponding distribution Range
NORMAL PDF N(µ, σ) R
GAMMA PDF G
2
(µ, σ, p
3
) [p
3
, +∞)
BETA PDF B(µ, σ, p
3
, p
4
) [p
3
, p
4
]
INV GAMMA PDF IG
1
(µ, σ) R
+
UNIFORM PDF U(p
3
, p
4
) [p
3
, p
4
]
where µ is the PRIOR MEAN, σ is the PRIOR STANDARD ERROR, p
3
is the PRIOR
3
rd
PARAMETER (whose default is 0) and p
4
is the PRIOR 4
th
PARAMETER (whose
default is 1). TIP! When specifying a uniform distribution between 0 and 1 as
a prior for a parameter, say α, you therefore have to put two empty spaces for
parameters µ and σ, and then specify parameters p3 and p4, since the uniform
distribution only takes p3 and p4 as arguments. For instance, you would write
alpha, uniform pdf, , , 0,1;
For a more complete review of all possible options for declaring priors, as
well as the syntax to declare priors for maximum likelihood estimation (not
Bayesian), see the Reference Manual. Note also that if some parameters in a
model are calibrated and are not to be estimated, you should declare them as
such, by using the parameters command and its related syntax, as explained
in chapter 3.
TIP! Choosing the appropriate prior for your parameters is a tricky, yet
very important endeavor. It is worth spending time on your choice of priors
and to test the robustness of your results to your priors. Some considerations
may prove helpful. First, think about the domain of your prior over each pa
rameter. Should it be bounded? Should it be opened on either or both sides?
Remember also that if you specify a probability of zero over a certain domain
in your prior, you will necessarily also ﬁnd a probability of zero in your pos
terior distribution. Then, think about the shape of your prior distribution.
Should it be symmetric? Skewed? If so, on which side? You may also go
one step further and build a distribution for each of your parameters in your
mind. Ask yourself, for instance, what is the probability that your parameter
is bigger than a certain value, and repeat the exercise by incrementally de
creasing that value. You can then pick the standard distribution that best ﬁts
5.6. DECLARING PRIORS 51
your perceived distribution. Finally, instead of describing here the shapes and
properties of each standard distribution available in Dynare, you are instead
encouraged to visualize these distributions yourself, either in a statistics book
or on the Web.
TIP! It may at times be desirable to estimate a transformation of a pa
rameter appearing in the model, rather than the parameter itself. In such a
case, it is possible to declare the parameter to be estimated in the parameters
statement and to deﬁne the transformation at the top of the model section,
as a Matlab expression, by adding a pound sign (#) at the beginning of the
corresponding line. For example, you may ﬁnd it easier to deﬁne a prior over
the discount factor, β, than its inverse which often shows up in Euler equa
tions. Thus you would write:
model;
# sig = 1/bet;
c = sig*c(+1)*mpk;
end;
estimated params;
bet,normal pdf,1,0.05;
end;
TIP! Finally, another useful command to use is the estimated params init
command which declares numerical initial values for the optimizer when these
are diﬀerent from the prior mean. This is especially useful when redoing an
estimation  if the optimizer got stuck the ﬁrst time around, or needing a
greater number of draws in the MetropolisHastings algorithm  and wanting
to enter the posterior mode as initial values for the parameters instead of a
prior. The Reference Manual gives more details as to the exact syntax of this
command.
Coming back to our basic example, we would write:
estimated params;
alpha, beta pdf, 0.35, 0.02;
beta, beta pdf, 0.99, 0.002;
delta, beta pdf, 0.025, 0.003;
psi, gamma pdf, 1.75, 0.02;
rho, beta pdf, 0.95, 0.05;
epsilon, gamma pdf, 10, 0.003;
stderr e, inv gamma pdf, 0.01, inf;
end;
52 CHAPTER 5. ESTIMATING DSGE MODELS  BASICS
5.7 Launching the estimation
To ask Dynare to estimate a model, all that is necessary is to add the com
mand estimation at the end of the .mod ﬁle. Easy enough. But the real
complexity comes from the options available for the command (to be entered
in parentheses and sequentially, separated by commas, after the command
estimation). Below, we list the most common and useful options, and en
courage you to view the Reference Manual for a complete list.
1. dataﬁle = FILENAME: the dataﬁle (a .m ﬁle, a .mat ﬁle, or an .xls
ﬁle). Note that observations do not need to show up in any order, but
vectors of observations need to be named with the same names as those
in var obs. In Excel ﬁles, for instance, observations could be ordered
in columns, and variable names would show up in the ﬁrst cell of each
column.
2. nobs = INTEGER: the number of observations to be used (default: all
observations in the ﬁle)
3. ﬁrst obs = INTEGER: the number of the ﬁrst observation to be used
(default = 1). This is useful when running loops, or instance, to divide
the observations into subperiods.
4. preﬁlter = 1: the estimation procedure demeans the data (default=0,
no preﬁltering). This is useful if model variables are in deviations from
steady state, for instance, and therefore have zero mean. Demeaning the
observations would also impose a zero mean on the observed variables.
5. nograph: no graphs should be plotted.
6. conf sig = {INTEGER — DOUBLE }: the level for the conﬁdence in
tervals reported in the results (default = 0.90)
7. mh replic = INTEGER: number of replication for Metropolis Hasting
algorithm. For the time being, mh replic should be larger than 1200
(default = 20000)
8. mh nblocks = INTEGER: number of parallel chains for Metropolis Hast
ing algorithm (default = 2). Despite this low default value, it is advisable
to work with a higher value, such as 5 or more. This improves the com
putation of between group variance of the parameter means, one of the
key criteria to evaluate the eﬃciency of the MetropolisHastings to eval
uate the posterior distribution. More details on this subject appear in
Chapter 6.
5.7. LAUNCHING THE ESTIMATION 53
9. mh drop = DOUBLE: the fraction of initially generated parameter vec
tors to be dropped before using posterior simulations (default = 0.5; this
means that the ﬁrst half of the draws from the MetropolisHastings are
discarded).
10. mh jscale = DOUBLE: the scale to be used for the jumping distribu
tion in MH algorithm. The default value is rarely satisfactory. This
option must be tuned to obtain, ideally, an acceptance rate of 25% in
the Metropolis Hastings algorithm (default = 0.2). The idea is not to
reject or accept too often a candidate parameter; the literature has set
tled on a value of between 0.2 and 0.4. If the acceptance rate were too
high, your MetropolisHastings iterations would never visit the tails of a
distribution, while if it were too low, the iterations would get stuck in a
subspace of the parameter range. Note that the acceptance rate drops if
you increase the scale used in the jumping distribution and vice a versa.
11. mh init scale=DOUBLE: the scale to be used for drawing the initial
value of the MetropolisHastings chain (default=2*mh jscale). The idea
here is to draw initial values from a stretched out distribution in order
to maximize the chances of these values not being too close together,
which would defeat the purpose of running several blocks of Metropolis
Hastings chains.
12. mode ﬁle=FILENAME: name of the ﬁle containing previous value for
the mode. When computing the mode, Dynare stores the mode (xparam1)
and the hessian (hh) in a ﬁle called MODEL NAME mode. This is a
particularly helpful option to speed up the estimation process if you have
already undertaken initial estimations and have values of the posterior
mode.
13. mode compute=INTEGER: speciﬁes the optimizer for the mode com
putation.
0: the mode isn’t computed. mode ﬁle must be speciﬁed
1: uses Matlab fmincon (see the Reference Manual to set options
for this command).
2: uses Lester Ingber’s Adaptive Simulated Annealing.
3: uses Matlab fminunc.
4 (default): uses Chris Sim’s csminwel.
14. mode check: when mode check is set, Dynare plots the minus of the
posterior density for values around the computed mode for each esti
mated parameter in turn. This is helpful to diagnose problems with the
optimizer. A clear indication of a problem would be that the mode is
not at the trough (bottom of the minus) of the posterior distribution.
54 CHAPTER 5. ESTIMATING DSGE MODELS  BASICS
15. load mh ﬁle: when load mh ﬁle is declared, Dynare adds to previous
MetropolisHastings simulations instead of starting from scratch. Again,
this is a useful option to speed up the process of estimation.
16. nodiagnostic: doesn’t compute the convergence diagnostics for Metropolis
Hastings (default: diagnostics are computed and displayed). Actually
seeing if the various blocks of MetropolisHastings runs converge is a
powerful and useful option to build conﬁdence in your model estima
tion. More details on these diagnostics are given in Chapter 6.
17. bayesian irf: triggers the computation of the posterior distribution of im
pulse response functions (IRFs). The length of the IRFs are controlled
by the irf option, as speciﬁed in chapter 3 when discussing the options for
stoch simul. To build the posterior distribution of the IRFs, Dynare
pulls parameter and shock values from the corresponding estimated dis
tributions and, for each set of draws, generates an IRF. Repeating this
process often enough generates a distribution of IRFs. TIP! If you stop
the estimation procedure after calculating the posterior mode, or carry
out maximum likelihood estimation, only the corresponding parameter
estimates will be used to generate the IRFs. If you instead carry out
a full MetropolisHastings estimation, on the other hand, the IRFs will
use the parameters the posterior distributions, including the variance of
the shocks.
18. All options available for stoch simul can simply be added to the above
options, separated by commas. To view a list of these options, either
see the Reference Manual or section 3.8 of chapter 3.
19. moments varendo: triggers the computation of the posterior distribution
of the theoretical moments of the endogenous variables as in stoch simul
(the posterior distribution of the variance decomposition is also in
cluded). ** will be implemented shortly  if not already  in Dynare
version 4.
20. ﬁltered vars: triggers the computation of the posterior distribution of
ﬁltered endogenous variables and shocks. See the note below on the
diﬀerence between ﬁltered and smoothed shocks. ** will be implemented
shortly  if not already  in Dynare version 4.
21. smoother: triggers the computation of the posterior distribution of smoothed
endogenous variables and shocks. Smoothed shocks are a reconstruction
of the values of unobserved shocks over the sample, using all the informa
tion contained in the sample of observations. Filtered shocks, instead,
are built only based on knowing past information. To calculate one pe
riod ahead prediction errors, for instance, you should use ﬁltered, not
smoothed variables.
5.8. THE COMPLETE .MOD FILE 55
22. forecast = INTEGER: computes the posterior distribution of a forecast
on INTEGER periods after the end of the sample used in estimation.
The corresponding graph includes one conﬁdence interval describing un
certainty due to parameters and one conﬁdence interval describing un
certainty due to parameters and future shocks. Note that Dynare cannot
forecast out of the posterior mode. You need to run MetropolisHastings
iterations before being able to run forecasts on an estimated model. Fi
nally, running a forecast is very similar to an IRF, as in bayesian irf,
except that the forecast does not begin at a steady state, but simply
at the point corresponding to the last set of observations. The goal
of undertaking a forecast is to see how the system returns to steady
state from this starting point. Of course, as observation do not exist
for all variables, those necessary are reconstructed by sampling out of
the posterior distribution of parameters. Again, repeating this step of
ten enough yields a posterior distribution of the forecast. ** will be
implemented shortly  if not already  in Dynare version 4.
TIP! Before launching estimation it is a good idea to make sure that your
model is correctly declared, that a steady state exists and that it can be sim
ulated for at least one set of parameter values. You may therefore want to
create a test version of your .mod ﬁle. In this test ﬁle, you would comment
out or erase the commands related to estimation, remove the prior estimates
for parameter values and replace them with actual parameter values in the
preamble, remove any nonstationary variables from your model, add a shocks
block, make sure you have steady and possibly check following the initval
block if you do not have exact steady state values and run a simulation using
stoch simul at the end of your .mod ﬁle. Details on model solution and sim
ulation can be found in Chapter 3.
Finally, coming back to our example, we could choose a standard option:
estimation(datafile=simuldataRBC,nobs=200,first obs=500,
mh replic=2000,mh nblocks=2,mh drop=0.45,mh jscale=0.8,
mode compute=6);
This ends our description of the .mod ﬁle.
5.8 The complete .mod ﬁle
To summarize and to get a complete perspective on our work so far, here is
the complete .mod ﬁle for the estimation of our very basic model. You can
ﬁnd the corresponding ﬁle in the models folder under UserGuide in your in
stallation of Dynare. The ﬁle is called RBC Est.mod.
56 CHAPTER 5. ESTIMATING DSGE MODELS  BASICS
var y c k i l y l w r z;
varexo e;
parameters beta psi delta alpha rho epsilon;
model;
(1/c) = beta*(1/c(+1))*(1+r(+1)delta);
psi*c/(1l) = w;
c+i = y;
y = (k(1)^alpha)*(exp(z)*l)^(1alpha);
w = y*((epsilon1)/epsilon)*(1alpha)/l;
r = y*((epsilon1)/epsilon)*alpha/k(1);
i = k(1delta)*k(1);
y l = y/l;
z = rho*z(1)+e;
end;
varobs y;
initval;
k = 9;
c = 0.7;
l = 0.3;
w = 2.0;
r = 0;
z = 0;
e = 0;
end;
steady;
check;
estimated params;
alpha, beta pdf, 0.35, 0.02;
beta, beta pdf, 0.99, 0.002;
delta, beta pdf, 0.025, 0.003;
psi, gamma pdf, 1.75, 0.02;
rho, beta pdf, 0.95, 0.05;
epsilon, gamma pdf, 10, 0.003;
stderr e, inv gamma pdf, 0.01, inf;
end;
estimation(datafile=simuldataRBC,nobs=200,first obs=500,
mh replic=2000,mh nblocks=2,mh drop=0.45,mh jscale=0.8,
mode compute=6);
5.9. INTERPRETING OUTPUT 57
5.9 Interpreting output
As in the case of model solution and simulation, Dynare returns both tabular
and graphical output. On the basis of the options entered in the example
.mod ﬁle above, Dynare will display the following results.
5.9.1 Tabular results
The ﬁrst results to be displayed (and calculated from a chronological stand
point) are the steady state results. Note the dummy values of 1 for the non
stationary variables Y obs and P obs. These results are followed by the eigen
values of the system, presented in the order in which the endogenous variables
are declared at the beginning of the .mod ﬁle. The table of eigenvalues is
completed with a statement about the BlanchardKahn condition being met
 hopefully!
The next set of results are for the numerical iterations necessary to ﬁnd
the posterior mode, as explained in more details in Chapter 6. The improve
ment from one iteration to the next reaches zero, Dynare give the value of
the objective function (the posterior Kernel) at the mode and displays two
important table summarizing results from posterior maximization.
The ﬁrst table summarizes results for parameter values. It includes: prior
means, posterior mode, standard deviation and tstat of the mode (based on
the assumption of a Normal, probably erroneous when undertaking Bayesian
estimation, as opposed to standard maximum likelihood), as well as the prior
distribution and standard deviation (pstdev). It is followed by a second table
summarizing the same results for the shocks. It may be entirely possible that
you get an inﬁnite value for a standard deviation, this is simply the limit case
of the inverse gamma distribution.
5.9.2 Graphical results
** corresponding graphs will be reproduced below.
The ﬁrst ﬁgure comes up soon after launching Dynare as little computa
tion is necessary to generate it. The ﬁgure returns a graphical representation
of the priors for each parameter of interest.
58 CHAPTER 5. ESTIMATING DSGE MODELS  BASICS
The second set of ﬁgures displays “MCMC univariate diagnostics”, where
MCMC stands for Monte Carlo Markov Chains. This is the main source
of feedback to gain conﬁdence, or spot a problem, with results. Recall that
Dynare completes several runs of MetropolisHastings simulations (as many as
determined in the option mh nblocks, each time starting from a diﬀerent ini
tial value). If the results from one chain are sensible, and the optimizer did not
get stuck in an odd area of the parameter subspace, two things should happen.
First, results within any of the however many iterations of MetropolisHastings
simulation should be similar. And second, results between the various chains
should be close. This is the idea of what the MCMC diagnostics track.
More speciﬁcally, the red and blue lines on the charts represent speciﬁc
measures of the parameter vectors both within and between chains. For the
results to be sensible, these should be relatively constant (although there
will always be some variation) and they should converge. Dynare reports
three measures: “interval”, being constructed from an 80% conﬁdence inter
val around the parameter mean, “m2”, being a measure of the variance and
“m3” based on third moments. In each case, Dynare reports both the within
and the between chains measures. The ﬁgure entitled “multivariate diagnos
tic” presents results of the same nature, except that they reﬂect an aggregate
measure based on the eigenvalues of the variancecovariance matrix of each
parameter.
In our example above, you can tell that indeed, we obtain convergence
and relative stability in all measures of the parameter moments. Note that
the horizontal axis represents the number of MetropolisHastings iterations
that have been undertaken, and the vertical axis the measure of the parame
ter moments, with the ﬁrst, corresponding to the measure at the initial value
of the MetropolisHastings iterations.
TIP! If the plotted moments are highly unstable or do not converge, you
may have a problem of poor priors. It is advisable to redo the estimation with
diﬀerent priors. If you have trouble coming up with a new prior, try starting
with a uniform and relatively wide prior and see where the data leads the
posterior distribution. Another approach is to undertake a greater number of
MetropolisHastings simulations.
The ﬁrst to last ﬁgure  ﬁgure 6 in our example  displays the most inter
esting set of results, towards which most of the computations undertaken by
Dynare are directed: the posterior distribution. In fact, the ﬁgure compares
the posterior to the prior distribution (black vs. grey lines). In addition,
on the posterior distribution, Dynare plots a green line which represents the
posterior mode. These allow you to make statements about your data other
than simply concerning the mean and variance of the parameters; you can also
5.9. INTERPRETING OUTPUT 59
discuss the probability that your parameter is larger or smaller than a certain
value.
TIP! These graphs are of course especially relevant and present key results,
but they can also serve as tools to detect problems or build additional conﬁ
dence in your results. First, the prior and the posterior distributions should
not be excessively diﬀerent. Second, the posterior distributions should be close
to normal, or at least not display a shape that is clearly nonnormal. Third,
the green mode (calculated from the numerical optimization of the posterior
kernel) should not be too far away from the mode of the posterior distribution.
If not, it is advisable to undertake a greater number of MetropolisHastings
simulations.
The last ﬁgure returns the smoothed estimated shocks in a useful illustra
tion to eyeball the plausibility of the size and frequency of the shocks. The
horizontal axis, in this case, represents the number of periods in the sample.
One thing to check is the fact that shocks should be centered around zero.
That is indeed the case for our example.
Chapter 6
Estimating DSGE models 
advanced topics
This chapter focusses on advanced topics and features of Dynare in the area of
model estimation. The chapter begins by presenting a more complex example
than the one used for illustration purposes in chapter 5. The goal is to show
how Dynare would be used in the more “realistic” setting of reproducing a
recent academic paper. The chapter then follows with sections on comparing
models to one another, and then to BVARs, and ends with a table summariz
ing where output series are stored and how these can be retrieved.
6.1 Alternative and nonstationary example
The example provided in chapter 5 is really only useful for illustration pur
poses. So we thought you would enjoy (and continue learning from!) a more
realistic example which reproduces the work in a recent  and highly regarded
 academic paper. The example shows how to use Dynare in a more realistic
setting, while emphasizing techniques to deal with nonstationary observations
and stochastic trends in dynamics.
6.1.1 Introducing the example
The example is drawn from Schorfheide (2000). This ﬁrst section introduces
the model, its basic intuitions and equations. We will then see in subsequent
sections how to estimate it using Dynare. Note that the original paper by
Schorfheide mainly focusses on estimation methodologies, diﬃculties and so
lutions, with a special interest in model comparison, while the mathematics
and economic intuitions of the model it evaluates are drawn from Nason and
Cogley (1994). That paper should serve as a helpful reference if anything is
61
62 CHAPTER 6. ESTIMATING DSGE MODELS  ADVANCED TOPICS
Figure 6.1: Continuous lines show the circulation of nominal funds, while
dashed lines show the ﬂow of real variables.
left unclear in the description below.
In essence, the model studied by Schorfheide (2000) is one of cash in ad
vance (CIA). The goal of the paper is to estimate the model using Bayesian
techniques, while observing only output and inﬂation. In the model, there are
several markets and actors to keep track of. So to clarify things, ﬁgure 6.1.1
sketches the main dynamics of the model. You may want to refer back to the
ﬁgure as you read through the following sections.
The economy is made up of three central agents and one secondary agent:
households, ﬁrms and banks (representing the ﬁnancial sector), and a mon
etary authority which plays a minor role. Households maximize their utility
function which depends on consumption, C
t
, and hours worked, H
t
, while
deciding how much money to hold next period in cash, M
t+1
and how much
to deposit at the bank, D
t
, in order to earn R
H,t
− 1 interest. Households
6.1. ALTERNATIVE AND NONSTATIONARY EXAMPLE 63
therefore solve the problem
max
{C
t
,H
t
,M
t+1
,D
t
}
E
0
_
∞
t=0
β
t
[(1 −φ) ln C
t
+φln(1 −H
t
)]
¸
s.t. P
t
C
t
≤ M
t
−D
t
+W
t
H
t
0 ≤ D
t
M
t+1
= (M
t
−D
t
+W
t
H
t
−P
t
C
t
) +R
H,t
D
t
+F
t
+B
t
where the second equation spells out the cash in advance constraint including
wage revenues, the third the inability to borrow from the bank and the fourth
the intertemporal budget constraint emphasizing that households accumulate
the money that remains after bank deposits and purchases on goods are de
ducted from total inﬂows made up of the money they receive from last period’s
cash balances, wages, interests, as well as dividends from ﬁrms, F
t
, and from
banks, B
t
, which in both cases are made up of net cash inﬂows deﬁned below.
Banks, on their end, receive cash deposits from households and a cash
injection, X
t
from the central bank (which equals the net change in nominal
money balances, M
t+1
− M
t
). It uses these funds to disburse loans to ﬁrms,
L
t
, on which they make a net return of R
F,t
− 1. Of course, banks are con
strained in their loans by a credit market equilibrium condition L
t
≤ X
t
+D
t
.
Finally, bank dividends, B
t
are simply equal to D
t
+R
F,t
L
t
−R
H,t
D
t
−L
t
+X
t
.
Finally, ﬁrms maximize the net present value of future dividends (dis
counted by the marginal utility of consumption, since they are owned by
households) by choosing dividends, next period’s capital stock, K
t+1
, labor
demand, N
t
, and loans. Its problem is summarized by
max
{F
t
,K
t+1
,N
t
,L
t
}
E
0
_
∞
t=0
β
t+1 F
t
C
t+1
P
t+1
_
s.t. F
t
≤ L
t
+P
t
_
K
α
t
(A
t
N
t
)
1−α
−K
t+1
+ (1 −δ)K
t
¸
−W
t
N
t
−L
t
R
F,t
W
t
N
t
≤ L
t
where the second equation makes use of the production function Y
t
= K
α
t
(A
t
N
t
)
1−α
and the real aggregate accounting constraint (goods market equilibrium) C
t
+I
t
= Y
t
,
where I
t
= K
t+1
− (1 − δ)K
t
, and where δ is the rate of depreciation. Note
that it is the ﬁrms that engage in investment in this model, by trading oﬀ
investment for dividends to consumers. The third equation simply speciﬁes
that bank loans are used to pay for wage costs.
To close the model, we add the usual labor and money market equilib
rium equations, H
t
= N
t
and P
t
C
t
= M
t
+ X
t
, as well as the condition that
R
H,t
= R
F,t
due to the equal risk proﬁles of the loans.
64 CHAPTER 6. ESTIMATING DSGE MODELS  ADVANCED TOPICS
More importantly, we add a stochastic elements to the model. The model
allows for two sources of perturbations, one real, aﬀecting technology and one
nominal, aﬀecting the money stock. These important equations are
ln A
t
= γ + ln A
t−1
+
A,t
,
A,t
∼ N(0, σ
2
A
)
and
ln m
t
= (1 −ρ) ln m
∗
+ρ ln m
t−1
+
M,t
,
M,t
∼ N(0, σ
2
M
)
where m
t
≡ M
T+1
/M
t
is the growth rate of the money stock. Note that theses
expressions for trends are not written in the most straightforward manner nor
very consistently. But we reproduced them nevertheless to make it easier to
compare this example to the original paper.
The ﬁrst equation is therefore a unit root with drift in the log of tech
nology, and the second an autoregressive stationary process in the growth
rate of money, but an AR(2) with a unit root in the log of the level of
money. This can be seen from the deﬁnition of m
t
which can be rewritten
as ln M
t+1
= ln M
t
+ ln m
t
.
1
When the above functions are maximized, we obtain the following set of
ﬁrst order and equilibrium conditions. We will not dwell on the derivations
here, to save space, but encourage you to browse Nason and Cogley (1994) for
additional details. We nonetheless give a brief intuitive explanation of each
1
Alternatively, we could have written the AR(2) process in state space form and realized
that the system has an eigenvalue of one. Otherwise said, one is a root of the second order
autoregressive lag polynomial. As usual, if the logs of a variable are speciﬁed to follow a unit
root process, the rate of growth of the series is a stationary stochastic process; see Hamilton
(1994), chapter 15, for details.
6.1. ALTERNATIVE AND NONSTATIONARY EXAMPLE 65
equation. The system comes down to
E
t
_
−
´
P
t
/
_
´
C
t+1
´
P
t+1
m
t
_
_
= βe
−α(γ+
A,t+1
)
P
t+1
_
α
´
K
α−1
t
N
1−α
t+1
+ (1 −δ)
_
/
_
´c
t+2
´
P
t+2
m
t+1
_
_
´
W
t
=
´
L
t
/N
t
φ
1 −φ
_
´
C
t
´
P
t
/ (1 −N
t
)
_
=
´
L
t
/N
t
R
t
= (1 −α)
´
P
t
e
−α(γ+
A,t+1
)
´
K
α
t−1
N
−α
t
/
´
W
t
_
´
C
t
´
P
t
_
−1
= β
_
(1 −α)
´
P
t
e
−α(γ+
A,t+1
)
´
K
α
t−1
N
1−α
t
_
×E
t
_
´
L
t
m
t
´
C
t+1
´
P
t+1
_
−1
´
C
t
+
´
K
t
= e
−α(γ+
A,t
)
´
K
α
t−1
N
1−α
+ (1 −δ)e
−(γ+
A,t
)
´
K
t−1
´
P
t
´
C = m
t
m
t
−1 +
´
D
t
=
´
L
t
´
Y
t
=
´
K
α
t−1
N
1−α
e
−α(γ+
A,t
)
ln(m
t
) = (1 −ρ) ln(m
) +ρ ln(m
t−1
) +
M,t
A
t
A
t−1
≡ dA
t
= exp(γ +
A,t
)
Y
t
/Y
t−1
= e
γ+
A,t ´
Y
t
/
´
Y
t−1
P
t
/P
t−1
= (
´
P
t
/
´
P
t−1
)(m
t−1
/e
γ+
A,t
)
where, importantly, hats over variables no longer mean deviations from steady
state, but instead represent variables that have been made stationary. We
come back to this important topic in details in section 6.1.3 below. For now,
we pause a moment to give some intuition for the above equations. In order,
these equations correspond to:
1. The Euler equation in the goods market, representing the tradeoﬀ to the
economy of moving consumption goods across time.
2. The ﬁrms’ borrowing constraint, also aﬀecting labor demand, as ﬁrms
use borrowed funds to pay for labor input.
3. The intertemporal labor market optimality condition, linking labor sup
ply, labor demand, and the marginal rate of substitution between con
sumption and leisure.
4. The equilibrium interest rate in which the marginal revenue product of
labor equals the cost of borrowing to pay for that additional unit of
labor.
66 CHAPTER 6. ESTIMATING DSGE MODELS  ADVANCED TOPICS
5. The Euler equation in the credit market, which ensures that giving up
one unit of consumption today for additional savings equals the net
present value of future consumption.
6. The aggregate resource constraint.
7. The money market equilibrium condition equating nominal consumption
demand to money demand to money supply to current nominal balances
plus money injection.
8. The credit market equilibrium condition.
9. The production function.
10. The stochastic process for money growth.
11. The stochastic process for technology.
12. The relationship between observable variables and stationary variables;
more details on these last two equations appear in the following section.
6.1.2 Declaring variables and parameters
This block of the .mod ﬁle follows the usual conventions and would look like:
var m P c e W R k d n l Y obs P obs y dA;
varexo e a e m;
parameters alp, bet, gam, mst, rho, psi, del;
where the choice of upper and lower case letters is not signiﬁcant, the ﬁrst set
of endogenous variables, up to l, are as speciﬁed in the model setup above,
and where the last ﬁve variables are deﬁned and explained in more details in
the section below on declaring the model in Dynare. The exogenous variables
are as expected and concern the shocks to the evolution of technology and
money balances.
6.1.3 The origin of nonstationarity
The problem of nonstationarity comes from having stochastic trends in tech
nology and money. The nonstationarity comes out clearly when attempting
to solve the model for a steady state and realizing it does not have one. It can
be shown that when shocks are null, real variables grow with A
t
(except for
labor, N
t
, which is stationary as there is no population growth), nominal vari
ables grow with M
t
and prices with M
t
/A
t
. Detrending therefore involves
6.1. ALTERNATIVE AND NONSTATIONARY EXAMPLE 67
the following operations (where hats over variables represent stationary vari
ables): for real variables, ˆ q
t
= q
t
/A
t
, where q
t
= [y
t
, c
t
, i
t
, k
t+1
]. For nominal
variables,
ˆ
Q
t
= Q
t
/M
t
, where Q
t
= [d
t
, l
t
, W
t
]. And for prices,
ˆ
P
t
= P
t
·A
t
/M
t
.
6.1.4 Stationarizing variables
Let’s illustrate this transformation on output, and leave the transformations
of the remaining equations as an exercise, if you wish (Nason and Cogley
(1994) includes more details on the transformations of each equation). We
stationarize output by dividing its real variables (except for labor) by A
t
. We
deﬁne
´
Y
t
to equal Y
t
/A
t
and
´
K
t
as K
t
/A
t
. NOTE! Recall from section 3.5 in
chapter 3), that in Dynare variables take the time subscript of the period in
which they are decided (in the case of the capital stock, today’s capital stock
is a result of yesterday’s decision). Thus, in the output equation, we should
actually work with
´
K
t−1
= K
t−1
/A
t−1
. The resulting equation made up of
stationary variables is
Y
t
A
t
=
_
K
t−1
A
t−1
_
α
A
1−α
t
N
1−α
t
A
−1
t
A
α
t−1
´
Y
t
=
´
K
α
t−1
N
1−α
t
_
A
t
A
t−1
_
−α
=
´
K
α
t−1
N
1−α
t
exp(−α(γ +
A,t
))
where we go from the second to the third line by taking the exponential of
both sides of the equation of motion of technology.
The above is the equation we retain for the .mod ﬁle of Dynare into which
we enter:
y=k(1)^alp*n^(1alp)*exp(alp*(gam+e a))
The other equations are entered into the .mod ﬁle after transforming them
in exactly the same way as the one above. A ﬁnal transformation to consider,
that turns out to be useful since we often deal with the growth rate of tech
nology, is to deﬁne
dA = exp(gam+e a)
by simply taking the exponential of both sides of the stochastic process of
technology deﬁned in the model setup above.
68 CHAPTER 6. ESTIMATING DSGE MODELS  ADVANCED TOPICS
6.1.5 Linking stationary variables to the data
And ﬁnally, we must make a decision as to our nonstationary observa
tions. We could simply stationarize them by working with rates of growth
(which we know are constant). In the case of output, the observable variable
would become Y
t
/Y
t−1
. We would then have to relate this observable, call it
gy obs, to our (stationary) model’s variables
´
Y
t
by using the deﬁnition that
´
Y
t
≡ Y
t
/A
t
. Thus, we add to the model block of the .mod ﬁle:
gy obs = dA*y/y(1);
where, the y of the .mod ﬁle are the stationary
´
Y
t
.
But, we could also work with nonstationary data in levels. This
complicates things somewhat, but illustrates several features of Dynare worth
highlighting; we therefore follow this path in the remainder of the example.
The result is not very diﬀerent, though, from what we just saw above. The
goal is to add a line to the model block of our .mod ﬁle that relates the non
stationary observables, call them Y
obs
, to our stationary output,
´
Y
t
. We could
simply write Y
obs
=
´
Y
t
A
t
. But since we don’t have an A
t
variable, but just a
dA
t
, we wewrite the above relationship in ratios. To the .mod ﬁle, we there
fore add:
Y obs/Y obs(1) = dA*y/y(1);
We of course do the same for prices, our other observable variable, except
that we use the relationship P
obs
=
´
P
t
M
t
/A
t
as noted earlier. The details
of the correct transformations for prices are left as an exercise and can be
checked against the results below.
6.1.6 The resulting model block of the .mod ﬁle
model;
dA = exp(gam+e a);
log(m) = (1rho)*log(mst) + rho*log(m(1))+e m;
P/(c(+1)*P(+1)*m)+bet*P(+1)*(alp*exp(alp*(gam+log(e(+1))))*k^(alp1)
*n(+1)^(1alp)+(1del)*exp((gam+log(e(+1)))))/(c(+2)*P(+2)*m(+1))=0;
W = l/n;
(psi/(1psi))*(c*P/(1n))+l/n = 0;
R = P*(1alp)*exp(alp*(gam+e a))*k(1)^alp*n^(alp)/W;
1/(c*P)bet*P*(1alp)*exp(alp*(gam+e a))*k(1)^alp*n^(1alp)/
(m*l*c(+1)*P(+1)) = 0;
c+k = exp(alp*(gam+e a))*k(1)^alp*n^(1alp)+(1del)
6.1. ALTERNATIVE AND NONSTATIONARY EXAMPLE 69
*exp((gam+e a))*k(1);
P*c = m;
m1+d = l;
e = exp(e a);
y = k(1)^alp*n^(1alp)*exp(alp*(gam+e a));
Y obs/Y obs(1) = dA*y/y(1);
P obs/P obs(1) = (p/p(1))*m(1)/dA;
end;
where, of course, the input conventions, such as ending lines with semicolons
and indicating the timing of variables in parentheses, are the same as those
listed in chapter 3.
TIP! In the above model block, notice that what we have done is in fact
relegated the nonstationarity of the model to just the last two equations,
concerning the observables which are, after all, nonstationary. The problem
that arises, though, is that we cannot linearize the above system in levels, as
the last two equations don’t have a steady state. If we ﬁrst take logs, though,
they become linear and it doesn’t matter anymore where we calculate their
derivative when taking a Taylor expansion of all the equations in the system.
Thus, when dealing with nonstationary observations, you must log
linearize your model (and not just linearize it); this is a point to which we
will return later.
6.1.7 Declaring observable variables
We begin by declaring which of our model’s variables are observables. In our
.mod ﬁle we write
varobs P obs Y obs;
to specify that our observable variables are indeed P obs and Y obs as noted
in the section above. NOTE! Recall from earlier that the number of observed
variables must be smaller or equal to the number of shocks such that the model
be estimated. If this is not the case, you should add measurement shocks to
your model where you deem most appropriate.
6.1.8 Declaring trends in observable variables
Recall that we decided to work with the nonstationary observable variables
in levels. Both output and prices exhibit stochastic trends. This can be seen
explicitly by taking the diﬀerence of logs of output and prices to compute
growth rates. In the case of output, we make use of the usual (by now!)
70 CHAPTER 6. ESTIMATING DSGE MODELS  ADVANCED TOPICS
relationship Y
t
=
´
Y
t
· A
t
. Taking logs of both sides and subtracting the same
equation scrolled back one period, we ﬁnd:
∆ln Y
t
= ∆ln
´
Y
t
+γ +
A,t
emphasizing clearly the drift term γ, whereas we know ∆ln
´
Y
t
is stationary in
steady state.
In the case of prices, we apply the same manipulations to show that:
∆ln P
t
= ∆ln
´
P
t
+ ln m
t−1
−γ −
A,t
Note from the original equation of motion of ln m
t
that in steady state,
ln m
t
= ln m
∗
, so that the drift terms in the above equation are ln m
∗
−γ.
2
In Dynare, any trends, whether deterministic or stochastic (the drift term)
must be declared up front. In the case of our example, we therefore write (in
a somewhat cumbersome manner)
observation trends;
P obs (log(mst)gam);
Y obs (gam);
end;
In general, the command observation trends speciﬁes linear trends as a
function of model parameters for the observed variables in the model.
6.1.9 Declaring unit roots in observable variables
And ﬁnally, since P obs and Y obs inherit the unit root characteristics of their
driving variables, technology and money, we must tell Dynare to use a diﬀuse
prior (inﬁnite variance) for their initialization in the Kalman ﬁlter. Note that
for stationary variables, the unconditional covariance matrix of these variables
is used for initialization. The algorithm to compute a true diﬀuse prior is taken
from Durbin and Koopman (2001). To give these instructions to Dynare, we
write in the .mod
unit root vars P obs Y obs;
NOTE! You don’t need to declare unit roots for any nonstationary model.
Unit roots are only related to stochastic trends. You don’t need to use a diﬀuse
2
This can also be see from substituting for ln m
t−1
in the above equation with the
equation of motion of ln m
t
to yield: ∆ln P
t
= ∆ln
P
t
+ln m
∗
+ρ(ln m
t−2
−ln m
∗
) +
M,t
−
γ −
A,t
where all terms on the right hand side are constant, except for ln m
∗
and γ.
6.1. ALTERNATIVE AND NONSTATIONARY EXAMPLE 71
initial condition in the case of a deterministic trend, since the variance is ﬁnite.
6.1.10 Specifying the steady state
Declaring the steady state is just as explained in details and according to
the same syntax explained in chapter 3, covering the initval, steady and
check commands. In chapter 5, section 5.5, we also discussed the usefulness
of providing an external Matlab ﬁle to solve for the steady state. In this
case, you can ﬁnd the corresponding steady state ﬁle in the models folder
under UserGuide. The ﬁle is called fs2000ns steadystate.m. There are some
things to notice. First, the output of the function is the endogenous variables
at steady state, the ys vector. The check=0 limits steady state values to
real numbers. Second, notice the declaration of parameters at the beginning;
intuitive, but tedious... This functionality may be updated in later versions of
Dynare. Third, note that the ﬁle is really only a sequential set of equalities,
deﬁning each variable in terms of parameters or variables solved in the lines
above. So far, nothing has changed with respect to the equivalent ﬁle of
chapter 5. The only novelty is the declaration of the nonstationary variables,
P obs and Y obs which take the value of 1. This is Dynare convention and
must be the case for all your nonstationary variables.
6.1.11 Declaring priors
We expand our .mod ﬁle with the following information:
estimated params;
alp, beta pdf, 0.356, 0.02;
bet, beta pdf, 0.993, 0.002;
gam, normal pdf, 0.0085, 0.003;
mst, normal pdf, 1.0002, 0.007;
rho, beta pdf, 0.129, 0.223;
psi, beta pdf, 0.65, 0.05;
del, beta pdf, 0.01, 0.005;
stderr e a, inv gamma pdf, 0.035449, inf;
stderr e m, inv gamma pdf, 0.008862, inf;
end;
6.1.12 Launching the estimation
We add the following commands to ask Dynare to run a basic estimation of
our model:
72 CHAPTER 6. ESTIMATING DSGE MODELS  ADVANCED TOPICS
estimation(datafile=fsdat,nobs=192,loglinear,mh replic=2000,
mode compute=6,mh nblocks=2,mh drop=0.45,mh jscale=0.65);
NOTE! As mentioned earlier, we need to instruct Dynare to loglinearize
our model, since it contains nonlinear equations in nonstationary variables.
A simple linearization would fail as these variables do not have a steady state.
Fortunately, taking the log of the equations involving nonstationary variables
does the job of linearizing them.
6.1.13 The complete .mod ﬁle
We have seen each part of the .mod separately; it’s now time to get a picture
of what the complete ﬁle looks like. For convenience, the ﬁle also appears in
the models folder under UserGuide in your Dynare installation. The ﬁle is
called fs2000ns.mod.
var m P c e W R k d n l Y obs P obs y dA;
varexo e a e m;
parameters alp, bet, gam, mst, rho, psi, del;
model;
dA = exp(gam+e a);
log(m) = (1rho)*log(mst) + rho*log(m(1))+e m;
P/(c(+1)*P(+1)*m)+bet*P(+1)*(alp*exp(alp*(gam+log(e(+1))))
*k^(alp1)*n(+1)^(1alp)+(1del)
*exp((gam+log(e(+1)))))/(c(+2)*P(+2)*m(+1))=0;
W = l/n;
(psi/(1psi))*(c*P/(1n))+l/n = 0;
R = P*(1alp)*exp(alp*(gam+e a))*k(1)^alp*n^(alp)/W;
1/(c*P)bet*P*(1alp)*exp(alp*(gam+e a))*k(1)^alp*n^(1alp)/
(m*l*c(+1)*P(+1)) = 0;
c+k = exp(alp*(gam+e a))*k(1)^alp*n^(1alp)+(1del)
exp((gam+e a))*k(1);
P*c = m;
m1+d = l;
e = exp(e a);
y = k(1)^alp*n^(1alp)*exp(alp*(gam+e a));
Y obs/Y obs(1) = dA*y/y(1);
P obs/P obs(1) = (p/p(1))*m(1)/dA;
end;
varobs P obs Y obs;
6.1. ALTERNATIVE AND NONSTATIONARY EXAMPLE 73
observation trends;
P obs (log(mst)gam);
Y obs (gam);
end;
unit root vars P obs Y obs;
initval;
k = 6;
m = mst;
P = 2.25;
c = 0.45;
e = 1;
W = 4;
R = 1.02;
d = 0.85;
n = 0.19;
l = 0.86;
y = 0.6;
dA = exp(gam);
end;
// the above is really only useful if you want to do a stoch simul
// of your model, since the estimation will use the Matlab
// steady state file also provided and discussed above.
steady;
estimated params;
alp, beta pdf, 0.356, 0.02;
bet, beta pdf, 0.993, 0.002;
gam, normal pdf, 0.0085, 0.003;
mst, normal pdf, 1.0002, 0.007;
rho, beta pdf, 0.129, 0.223;
psi, beta pdf, 0.65, 0.05;
del, beta pdf, 0.01, 0.005;
stderr e a, inv gamma pdf, 0.035449, inf;
stderr e m, inv gamma pdf, 0.008862, inf;
end;
estimation(datafile=fsdat,nobs=192,loglinear,mh replic=2000,
mode compute=6,mh nblocks=2,mh drop=0.45,mh jscale=0.65);
74 CHAPTER 6. ESTIMATING DSGE MODELS  ADVANCED TOPICS
Figure 6.2: At a high level, there are ﬁve basic steps to translate a model into
Dynare for successful estimation.
6.1.14 Summing it up
The explanations given above of each step necessary to translate the Schorfheide
(2000) example into language that Dynare can understand and process was
quite lengthy and involved a slew of new commands and information. It may
therefore be useful, to gain a “bird’s eyeview” on what we have just accom
plished, and summarize the most important steps at a high level. This is done
in ﬁgure 6.1.14.
6.2 Comparing models based on their posterior
distributions
** TBD
6.3. WHERE IS YOUR OUTPUT STORED? 75
6.3 Where is your output stored?
The output from estimation can be extremely varied, depending on the in
structions you give Dynare. The Reference Manual overviews the complete
set of potential output ﬁles and describes where you can ﬁnd each one.
Chapter 7
Solving DSGE models 
Behind the scenes of Dynare
7.1 Introduction
The aim of this chapter is to peer behind the scenes of Dynare, or under its
hood, to get an idea of the methodologies and algorithms used in its com
putations. Going into details would be beyond the scope of this User Guide
which will instead remain at a high level. What you will ﬁnd below will
either comfort you in realizing that Dynare does what you expected of it 
and what you would have also done if you had had to code it all yourself
(with a little extra time on your hands!), or will spur your curiosity to have
a look at more detailed material. If so, you may want to go through Michel
Juillard’s presentation on solving DSGE models to a ﬁrst and second order
(available on Michel Juillard’s website), or read Collard and Juillard (2001a)
or SchmittGrohe and Uribe (2004) which gives a good overview of the most
recent solution techniques based on perturbation methods. Finally, note that
in this chapter we will focus on stochastic models  which is where the major
complication lies, as explained in section 3.1.1 of chapter 3. For more details
on the NewtonRaphson algorithm used in Dynare to solve deterministic mod
els, see Juillard (1996).
7.2 What is the advantage of a second order
approximation?
As noted in chapter 3 and as will become clear in the section below, lin
earizing a system of equations to the ﬁrst order raises the issue of certainty
equivalence. This is because only the ﬁrst moments of the shocks enter the
linearized equations, and when expectations are taken, they disappear. Thus,
77
78
CHAPTER 7. SOLVING DSGE MODELS  BEHIND THE SCENES OF
DYNARE
unconditional expectations of the endogenous variables are equal to their non
stochastic steady state values.
This may be an acceptable simpliﬁcation to make. But depending on the
context, it may instead be quite misleading. For instance, when using sec
ond order welfare functions to compare policies, you also need second order
approximations of the policy function. Yet more clearly, in the case of asset
pricing models, linearizing to the second order enables you to take risk (or the
variance of shocks) into consideration  a highly desirable modeling feature. It
is therefore very convenient that Dynare allows you to choose between a ﬁrst
or second order linearization of your model in the option of the stoch simul
command.
7.3 How does dynare solve stochastic DSGE
models?
In this section, we shall brieﬂy overview the perturbation methods employed
by Dynare to solve DSGE models to a ﬁrst order approximation. The sec
ond order follows very much the same approach, although at a higher level
of complexity. The summary below is taken mainly from Michel Juillard’s
presentation “Computing ﬁrst order approximations of DSGE models with
Dynare”, which you should read if interested in particular details, especially
regarding second order approximations (available on Michel Juillard’s web
site).
To summarize, a DSGE model is a collection of ﬁrst order and equilibrium
conditions that take the general form:
E
t
{f(y
t+1
, y
t
, y
t−1
, u
t
)} = 0
E(u
t
) = 0
E(u
t
u
t
) = Σ
u
and where:
y : vector of endogenous variables of any dimension
u : vector of exogenous stochastic shocks of any dimension
The solution to this system is a set of equations relating variables in the
current period to the past state of the system and current shocks, that satisfy
the original system. This is what we call the policy function. Sticking to the
above notation, we can write this function as:
y
t
= g(y
t−1
, u
t
)
7.3. HOW DOES DYNARE SOLVE STOCHASTIC DSGE MODELS? 79
Then, it is straightforward to rewrite y
t+1
as
y
t+1
= g(y
t
, u
t+1
)
= g(g(y
t−1
, u
t
), u
t+1
)
We can then deﬁne a new function F, such that:
F(y
t−1
, u
t
, u
t+1
) = f(g(g(y
t−1
, u
t
), u
t+1
), g(y
t−1
, u
t
), y
t−1
, u
t
)
which enables us to rewrite our system in terms of past variables, and current
and future shocks:
E
t
[F(y
t−1
, u
t
, u
t+1
)] = 0
We then venture to linearize this model around a steady state deﬁned as:
f(¯ y, ¯ y, ¯ y, 0) = 0
having the property that:
¯ y = g(¯ y, 0)
The ﬁrst order Taylor expansion around ¯ y yields:
E
t
_
F
(1)
(y
t−1
, u
t
, u
t+1
)
_
=
E
t
_
f(¯ y, ¯ y, ¯ y) +f
y
+
_
g
y
(g
y
ˆ y +g
u
u) +g
u
u
_
+f
y
0
(g
y
ˆ y +g
u
u) +f
y
−
ˆ y +f
u
u
_
= 0
with ˆ y = y
t−1
− ¯ y, u = u
t
, u
= u
t+1
, f
y
+
=
∂f
∂y
t+1
, f
y
0
=
∂f
∂y
t
, f
y
−
=
∂f
∂y
t−1
,
f
u
=
∂f
∂u
t
, g
y
=
∂g
∂y
t−1
, g
u
=
∂g
∂u
t
.
Taking expectations (we’re almost there!):
E
t
_
F
(1)
(y
t−1
, u
t
, u
t+1
)
_
=
f(¯ y, ¯ y, ¯ y) +f
y
+
(g
y
(g
y
ˆ y +g
u
u))
+f
y
0
(g
y
ˆ y +g
u
u) +f
y
−
ˆ y +f
u
u
_
=
_
f
y
+
g
y
g
y
+f
y
0
g
y
+f
y
−
_
ˆ y +
_
f
y
+
g
y
g
u
+f
y
0
g
u
+f
u
_
u
= 0
As you can see, since future shocks only enter with their ﬁrst moments
(which are zero in expectations), they drop out when taking expectations of
the linearized equations. This is technically why certainty equivalence holds
80
CHAPTER 7. SOLVING DSGE MODELS  BEHIND THE SCENES OF
DYNARE
in a system linearized to its ﬁrst order. The second thing to note is that we
have two unknown variables in the above equation: g
y
and g
u
each of which
will help us recover the policy function g.
Since the above equation holds for any ˆ y and any u, each parenthesis must
be null and we can solve each at a time. The ﬁrst, yields a quadratic equation
in g
y
, which we can solve with a series of algebraic trics that are not all imme
diately apparent (but detailed in Michel Juillard’s presentation). Incidentally,
one of the conditions that comes out of the solution of this equation is the
BlanchardKahn condition: there must be as many roots larger than one in
modulus as there are forwardlooking variables in the model. Having recov
ered g
y
, recovering g
u
is then straightforward from the second parenthesis.
Finally, notice that a ﬁrst order linearization of the function g yields:
y
t
= ¯ y +g
y
ˆ y +g
u
u
And now that we have g
y
and g
u
, we have solved for the (approximate) policy
(or decision) function and have succeeded in solving our DSGE model. If we
were interested in impulse response functions, for instance, we would simply
iterate the policy function starting from an initial value given by the steady
state.
The second order solution uses the same “perturbation methods” as above
(the notion of starting from a function you can solve  like a steady state 
and iterating forward), yet applies more complex algebraic techniques to re
cover the various partial derivatives of the policy function. But the general
approach is perfectly isomorphic. Note that in the case of a second order
approximation of a DSGE model, the variance of future shocks remains after
taking expectations of the linearized equations and therefore aﬀects the level
of the resulting policy function.
Chapter 8
Estimating DSGE models 
Behind the scenes of Dynare
This chapter focuses on the theory of Bayesian estimation. It begins by mo
tivating Bayesian estimation by suggesting some arguments in favor of it as
opposed to other forms of model estimation. It then attempts to shed some
light on what goes on in Dynare’s machinery when it estimates DSGE models.
To do so, this section surveys the methodologies adopted for Bayesian estima
tion, including deﬁning what are prior and posterior distributions, using the
Kalman ﬁlter to ﬁnd the likelihood function, estimating the posterior function
thanks to the MetropolisHastings algorithm, and comparing models based on
posterior distributions.
8.1 Advantages of Bayesian estimation
Bayesian estimation is becoming increasingly popular in the ﬁeld of macro
economics. Recent papers have attracted signiﬁcant attention; some of these
include: Schorfheide (2000) which uses Bayesian methods to compare the
ﬁt of two competing DSGE models of consumption, Lubik and Schorfheide
(2003) which investigates whether central banks in small open economies re
spond to exchange rate movements, Smets and Wouters (2003) which ap
plies Bayesian estimation techniques to a model of the Eurozone, Ireland
(2004) which emphasizes instead maximum likelihood estimation, Fernandez
Villaverde and RubioRamirez (2004) which reviews the econometric proper
ties of Bayesian estimators and compare estimation results with maximum
likelihood and BVAR methodologies, Lubik and Schorfheide (2005) which ap
plies Bayesian estimation methods to an open macro model focussing on issues
of misspeciﬁcation and identiﬁcation, and ﬁnally Rabanal and RubioRamirez
(2005) which compares the ﬁt, based on posterior distributions, of four com
peting speciﬁcations of New Keynesian monetary models with nominal rigidi
ties.
81
82
CHAPTER 8. ESTIMATING DSGE MODELS  BEHIND THE SCENES
OF DYNARE
There are a multitude of advantages of using Bayesian methods to esti
mate a model, but ﬁve of these stand out as particularly important and general
enough to mention here.
First, Bayesian estimation ﬁts the complete, solved DSGE model, as op
posed to GMM estimation which is based on particular equilibrium relation
ships such as the Euler equation in consumption. Likewise, estimation in the
Bayesian case is based on the likelihood generated by the DSGE system, rather
than the more indirect discrepancy between the implied DSGE and VAR im
pulse response functions. Of course, if your model is entirely misspeciﬁed,
estimating it using Bayesian techniques could be a disadvantage.
Second, Bayesian techniques allow the consideration of priors which work
as weights in the estimation process so that the posterior distribution avoids
peaking at strange points where the likelihood peaks. Indeed, due to the
stylized and often misspeciﬁed nature of DSGE models, the likelihood often
peaks in regions of the parameter space that are contradictory with common
observations, leading to the “dilemma of absurd parameter estimates”.
Third, the inclusion of priors also helps identifying parameters. Unfortu
nately, when estimating a model, the problem of identiﬁcation often arises. It
can be summarized by diﬀerent values of structural parameters leading to the
same joint distribution for observables. More technically, the problem arises
when the posterior distribution is ﬂat over a subspace of parameter values.
But the weighting of the likelihood with prior densities often leads to adding
just enough curvature in the posterior distribution to facilitate numerical max
imization.
Fourth, Bayesian estimation explicitly addresses model misspeciﬁcation by
including shocks, which can be interpreted as observation errors, in the struc
tural equations.
Sixth, Bayesian estimation naturally leads to the comparison of models
based on ﬁt. Indeed, the posterior distribution corresponding to competing
models can easily be used to determine which model best ﬁts the data. This
procedure, as other topics mentioned above, is discussed more technically in
the subsection below.
8.2 The basic mechanics of Bayesian estimation
This and the following subsections are based in great part on work by, and
discussions with, St´ephane Adjemian, a member of the Dynare development
8.2. THE BASIC MECHANICS OF BAYESIAN ESTIMATION 83
team. Some of this work, although summarized in presentation format, is
available in the “Events” page of the Dynare website. Other helpful material
includes An and Schorfheide (2006), which includes a clear and quite complete
introduction to Bayesian estimation, illustrated by the application of a sim
ple DSGE model. Also, the appendix of Schorfheide (2000) contains details
as to the exact methodology and possible diﬃculties encountered in Bayesian
estimation. You may also want to take a glance at Hamilton (1994), chapter
12, which provides a very clear, although somewhat outdated, introduction
to the basic mechanics of Bayesian estimation. Finally, the websites of Frank
Schorfheide and Jesus FernandezVillaverde contain a wide variety of very
helpful material, from example ﬁles to lecture notes to related papers. Fi
nally, remember to also check the open online examples of the Dynare website
for examples of .mod ﬁles touching on Bayesian estimation.
At its most basic level, Bayesian estimation is a bridge between calibra
tion and maximum likelihood. The tradition of calibrating models is inherited
through the speciﬁcation of priors. And the maximum likelihood approach en
ters through the estimation process based on confronting the model with data.
Together, priors can be seen as weights on the likelihood function in order to
give more importance to certain areas of the parameter subspace. More tech
nically, these two building blocks  priors and likelihood functions  are tied
together by Bayes’ rule. Let’s see how.
First, priors are described by a density function of the form
p(θ
A
A)
where A stands for a speciﬁc model, θ
A
represents the parameters of model A,
p(•) stands for a probability density function (pdf) such as a normal, gamma,
shifted gamma, inverse gamma, beta, generalized beta, or uniform function.
Second, the likelihood function describes the density of the observed data,
given the model and its parameters:
L(θ
A
Y
T
, A) ≡ p(Y
T
θ
A
, A)
where Y
T
are the observations until period T, and where in our case the
likelihood is recursive and can be written as:
p(Y
T
θ
A
, A) = p(y
0
θ
A
, A)
T
t=1
p(y
t
Y
t−1
, θ
A
, A)
We now take a step back. Generally speaking, we have a prior density p(θ)
on the one hand, and on the other, a likelihood p(Y
T
θ). In the end, we are
interested in p(θY
T
), the posterior density. Using the Bayes theorem
84
CHAPTER 8. ESTIMATING DSGE MODELS  BEHIND THE SCENES
OF DYNARE
twice we obtain this density of parameters knowing the data. Generally, we
have
p(θY
T
) =
p(θ; Y
T
)
p(Y
T
)
We also know that
p(Y
T
θ) =
p(θ; Y
T
)
p(θ)
⇔p(θ; Y
T
) = p(Y
T
θ) ×p(θ)
By using these identities, we can combine the prior density and the
likelihood function discussed above to get the posterior density:
p(θ
A
Y
T
, A) =
p(Y
T
θ
A
, A)p(θ
A
A)
p(Y
T
A)
where p(Y
T
A) is the marginal density of the data conditional on the model:
p(Y
T
A) =
_
Θ
A
p(θ
A
; Y
T
A)dθ
A
Finally, the posterior kernel (or unnormalized posterior density, given
that the marginal density above is a constant or equal for any parameter),
corresponds to the numerator of the posterior density:
p(θ
A
Y
T
, A) ∝ p(Y
T
θ
A
, A)p(θ
A
A) ≡ K(θ
A
Y
T
, A)
This is the fundamental equation that will allow us to rebuild all posterior mo
ments of interest. The trick will be to estimate the likelihood function with
the help of the Kalman ﬁlter and then simulate the posterior kernel using
a samplinglike or Monte Carlo method such as the MetropolisHastings.
These topics are covered in more details below. Before moving on, though,
the subsection below gives a simple example based on the above reasoning
of what we mean when we say that Bayesian estimation is “somewhere in
between calibration and maximum likelihood estimation”. The example is
drawn from Zellner (1971), although other similar examples can be found in
Hamilton (1994), chapter 12.
8.2.1 Bayesian estimation: somewhere between calibration
and maximum likelihood estimation  an example
Suppose a data generating process y
t
= µ + ε
t
for t = 1, ..., T, where ε
t
∼
N(0, 1) is gaussian white noise. Then, the likelihood is given by
p(Y
T
µ) = (2π)
−
T
2
e
−
1
2
T
t=1
(y
t
−µ)
2
We know from the above that ´ µ
ML,T
=
1
T
T
t=1
y
t
≡ y and that V[´ µ
ML,T
] =
1
T
.
8.3. DSGE MODELS AND BAYESIAN ESTIMATION 85
In addition, let our prior be a gaussian distribution with expectation µ
0
and variance σ
2
µ
. Then, the posterior density is deﬁned, up to a constant, by:
p (µY
T
) ∝ (2πσ
2
µ
)
−
1
2
e
−
1
2
(µ−µ
0
)
2
σ
2
µ
×(2π)
−
T
2
e
−
1
2
T
t=1
(y
t
−µ)
2
Or equivalently, p (µY
T
) ∝ e
−
(µ−E[µ])
2
V[µ]
, with
V[µ] =
1
_
1
T
_
−1
+σ
−2
µ
and
E[µ] =
_
1
T
_
−1
´ µ
ML,T
+σ
−2
µ
µ
0
_
1
T
_
−1
+σ
−2
µ
From this, we can tell that the posterior mean is a convex combination of
the prior mean and the ML estimate. In particular, if σ
2
µ
→ ∞ (ie, we have
no prior information, so we just estimate the model) then E[µ] → ´ µ
ML,T
, the
maximum likelihood estimator. But if σ
2
µ
→0 (ie, we’re sure of ourselves and
we calibrate the parameter of interest, thus leaving no room for estimation)
then E[µ] → µ
0
, the prior mean. Most of the time, we’re somewhere in the
middle of these two extremes.
8.3 DSGE models and Bayesian estimation
8.3.1 Rewriting the solution to the DSGE model
Recall from chapter 7 that any DSGE model, which is really a collection of ﬁrst
order and equilibrium conditions, can be written in the formE
t
{f(y
t+1
, y
t
, y
t−1
, u
t
)} =
0, taking as a solution equations of the type y
t
= g(y
t−1
, u
t
), which we call
the decision rule. In more appropriate terms for what follows, we can rewrite
the solution to a DSGE model as a system in the following manner:
y
∗
t
= M¯ y(θ) +Mˆ y
t
+N(θ)x
t
+η
t
ˆ y
t
= g
y
(θ)ˆ y
t−1
+g
u
(θ)u
t
E(η
t
η
t
) = V (θ)
E(u
t
u
t
) = Q(θ)
where ˆ y
t
are variables in deviations from steady state, ¯ y is the vector of steady
state values and θ the vector of deep (or structural) parameters to be esti
mated. Other variables are described below.
The second equation is the familiar decision rule mentioned above. But
the equation expresses a relationship among true endogenous variables that
are not directly observed. Only y
∗
t
is observable, and it is related to the true
86
CHAPTER 8. ESTIMATING DSGE MODELS  BEHIND THE SCENES
OF DYNARE
variables with an error η
t
. Furthermore, it may have a trend, which is captured
with N(θ)x
t
to allow for the most general case in which the trend depends on
the deep parameters. The ﬁrst and second equations above therefore naturally
make up a system of measurement and transition or state equations, respec
tively, as is typical for a Kalman ﬁlter (you guessed it, it’s not a coincidence!).
8.3.2 Estimating the likelihood function of the DSGE model
The next logical step is to estimate the likelihood of the DSGE solution system
mentioned above. The ﬁrst apparent problem, though, is that the equations
are non linear in the deep parameters. Yet, they are linear in the endogenous
and exogenous variables so that the likelihood may be evaluated with a linear
prediction error algorithm like the Kalman ﬁlter. This is exactly what Dynare
does. As a reminder, here’s what the Kalman ﬁlter recursion does.
For t = 1, . . . , T and with initial values y
1
and P
1
given, the recursion
follows
v
t
= y
∗
t
− ¯ y
∗
−Mˆ y
t
−Nx
t
F
t
= MP
t
M
+V
K
t
= g
y
P
t
g
y
F
−1
t
ˆ y
t+1
= g
y
ˆ y
t
+K
t
v
t
P
t+1
= g
y
P
t
(g
y
−K
t
M)
+g
u
Qg
u
For more details on the Kalman ﬁlter, see Hamilton (1994), chapter 13.
From the Kalman ﬁlter recursion, it is possible to derive the loglikelihood
given by
ln L(θY
∗
T
) = −
Tk
2
ln(2π) −
1
2
T
t=1
F
t
 −
1
2
v
t
F
−1
t
v
t
where the vector θ contains the parameters we have to estimate: θ, V (θ) and
Q(θ) and where Y
∗
T
expresses the set of observable endogenous variables y
∗
t
found in the measurement equation.
The loglikelihood above gets us one step closer to our goal of ﬁnding the
posterior distribution of our parameters. Indeed, the log posterior kernel
can be expressed as
ln K(θY
∗
T
) = ln L(θY
∗
T
) + ln p(θ)
where the ﬁrst term on the right hand side is now known after carrying out the
Kalman ﬁlter recursion. The second, recall, are the priors, and are also known.
8.3. DSGE MODELS AND BAYESIAN ESTIMATION 87
8.3.3 Finding the mode of the posterior distribution
Next, to ﬁnd the mode of the posterior distribution  a key parameter and
an important output of Dynare  we simply maximize the above log posterior
kernel with respect to θ. This is done in Dynare using numerical methods.
Recall that the likelihood function is not Gaussian with respect to θ but to
functions of θ as they appear in the state equation. Thus, this maximization
problem is not completely straightforward, but fortunately doable with mod
ern computers.
8.3.4 Estimating the posterior distribution
Finally, we are now in a position to ﬁnd the posterior distribution of our
parameters. The distribution will be given by the kernel equation above,
but again, it is a nonlinear and complicated function of the deep parameters
θ. Thus, we cannot obtain an explicit form for it. We resort, instead, to
samplinglike methods, of which the MetropolisHastings has been retained in
the literature as particularly eﬃcient. This is indeed the method adopted by
Dynare.
The general idea of the MetropolisHastings algorithm is to simulate the
posterior distribution. It is a “rejection sampling algorithm” used to generate
a sequence of samples (also known as a “Markov Chain” for reasons that will
become apparent later) from a distribution that is unknown at the outset.
Remember that all we have is the posterior mode; we are instead more often
interested in the mean and variance of the estimators of θ. To do so, the
algorithm builds on the fact that under general conditions the distribution
of the deep parameters will be asymptotically normal. The algorithm, in the
words of An and Shorfheide, “constructs a Gaussian approximation around
the posterior mode and uses a scaled version of the asymptotic covariance
matrix as the covariance matrix for the proposal distribution. This allows for
an eﬃcient exploration of the posterior distribution at least in the neighbor
hood of the mode” (An and Schorfheide (2006), p. 18). More precisely, the
MetropolisHastings algorithm implements the following steps:
1. Choose a starting point θ
◦
, where this is typically the posterior mode,
and run a loop over 234.
2. Draw a proposal θ
∗
from a jumping distribution
J(θ
∗
θ
t−1
) = N(θ
t−1
, cΣ
m
)
where Σ
m
is the inverse of the Hessian computed at the posterior mode.
88
CHAPTER 8. ESTIMATING DSGE MODELS  BEHIND THE SCENES
OF DYNARE
3. Compute the acceptance ratio
r =
p(θ
∗
Y
T
)
p(θ
t−1
Y
T
)
=
K(θ
∗
Y
T
)
K(θ
t−1
Y
T
)
4. Finally accept or discard the proposal θ
∗
according to the following rule,
and update, if necessary, the jumping distribution:
θ
t
=
_
θ
∗
with probability min(r, 1)
θ
t−1
otherwise.
Figure 8.3.4 tries to clarify the above. In step 1, choose a candidate
paramter, θ
∗
from a Normal distribution, whose mean has been set to θ
t−1
(this will become clear in just a moment). In step 2, compute the value of
the posterior kernel for that candidate parameter, and compare it to the value
of the kernel from the mean of the drawing distribution. In step 3, decide
whether or not to hold on to your candidate parameter. If the acceptance
ratio is greater than one, then deﬁnitely keep your candidate. Otherwise, go
back to the candidate of last period (this is true in very coarse terms, notice
that in fact you would keep your candidate only with a probability less than
one). Then, do two things. Update the mean of your drawing distribution,
and note the value of the parameter your retain. After having repeated these
steps often enough, in the ﬁnal step, build a histogram of those retained val
ues. Of course, the point is for each “bucket” of the histogram to shrink to
zero. This “smoothed histogram” will eventually be the posterior distribution
after suﬃcient iterations of the above steps.
But why have such a complicated acceptance rule? The point is to be able
to visit the entire domain of the posterior distribution. We should not be too
quick to simply throw out the candidate giving a lower value of the posterior
kernel, just in case using that candidate for the mean of the drawing distri
bution allows us to to leave a local maximum and travel towards the global
maximum. Metaphorically, the idea is to allow the search to turn away from
taking a small step up, and instead take a few small steps down in the hope
of being able to take a big step up in the near future. Of course, an important
parameter in this searching procedure is the variance of the jumping distri
bution and in particular the scale factor. If the scale factor is too small,
the acceptance rate (the fraction of candidate parameters that are accepted
in a window of time) will be too high and the Markov Chain of candidate
parameters will “mix slowly”, meaning that the distribution will take a long
time to converge to the posterior distribution since the chain is likely to get
“stuck” around a local maximum. On the other hand, if the scale factor is
too large, the acceptance rate will be very low (as the candidates are likely to
land in regions of low probability density) and the chain will spend too much
8.3. DSGE MODELS AND BAYESIAN ESTIMATION 89
Figure 8.1: The above sketches the MetropolisHastings algorithm, used to
build the posterior distribution function. Imagine repeating these steps a large
number of times, and smoothing the “histogram” such that each “bucket” has
inﬁnitely small width.
time in the tails of the posterior distribution.
While these steps are mathematically clear, at least to a machine needing
to undertake the above calculations, several practical questions arise when
carrying out Bayesian estimation. These include: How should we choose the
scale factor c (variance of the jumping distribution)? What is a satisfactory
acceptance rate? How many draws are ideal? How is convergence of the
MetropolisHastings iterations assessed? These are all important questions
that will come up in your usage of Dynare. They are addressed as clearly as
possible in section 5.7 of Chapter 5.
90
CHAPTER 8. ESTIMATING DSGE MODELS  BEHIND THE SCENES
OF DYNARE
8.4 Comparing models using posterior
distributions
As mentioned earlier, while touting the advantages of Bayesian estimation,
the posterior distribution oﬀers a particularly natural method of comparing
models. Let’s look at an illustration.
Suppose we have a prior distribution over two competing models: p(A)
and p(B). Using Bayes’ rule, we can compute the posterior distribution over
models, where I = A, B
p(IY
T
) =
p(I)p(Y
T
I)
I=A,B
p(I)p(Y
T
I)
where this formula may easily be generalized to a collection of N models.
Then, the comparison of the two models is done very naturally through the
ratio of the posterior model distributions. We call this the posterior odds
ratio:
p(AY
T
)
p(BY
T
)
=
p(A)
p(B)
p(Y
T
A)
p(Y
T
B)
The only complication is ﬁnding the magrinal density of the data condi
tional on the model, p(Y
T
I), which is also the denominator of the posterior
density p(θY
T
) discussed earlier. This requires some detailed explanations
of their own.
For each model I = A, B we can evaluate, at least theoretically, the
marginal density of the data conditional on the model by integrating out
the deep parameters θ
I
from the posterior kernel:
p(Y
T
I) =
_
Θ
I
p(θ
I
; Y
T
θ
I
, I)dθ
I
=
_
Θ
I
p(θ
I
I) ×p(Y
T
θ
I
, I)dθ
I
Note that the expression inside the integral sign is exactly the posterior kernel.
To remind you of this, you may want to glance back at the ﬁrst subsection
above, specifying the basic mechanics of Bayesian estimation.
To obtain the marginal density of the data conditional on the model, there
are two options. The ﬁrst is to assume a functional form of the posterior kernel
that we can integrate. The most straightforward and the best approximation,
especially for large samples, is the Gaussian (called a Laplace approxima
tion). In this case, we would have the following estimator:
´ p(Y
T
I) = (2π)
k
2
Σ
θ
m
I

1
2
p(θ
m
I
Y
T
, I)p(θ
m
I
I)
where θ
m
I
is the posterior mode. The advantage of this technique is its com
putational eﬃciency: time consuming MetropolisHastings iterations are not
8.4. COMPARING MODELS USING POSTERIOR DISTRIBUTIONS 91
necessary, only the numerically calculated posterior mode is required.
The second option is instead to use information from the Metropolis
Hastings runs and is typically referred to as the Harmonic Mean Esti
mator. The idea is to simulate the marginal density of interest and to simply
take an average of these simulated values. To start, note that
p(Y
T
I) = E
_
f(θ
I
)
p(θ
I
I)p(Y
T
θ
I
, I)
¸
¸
¸
¸
θ
I
, I
_
−1
where f is a probability density function, since
E
_
f(θ
I
)
p(θ
I
I)p(Y
T
θ
I
, I)
¸
¸
¸
¸
θ
I
, I
_
=
_
Θ
I
f(θ)dθ
_
Θ
I
p(θ
I
I)p(Y
T
θ
I
, I)dθ
I
and the numerator integrates out to one (seeGeweke (1999) for more details).
This suggests the following estimator of the marginal density
´ p(Y
T
I) =
_
1
B
B
b=1
f(θ
(b)
I
)
p(θ
(b)
I
I)p(Y
T
θ
(b)
I
, I)
_
−1
where each drawn vector θ
(b)
I
comes from the MetropolisHastings iterations
and where the probability density function f can be viewed as a weights on
the posterior kernel in order to downplay the importance of extreme values of
θ. Geweke (1999) suggests to use a truncated Gaussian function, leading to
what is typically referred to as the Modiﬁed Harmonic Mean Estimator.
Chapter 9
Optimal policy under
commitment
93
Chapter 10
Troubleshooting
To make sure this section is as user friendly as possible, the best is to compile
what users have to say! Please let me know what your most common problem
is with Dynare, how Dynare tells you about it and how you solve it. Thanks
for your precious help!
95
Bibliography
An, S., and F. Schorfheide (2006): “Bayesian Analysis of DSGE Models,”
Econometric Review, Forthcoming.
Clarida, R., J. Gali, and M. Gertler (1999): “The Science of Monetary
Policy: A New Keynesian Perspective,” Journal of Economic Literature,
XXXVII, 1661–1707.
Collard, F., and M. Juillard (2001a): “Accuracy of stochastic pertur
bation methods: The case of asset pricing models,” Journal of Economic
Dynamics and Control, 25(67), 979–999.
(2001b): “A HigherOrder Taylor Expansion Approach to Simulation
of Stochastic ForwardLooking Models with an Application to a Nonlinear
Phillips Curve Model,” Computational Economics, 17(23), 125–39.
(2003): “Stochastic simulations with DYNARE. A practical guide.,”
CEPREMAP mimeo.
Durbin, J., and S. Koopman (2001): Time Series Analysis by State Space
Methods. Oxford University Press, Oxford, U.K.
FernandezVillaverde, J., and J. F. RubioRamirez (2004): “Compar
ing dynamic equilibrium models to data: a Bayesian approach,” Journal of
Econometrics, 123(1), 153–187.
Geweke, J. (1999): “Using Simulation Methods for Bayesian Econometric
Models: In ference, Development and Communication,” Econometric Re
view, 18(1), 1–126.
Hamilton, J. D. (1994): Time Series Analysis. Princeton University Press,
Princeton, NJ.
Ireland, P. N. (2004): “A method for taking models to the data,” Journal
of Economic Dynamics and Control, 28(6), 1205–1226.
Juillard, M. (1996): “Dynare : a program for the resolution and simulation
of dynamic models with forward variables through the use of a relaxation
algorithm,” CEPREMAP working papers 9602, CEPREMAP.
97
98 BIBLIOGRAPHY
Lubik, T., and F. Schorfheide (2003): “Do Central Banks Respond to
Exchange Rate Movements? A Structural Investigation,” Economics Work
ing Paper Archive 505, The Johns Hopkins University,Department of Eco
nomics.
(2005): “A Bayesian Look at New Open Economy Macroeco
nomics,” Economics Working Paper Archive 521, The Johns Hopkins Uni
versity,Department of Economics.
Nason, J. M., and T. Cogley (1994): “Testing the Implications of Long
Run Neutrality for Monetary Business Cycle Models,” Journal of Applied
Econometrics, 9(S), S37–70.
Rabanal, P., and J. F. RubioRamirez (2005): “Comparing New Key
nesian models of the business cycle: A Bayesian approach,” Journal of
Monetary Economics, 52(6), 1151–1166.
SchmittGrohe, S., and M. Uribe (2004): “Solving dynamic general equi
librium models using a secondorder approximation to the policy function,”
Journal of Economic Dynamics and Control, 28(4), 755–775.
Schorfheide, F. (2000): “Loss functionbased evaluation of DSGE models,”
Journal of Applied Econometrics, 15(6), 645–670.
Smets, F., and R. Wouters (2003): “An Estimated Dynamic Stochastic
General Equilibrium Model of the Euro Area,” Journal of the European
Economic Association, 1(5), 1123–1175.
Zellner, A. (1971): An Introduction to Bayesian Inference in Econometrics.
John Wiley & Sons, Inc., New York.
Dynare v4  User Guide Public beta version
Tommaso Mancini Griﬀoli tommaso.mancini@stanfordalumni.org This draft: June 2010
iii
Copyright c 20072010 Tommaso Mancini Griﬀoli Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.3 or any later version published by the Free Software Foundation; with no Invariant Sections, no FrontCover Texts, and no BackCover Texts. A copy of the license can be found at: http://www.gnu.org/licenses/ fdl.txt
Contents
Contents List of Figures 1 Introduction 1.1 About this Guide  approach and structure 1.2 What is Dynare? . . . . . . . . . . . . . . . 1.3 Additional sources of help . . . . . . . . . . 1.4 Nomenclature . . . . . . . . . . . . . . . . . 1.5 v4, what’s new and backward compatibility 2 Installing Dynare 2.1 Dynare versions . . . . . 2.2 System requirements . . 2.3 Installing Dynare . . . . 2.4 MATLAB particularities iv vii 1 1 2 4 5 5 7 7 7 7 8 9 9 10 11 15 15 16 16 17 18 18 19 19 19 20 20 21
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
3 Solving DSGE models  basics 3.1 A fundamental distinction . . . . . . . . . . . . . . . . . . . 3.1.1 NOTE! Deterministic vs stochastic models . . . . . . 3.2 Introducing an example . . . . . . . . . . . . . . . . . . . . 3.3 Dynare .mod ﬁle structure . . . . . . . . . . . . . . . . . . . 3.4 Filling out the preamble . . . . . . . . . . . . . . . . . . . . 3.4.1 The deterministic case . . . . . . . . . . . . . . . . . 3.4.2 The stochastic case . . . . . . . . . . . . . . . . . . . 3.4.3 Comments on your ﬁrst lines of Dynare code . . . . 3.5 Specifying the model . . . . . . . . . . . . . . . . . . . . . . 3.5.1 Model in Dynare notation . . . . . . . . . . . . . . . 3.5.2 General conventions . . . . . . . . . . . . . . . . . . 3.5.3 Notational conventions . . . . . . . . . . . . . . . . . 3.5.4 Timing conventions . . . . . . . . . . . . . . . . . . 3.5.5 Conventions specifying nonpredetermined variables 3.5.6 Linear and loglinearized models . . . . . . . . . . . 3.6 Specifying steady states and/or initial values . . . . . . . . iv
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
CONTENTS 3.6.1 Stochastic models and steady states . . . . 3.6.2 Deterministic models and initial values . . . 3.6.3 Finding a steady state . . . . . . . . . . . . 3.6.4 Checking system stability . . . . . . . . . . 3.7 Adding shocks . . . . . . . . . . . . . . . . . . . . 3.7.1 Deterministic models  temporary shocks . 3.7.2 Deterministic models  permanent shocks . 3.7.3 Stochastic models . . . . . . . . . . . . . . 3.8 Selecting a computation . . . . . . . . . . . . . . . 3.8.1 For deterministic models . . . . . . . . . . . 3.8.2 For stochastic models . . . . . . . . . . . . 3.9 The complete .mod ﬁle . . . . . . . . . . . . . . . . 3.9.1 The stochastic model . . . . . . . . . . . . . 3.9.2 The deterministic model (case of temporary 3.10 File execution and results . . . . . . . . . . . . . . 3.10.1 Results  stochastic models . . . . . . . . . 3.10.2 Results  deterministic models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . shock) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
v 21 23 23 24 25 25 25 27 27 28 28 31 31 32 33 33 34 37 37 37 38 41 42 43 43 44 44 44 44 46 47 47 48 48 49 49 49 52 55 57 57 57
4 Solving DSGE models  advanced topics 4.1 Dynare features and functionality . . . . . . . . . . . . . . 4.1.1 Other examples . . . . . . . . . . . . . . . . . . . . 4.1.2 Alternative, complete example . . . . . . . . . . . 4.1.3 Finding, saving and viewing your output . . . . . . 4.1.4 Referring to external ﬁles . . . . . . . . . . . . . . 4.1.5 Inﬁnite eigenvalues . . . . . . . . . . . . . . . . . . 4.2 Files created by Dynare . . . . . . . . . . . . . . . . . . . 4.3 Modeling tips . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Stationarizing your model . . . . . . . . . . . . . . 4.3.2 Expectations taken in the past . . . . . . . . . . . 4.3.3 Inﬁnite sums . . . . . . . . . . . . . . . . . . . . . 4.3.4 Inﬁnite sums with changing timing of expectations 5 Estimating DSGE models  basics 5.1 Introducing an example . . . . . . 5.2 Declaring variables and parameters 5.3 Declaring the model . . . . . . . . 5.4 Declaring observable variables . . . 5.5 Specifying the steady state . . . . 5.6 Declaring priors . . . . . . . . . . . 5.7 Launching the estimation . . . . . 5.8 The complete .mod ﬁle . . . . . . . 5.9 Interpreting output . . . . . . . . . 5.9.1 Tabular results . . . . . . . 5.9.2 Graphical results . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
6 Estimating DSGE models  advanced topics 6.1 Alternative and nonstationary example . . . . . . . . . 6.1.1 Introducing the example . . . . . . . . . . . . . . 6.1.2 Declaring variables and parameters . . . . . . . . 6.1.3 The origin of nonstationarity . . . . . . . . . . . 6.1.4 Stationarizing variables . . . . . . . . . . . . . . 6.1.5 Linking stationary variables to the data . . . . . 6.1.6 The resulting model block of the .mod ﬁle . . . . 6.1.7 Declaring observable variables . . . . . . . . . . . 6.1.8 Declaring trends in observable variables . . . . . 6.1.9 Declaring unit roots in observable variables . . . 6.1.10 Specifying the steady state . . . . . . . . . . . . 6.1.11 Declaring priors . . . . . . . . . . . . . . . . . . 6.1.12 Launching the estimation . . . . . . . . . . . . . 6.1.13 The complete .mod ﬁle . . . . . . . . . . . . . . . 6.1.14 Summing it up . . . . . . . . . . . . . . . . . . . 6.2 Comparing models based on their posterior distributions 6.3 Where is your output stored? . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
61 61 61 66 66 67 68 68 69 69 70 71 71 71 72 74 74 75
7 Solving DSGE models  Behind the scenes of Dynare 77 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 7.2 What is the advantage of a second order approximation? . . . . 77 7.3 How does dynare solve stochastic DSGE models? . . . . . . . . 78 8 Estimating DSGE models  Behind the scenes of Dynare 81 8.1 Advantages of Bayesian estimation . . . . . . . . . . . . . . . . 81 8.2 The basic mechanics of Bayesian estimation . . . . . . . . . . . 82 8.2.1 Bayesian estimation: somewhere between calibration and maximum likelihood estimation  an example . . . . . . 84 8.3 DSGE models and Bayesian estimation . . . . . . . . . . . . . . 85 8.3.1 Rewriting the solution to the DSGE model . . . . . . . 85 8.3.2 Estimating the likelihood function of the DSGE model . 86 8.3.3 Finding the mode of the posterior distribution . . . . . 87 8.3.4 Estimating the posterior distribution . . . . . . . . . . . 87 8.4 Comparing models using posterior distributions . . . . . . . . . 90 9 Optimal policy under commitment 10 Troubleshooting vi 93 95
. a bird’s eyeview . . . . . . . . . . . . . . . . . . . .1 Dynare. . . . . . . . . . . . Structure of the . . . .1 6. .1 6. . . . . . . . .mod ﬁle . . . . . . . . . . . . . . . . . . . . . . . Steps of model estimation . . . . . . Illustration of the MetropolisHastings algorithm . . . . . . .1 3. . . . . . 3 16 62 74 89 . .List of Figures Bibliography vii 97 List of Figures 1. . . . CIA model illustration . . . . . . . . . . . . . . .2 8. .
.
our readers . Please write either direclty to myself: tommaso. please read this with a critical eye and send me comments! Are some areas unclear? Is anything plain wrong? Are some sections too wordy. are there certain parts that just click particularly well? How can others be improved? I’m very interested to get your feedback. Any such notes are marked with two stars (**).Work in Progress! This is the second version of the Dynare User Guide which is still work in progress! This means two things.org. First. These are mostly placeholders for future work. are these clear? On the contrary. ix . are there enough examples.mancini@stanfordalumni. Thanks very much for your patience and good ideas. The second thing that a work in progress manuscript comes with is a few internal notes. or preferably on the Dynare Documentation Forum available in the Dynare Forums. notes to myself or others of the Dynare development team. or at times notes to you .to highlight a feature not yet fully stable.
.
juillard“AT”mjui. Klein. almost no question is speciﬁc enough to interest just one person.ratto“AT”jrc. S.adjemian“AT”ens. Schorfheide.fr) e • Houtan Bastani (houtan.uk) • Marco Ratto (marco. P. C. Kamenik. the development team of Dynare is composed of • St´phane Adjemian (stephane.Contacts and Credits Dynare was originally developed by Michel Juillard in Paris.co. The email addresses above are provided in case you wish to contact any one of the authors of Dynare directly. Currently. F. We nonetheless encourage you to ﬁrst use the Dynare forums to ask your questions so that other users can beneﬁt from them as well. Sakata. Sims. Soederlind and R. Wouters. Ingber. L.bastani“AT”ens. remember. O. F. and yours is not the exception! xi .eu) • S´bastien Villemot (sebastien.fr) e Several parts of Dynare use or have strongly beneﬁted from publicly available programs by G.europa. Finally.orangehome.fr) • Michel Juillard (michel. Anderson.villemot“AT”ens.fr) • Ferhat Mihoubi (ferhat. France. P.fr) • George Perendia (george“AT”perendia. Collard.ec. the development of Dynare could not have come such a long ways withough an active community of users who continually pose questions. report bugs and suggest new features.mihoubi“AT”univevry. The help of this community is gratefully acknowledged.
.
The sophisticated computer programmer. may not ﬁnd this Guide suﬃciently detailed. This Guide is written mainly for an advanced economist .approach and structure This User Guide aims to help you master Dynare’s main functionalities. this Guide is structured around examples and oﬀers practical advice. To do 1 .Chapter 1 Introduction Welcome to Dynare! 1. we recommend starting with chapters 3 and 5. from getting started to implementing advanced features. on the other. though. on the one hand. The idea is to get you to use 90% of the program well and then tell you where else to look if you’re interested in ﬁne tuning or advanced customization. which introduce the program’s basic features to solve (including running impulse response functions) and estimate DSGE models. this Guide also gives some background on Dynare’s algorithms. This Guide will focus on the most common or useful features of the program. methodologies and underlying theory. Thus.needing a powerful and ﬂexible program to support and facilitate his or her research activities in a variety of ﬁelds.like a professor. This Guide is written to accommodate both. If you’re new to Dynare. respectively. We recognize that the “advanced economist” may be either a beginning or intermediate user of Dynare. To do so.1 About this Guide . To root this understanding more deeply. a secondary function of this Guide is to serve as a basic primer on DSGE model solving and Bayesian estimation. graduate student or central banker . thus emphasizing depth over breadth. or the specialist of computational economics.
you will know the crux of Dynare’s functionality and (hopefully!) feel comfortable using Dynare for your own work. We also recognize that you probably have had repeated if not active exposure to programming and are likely to have a strong economic background. a black box solution to your needs is inadequate. besides breaking up content into short chapters. we’ve introduced two diﬀerent markers throughout the Guide to help streamline your reading. let’s have a look at the “forest” from the top . These chapters cover more advanced features of Dynare and more complicated usage scenarios. Once you have read these two chapters. these chapters read a bit more like a reference manual. though. These are available in the “behind the scenes of Dynare” chapters 7 and 8.2 What is Dynare? Before we dive into the thick of the “trees”. To hopefully address this issue. more appropriate. . to solve. • TIP! introduces advice to help you work more eﬃciently with Dynare or solve common problems. . in order to “learn by doing”. these chapters lead you through a complete handson example. 4 and 6. Thus. INTRODUCTION so. 1.2 CHAPTER 1. If you’re instead an intermediate user of Dynare. with an intuitive frontend interface. simulate and estimate DSGE models. At that point. . the User Guide goes into some depth in covering the theoretical underpinnings and methodologies that Dynare follows to solve and estimate DSGE models. just what is Dynare? Dynare is a powerful and highly customizable engine. • NOTE! is used to draw your attention to particularly important information you should keep in mind when using Dynare. These chapters can also serve as a basic primer if you are new to the practice of DSGE model solving and Bayesian estimation. The presumption is that you would skip around these chapters to focus on the topics most applicable to your needs and curiosity. which we recommend following from A to Z. Examples are therefore more concise and speciﬁc to each feature. you will probably ﬁnd yourself coming back to the User Guide to skim over some of the content in the advanced chapters to iron out details and potential complications you may run into. you will most likely ﬁnd the advanced chapters. Finally.
This initiates the Dynare preprocessor which translates the .mod ﬁle being read by the Dynare preprocessor. Some more details on the internal ﬁles generated by Dynare is given in section 4.2 in chapter 4. is written equation by equation in an editor of your choice. the model and its related attributes.1: The . but also enables you to easily share your code as it is straightforward to read by anyone. WHAT IS DYNARE? 3 Figure 1. This not only facilitates the inputting of a model. it is a preprocessor and a collection of Matlab routines that has the great advantages of reading DSGE model equations written almost as in an academic paper.2. Finally.mod ﬁle into a suitable input for the Matlab routines (more precisely. it creates intermediary Matlab or C ﬁles which are then used by Matlab code) used to either solve or estimate the model. but for now it may be helpful to summarize what Dynare is able to do: . like a shock structure for instance.mod ﬁle. Each of these steps will become clear as you read through the User Guide. which then calls the relevant Matlab routines to carry out the desired operations and display the results. Basically.2 gives you an overview of the way Dynare works. The resulting ﬁle will be called the .1. Figure 1. results are presented in Matlab. That ﬁle is then called from Matlab. In slightly less ﬂowery words.
usually well documented . Besides allowing you to stay up to date with the most recent papers and possibly make new contacts. so reading corresponding command descriptions in the Reference Manual is a good idea to cover all relevant details. • Dynare forums: this lively online discussion forum allows you to ask your questions openly and read threads from others who might have run into similar diﬃculties. giving a clear deﬁnition and explanation of usage for each. • Frequently Asked Questions (FAQ): this section of the Dynare site emphasizes a few of the most popular questions in the forums. workshops and seminars that may be of interest. • Oﬃcial online examples: the Dynare website includes other examples . The User Guide will often introduce you to a command in a rather loose manner (mainly through examples). and yearn to ask that one question whose answer seems to exist nowhere.mod ﬁles covering models and methodologies introduced in recent papers.3 Additional sources of help While this User Guide tries to be as complete and thorough as possible. it conveniently lists conferences. as you learn about new features. you have the following additional sources of help: • Reference Manual: this manual covers all Dynare commands.of . run my members of the Dynare team. struggle with adapting examples to your own work.4 • compute the steady state of a model CHAPTER 1.net: this website. At your disposal. • DSGE. you will certainly want to browse other material for help. INTRODUCTION • compute the solution of deterministic models • compute the ﬁrst and second order approximation to solutions of stochastic models • estimate parameters of DSGE models using either a maximum likelihood or a Bayesian approach • compute optimal policies in linearquadratic models 1. . is a resource for all scholars working in the ﬁeld of DSGE modeling.
this new version introduces several important features. • Variable name indicates a variable name. NOMENCLATURE 5 1.g. 1.1E3. the option must be speciﬁed in parenthesis in Dynare. All other characters. • Options or optional arguments for a command are listed in square brackets [ ] unless otherwise noted. The following syntaxes are valid: 1.1e3. • Typewritten text indicates text as it should appear in Dynare code. what’s new and backward compatibility The current version of Dynare .5 v4. it is worthwhile to agree on a few deﬁnitions of terms.1d3. for instance. this concerns your Dynare .1E3. With respect to version 3.mod ﬁles. This increases computational speed signiﬁcantly.for which this guide is written .4. NOTE! These must start with an alphabetical character and can only contain other alphabetical characters and digits. 1. • Filename indicates a ﬁle name valid in your operating system. . • Integer indicates an integer number. Many of these are shared with the Reference Manual.1D3 • Expression indicates a mathematical expression valid in the underlying language (e. The drawback is that Dynare can now handle only a limited set of functions. although in nearly all economic applications this should not be a constraint. Note that Matlab requires that names of ﬁles or functions start with alphabetical characters. Matlab).4 Nomenclature To end this introduction and avoid confusion in what follows. in the Newton algorithm for deterministic models and in the linearizations necessary to solve stochastic models).1. • Command is an instruction to Dynare or other program when speciﬁed. 1. 1.is version 4. are forbidden. 1. • Double indicates a double precision number. If. as well as underscores ( ). including accents. optimizations of routines and bug ﬁxes. as well as improvements. • Parameter name indicates a parameter name which must follow the same naming conventions as above. it will show up in the Guide as [(option)]. and spaces. The major new features are the following: • Analytical derivatives are now used everywhere (for instance.
• Speed. This is covered in more details in chapter 3. . in section 3. but also apparent in other functionality. variables and parameters where at times in their order of declaration and at times in alphabetical order. Several largescale improvements have been implemented to speed up Dynare. The names of the ﬁles internally generated by Dynare have also changed. NOTE! You will unfortunately have to slightly amend any old steady state ﬁles you may have written. (** more on this when explaining internal ﬁle structure TBD) • The syntax for the external steady state ﬁle has changed. This should be most noticeable when solving deterministic models.6 CHAPTER 1. These are enumerated in details in the relevant chapters. Recall that in version 3.6. • The names of many internal variables and the organization of output variables has changed. INTRODUCTION • Variables and parameters are now kept in the order in which they are declared whenever displayed and when used internally by Dynare. NOTE! This may cause some problems of backward compatibility if you wrote programs to run oﬀ Dynare v3 output.3.
Note that Dynare++ is distributed along with Dynare since version 4. There used to be versions of Dynare for Scilab and Gauss.Chapter 2 Installing Dynare 2. Depending on the type of computations required. 7 .02 and that for Gauss after Dynare version 1. as well as Unixlike operating systems. which is a standalone C++ program specialized in computing korder approximations of dynamic stochastic general equilibrium models. Development of the Scilab version stopped after Dynare version 3.1 Dynare versions The current version of Dynare (4. Dynare++.1) runs on both MATLAB and GNU Octave.0 and later. although 512MB is preferred. you may need up to 1GB of RAM to obtain acceptable computational times. 2. in particular GNU/Linux and Mac OS X. please ask your question on Dynare forums.2. This User Guide will exclusively focus on Dynare version 4. like the very processor intensive Metropolis Hastings algorithm. 2. See the Dynare++ webpage for more information. You may also be interested by another program.3 Installing Dynare Please refer to the section entitled “Installation and conﬁguration” in the Dynare reference manual.2 System requirements Dynare can run on Microsoft Windows.1. it is recommended to allocate at least 256MB of RAM to the platform running Dynare. To run Dynare. If you have questions about the support of a particular platform.
no additional toolbox is necessary for running most of Dynare. INSTALLING DYNARE 2. both of which are deﬁned later. you will have additional options for solving for the steady state (solve algo option) and for searching for the posterior mode (mode compute option). except maybe for optimal simple rules (see chapter 9). . but even then remedies exist (see the Dynare forums for discussions on this.4 MATLAB particularities A question often comes up: what special MATLAB toolboxes are necessary to run Dynare? In fact. or to ask your particular question).8 CHAPTER 2. But if you do have the ‘optimization toolbox’ installed.
in fact.mod ﬁle . we recommend that you read this chapter chronologically. That is to say that the term “solution” in the title of the chapter is used rather broadly. it is so fundamental. This chapter covers all these topics. Let’s consider a shock to a model’s innovation only in period 1. it is important to recognize a distinction in model types. here is the important question: is your model stochastic or deterministic? The distinction hinges on whether future shocks are known. but you may also want to go a bit further.1 A fundamental distinction Before speaking of Dynare. In a deterministic context.Dynare commands and syntax . Likewise. But the amount of common material . This distinction will appear throughout the chapter. you may want to explore how the system comes back to its steady state or moves to a new one. you may be interested in how this system behaves in response to shocks. to learn basic Dynare commands and the process of writing a proper .Chapter 3 Solving DSGE models . whether temporary or permanent. In stochastic models. that we considered writing separate chapters altogether. 3. agents will take their decisions knowing that future values of the innovations will be zero in all periods to come. and stems from. agents will take their decisions 9 . But instead of skipping to the topic closest to your needs. Typically. only the distribution of future shocks is known. a vast terrain. instead.basics This chapter covers everything that leads to. Enough suspense. In deterministic models. You may be interested in simply ﬁnding the solution functions to a set of ﬁrst order conditions stemming from your model.this will serve as a base to carry out any of the above computations. In a stochastic context. the occurrence of all future shocks is known exactly at the time of computing the model’s solution.is notable and writing two chapters would have been overly repetitive. the solution of DSGE models.
“stochastic”.at the time of making her decision . the two cases become practically the same. deterministic models have become somewhat rare. 2.1 NOTE! Deterministic vs stochastic models Deterministic models have the following characteristics: 1. as in the introduction of a new tax. though. the following gives some additional details. 4. if you consider only a ﬁrst order linear approximation of the stochastic model. instead. This isn’t the same thing because of Jensen’s inequality. The solution is nothing more than a series of numbers that match a given set of equations. as the variance of shocks will matter. Shocks can hit the economy today or at any time in the future. a highly accurate solution can be found by numerical methods. Of course. Models assume full information. In this case. i. due to certainty equivalence. not deterministic!) literature has gained attention in economics. The solution method for each of these model types diﬀers signiﬁcantly. solutions to deterministic models are usually called “closed loop” solutions. . In deterministic models. 3. models introduce a positive shock today and zero shocks thereafter (with certainty). this function may be nonlinear and thus needs to be approximated. policy or feedback rule for the future: what will her optimal actions be contingent on each possible realization of shocks.1. These models are usually introduced to study the impact of a change in regime. In a stochastic environment. or a linear model. In control theory. Examples include OLG models without aggregate uncertainty. perfect foresight and no uncertainty around shocks.e. for instance. Because this distinction will resurface again and again throughout the chapter. 3. SOLVING DSGE MODELS .BASICS knowing that the future value of innovations are random but will have zero mean. Most often. Intuitively. and those to stochastic models are referred to as “open loop”. she can specify today . in which case they would be expected with perfect foresight.what each of her precise actions will be in the future. 5. if an agent has perfect foresight. As the DSGE (read.10 CHAPTER 3. we therefore search for a function satisfying the model’s ﬁrst order conditions. To complicate things. They can also last one or several periods. A second order approximation will instead lead to very diﬀerent results. but also because it has been a source of signiﬁcant confusion in the past. the best the agent can do is specify a decision.
you’ll have to use a bit of imagination (on top of that needed to think you have perfect foresight!) to .2 Introducing an example The goal of this ﬁrst section is to introduce a simple example. or new keynesian monetary models. The model introduced here is a basic RBC model with monopolistic competition. More complex examples are instead presented in the advanced chapters. Examples include most RBC models. In these models. Thus. 3 and 5. used widely in the literature. We will nonetheless illustrate both the stochastic and the deterministic settings on the basis of this example. 2. 7. it doesn’t even really need a steady state. This is an often overlooked point in the literature which misleads readers in supposing their models may be deterministic. Note that as a general rule. Its particular notation adopted below is drawn mostly from notes available on Jesus FernandezVillaverde’s very instructive website. this is a good place to look for additional information on any of the following model setup and discussion. in fact. These types of models tend to be more popular in the literature. Future sections will aim to code this example into Dynare and analyze its salient features under the inﬂuence of shocks . This solution method can therefore be useful when the economy is far away from steady state (when linearization oﬀers a poor approximation). with just enough features to help illustrate Dynare commands and functionalities. The solution does not require linearization. but thereafter their expected value is zero. it involves numerical simulation to ﬁnd the exact paths of endogenous variables that meet the model’s ﬁrst order conditions and shock structure.3. which is the certainty equivalence property. are kept as bare as possible.both in a stochastic and a deterministic environment. the examples in the basic chapters. Note throughout this model description that the use of expectation signs is really only relevant in a stochastic setting. or permanent changes in the exogenous variables cannot be handled due to the use of Taylor approximations around a steady state. agents behave as if future shocks where equal to zero (since their expectation is null). Stochastic models. INTRODUCING AN EXAMPLE 11 6. instead. Instead. 3. Note that when these models are linearized to the ﬁrst order. have the following characteristics: 1.2. 3. Expected future shocks. as per the earlier discussion. shocks hit today (with a surprise). when thinking of the latter.
capturing the intertemporal tradeoﬀ mentioned above. with a little more imagination. but is presented below in the simplest possible terms. the more leisure due to the decreasing marginal utility of consumption). with a little handwaiving involved. we obtain the intuitive result that it = kt+1 − (1 − δ)kt . the consumer therefore faces a tradeoﬀ between consuming and investing in order to increase the capital stock and consuming more in following periods (as we will see later. according to the following utility function ∞ Et t=0 β [log ct + ψ log(1 − lt )] and subject to the following budget constraint ct + kt+1 = wt lt + rt kt + (1 − δ)kt . and the labor supply equation linking labor positively to wages and negatively to consumption (the wealthier. Alternatively. 1−lt . The above equation can be seen as an accounting identity. Households maximize utility over consumption. by the zero proﬁt condition. total payments to factors. leisure and capital stock.BASICS ignore the expectation signs. ∀t > 0 where kt is capital stock. where lt is labor input. As a consequence. rt real interest rates or cost of capital and δ the depreciation rate. if we deﬁne investment as it = yt − ct . as the derivations are relatively standard. In any given period. yields the Euler equation in consumption. wt real wages. Maximization of the household problem with respect to consumption. equals yt . due to monopolistic competition. These equation are 1 1 = βEt (1 + rt+1 − δ) ct ct+1 and ψ ct = wt 1 − lt The ﬁrm side of the problem is slightly more involved. or aggregate output.12 CHAPTER 3.including the liquidation value of the capital stock . production depends on capital). or that investment replenishes the capital stock thereby countering the eﬀects of depreciation. the equation can also be interpreted as a capital accumulation equation after bringing ct to the right hand side and noticing that wt lt + rt kt .on the right hand side. . ct and leisure. SOLVING DSGE MODELS . with total expenditures on the left hand side and revenues .
face a two pronged decision: how much labor and capital to employ given these factors’ perfectly competitive prices and how to price the variety they produce. zt captures technology which evolves according to zt = ρzt−1 + et where ρ is a parameter capturing the persistence of technological progress and et ∼ N (0.3. The solution to the sourcing problem yields an optimal capital to labor ratio. we therefore have: mct = ( − 1)/ But what are marginal costs equal to? To ﬁnd the answer. An additional step simpliﬁes this expression: symmetric ﬁrms implies that all ﬁrms charge the same price and thus pit = pt . we combine the optimal capital to labor ratio into the production function and take advantage of its CRS property to solve for the amount of labor or capital required to produce one unit of output. mct is real marginal cost and pt is the aggregate CES price or average price. instead.2. If we follow the second route. yielding the DixitStiglitz downward sloping demand curve. σ). Intermediate producers. Also. the ﬁnal goods producer chooses his or her optimal demand for each variety. INTRODUCING AN EXAMPLE 13 There are two ways to introduce monopolistic competition. or relationship between payments to factors: kit rt = α wt lit 1−α The solution to the pricing problem. Production of intermediate goods follows a CRS production function deﬁned as α yit = kit (ezt lit )1−α where the i subscript stands for ﬁrm i of a continuum of ﬁrms between zero and one and where α is the capital elasticity in the production function. instead. with 0 < α < 1. The real cost of using this amount of any one . We can either assume that ﬁrms sell diﬀerentiated varieties of a good to consumers who aggregate these according to a CES index. yields the wellknown constant markup pricing condition of monopolistic competition: pit = −1 mct pt where pit is ﬁrm i’s speciﬁc price. Or we can postulate that there is a continuum of intermediate producers with market power who each sell a diﬀerent variety to a competitive ﬁnal goods producer whose production function is a CES aggregate of intermediate varieties.
we have the DixitStiglitz demand for each variety. or wt ∂yit . which is the deﬁnition it of marginal cost: the cost in terms of labor input of producing an additional unit of output. SOLVING DSGE MODELS . On the other side. . to yield wt [(1 − α)yit /lit ]−1 . This should not be a surprise since the optimal capital to labor ratio follows from the maximization of the production function (minus real costs) with respect to its factors. Interestingly. in turn. we can rewrite the above two equations for wt and rt without the i subscripts on the right hand side. we obtain mct = 1 1−α 1−α 1 α α 1 1−α α w rt At t which does not depend on i. pit = pt . we aggregate the production of each individual ﬁrm to ﬁnd an aggregate production function. Combining this result for marginal cost. which is the same for all ﬁrms and thus does not depend on i. we factor out the capital to labor ratio. that because the ratio of output to each factor is the same for each intermediate ﬁrm and that ﬁrm speciﬁc as well as aggregate production is CRS. itself equal to household consumption. equal to the aggregate output of ﬁnal good.14 CHAPTER 3.we obtain aggregate production rt = α α 1−α yt = At kt lt which can be shown is equal to the aggregate amount of varieties bought by the ﬁnal good producer (according to a CES aggregation index) and. kt /lt . as well as its counterpart in terms of capital. with the optimal pricing condition yields the ﬁnal two important equations of our model wt = (1 − α) and yit ( − 1) lit yit ( − 1) kit To end. as hinted earlier. This ends the exposition of the example. for instance. On the supply side. Note. and noting that price dispersion is null . it is thus the same for all ﬁrms. When solving for labor. Now. let’s roll up our sleeves and see how we can input the model into Dynare and actually test how the model will respond to shocks. By equating the two and integrating both side. by using the optimal capital ∂l to labor ratio. the above can be worked out. to close.or that.BASICS factor is given by wt lit + rt kit where we substitute out the payments to the other factor using again the optimal capital to labor ratio.
3. we lay these out.g.mod ﬁle.mod.mod is not necessary). • varexo starts the list of exogenous variables that will be shocked. as mentioned loosely in the introduction of this Guide.mod ﬁle. The commands are: • var starts the list of endogenous variables.mod ﬁle as containing four distinct blocks. forecasting. or the starting point for simulations or impulse response functions based on the model’s solution. then we discuss them. 3. let’s diﬀerentiate between the stochastic and deterministic cases.mod ﬁle is stored and then by typing in the Matlab command line Dynare filename.3.MOD FILE STRUCTURE 15 3.4 Filling out the preamble The preamble generally involves three commands that tell Dynare what are the model’s variables. But before we get into executing a . It will then be read by Matlab by ﬁrst navigating within Matlab to the directory where the . In the case of our example.3: • preamble: lists variables and parameters • model: spells out the model • steady state or initial value: gives indications to ﬁnd the steady state of a model. The . estimating impulse response functions) Our exposition below will structured according to each of these blocks. which are endogenous and what are the parameters. DYNARE .mod ﬁle can be written in any editor. • parameters starts the list of parameters and assigns values to each. First. • shocks: deﬁnes the shocks to the system • computation: instructs Dynare to undertake speciﬁc operations (e. . illustrated in ﬁgure 3. external or internal to Matlab. to be separated by commas. let’s start by writing one! It is convenient to think of the .3 Dynare . (although actually typing the extension .mod ﬁle structure Input into Dynare involves the .
except that we no longer need the et variable. epsilon = 10.mod ﬁle contains ﬁve logically distinct parts. Thus.023. With respect to the above.007/(1alpha)).1 The deterministic case The model is inherited exactly as speciﬁed in the earlier description. alpha = 0. et .BASICS Figure 3. parameters beta psi delta alpha sigma epsilon. beta = 0. psi = 1.33.4.16 CHAPTER 3. delta = 0.2 The stochastic case In this case. varexo z. 3. sigma = (0.4. consisting of an exogenous shock.1: The . we therefore . SOLVING DSGE MODELS .75. 3. as we can make zt directly exogenous. the preamble would look like: var y c k i l y l w r. we go back to considering the law of motion for technology.99.
mod ﬁle must be terminated by a semicolon (. Here’s what the preamble would look like: var y c k i l y l w r z.). . parameters beta psi delta alpha rho sigma epsilon.75. sigma = (0. rho = 0. although a single instruction can span two lines if you need extra space (just don’t put a semicolon at the end of the ﬁrst line). psi = 1.007/(1alpha)). parameters beta psi delta alpha rho sigma epsilon.99. beta = 0. alpha = 0.33.75.4. and add the parameter ρ.95.95. Ask her all the difficult questions! */ alpha = 0. For example: var y c k i l y l w r z. TIP! You can also comment out any line by starting the line with two forward slashes (//). delta = 0.99. varexo e.3 Comments on your ﬁrst lines of Dynare code As you can tell. or comment out an entire section by starting the section with /* and ending with */. beta = 0.4.33. delta = 0.023.3. rho = 0. FILLING OUT THE PREAMBLE 17 adjust the list of endogenous and exogenous variables. psi = 1. epsilon = 10.mod ﬁle is really quite straightforward. writing a . 3.023. // the above instruction reads over two lines /* the following section lists several parameters which were calibrated by my coauthor. varexo e. Two quick comments: NOTE! Remember that each instruction of the .
SOLVING DSGE MODELS . end. The ﬁfth and sixth are the marginal cost equal to markup equations. The eighth an identity that may be useful and the last the equation of motion of technology. The fourth is the production function. notice that the law of motion for technology is included. c+i = y. The third the accounting identity. NOTE! that the above model speciﬁcation corresponds to the stochastic case. What you can already try to do is glance at the model block below and see if you can recognize the equations from the earlier example.5 3.5. y l = y/l. almost as if you were writing them in an academic paper.18 CHAPTER 3. z = rho*z(1)+e. epsilon = 10.007/(1alpha)). . See how easy it is to read Dynare code? model. Let’s ﬁrst have a look at our model in Dynare notation.1 Specifying the model Model in Dynare notation One of the beauties of Dynare is that you can input your model’s equations naturally. as your colleagues will be able to understand your code in notime. 3. here’s a brief description: the ﬁrst equation is the Euler equation in consumption. y = (k(1)^alpha)*(exp(z)*l)^(1alpha). r = y*((epsilon1)/epsilon)*alpha/k(1). indeed. The seventh is the investment equality. Just in case you need a hint or two to recognize these equations. w = y*((epsilon1)/epsilon)*(1alpha)/l. There are just a few conventions to follow. The corresponding model for the deterministic casce would simply loose the last equation. (1/c) = beta*(1/c(+1))*(1+r(+1)delta).BASICS sigma = (0. The second the labor supply function. psi*c/(1l) = w. i = k(1delta)*k(1). as per our discussion of the preamble. and then go through the various Dynare input conventions. This greatly facilitates the sharing of your Dynare ﬁles.
it is what we call in the jargon a predetermined variable. • Fourth.5. TIP! remember that variable and parameter names are case sensitive. 3. as in many papers.3.mod ﬁle. 3. • Second. our capital stock is not decided today. • The ﬁrst thing to notice.5.mod ﬁle begins with the command model and ends with the command end. • Variables entering the system with a time t − n subscript are written with (−n) following them. xt would be written x. eventhough in the example presented above we wrote kt+1 = it + (1 − δ)kt . ). For example.2 General conventions The above example illustrates the use of a few important commands and conventions to translate a model into a Dynarereadable . This is unlike Matlab where if you break a line you need to add . .mod ﬁle. • Third. but yesterday (recall that it is a function of yesterday’s investment and capital stock). there need to be as many equations as you declared endogenous variables (this is actually one of the ﬁrst things that Dynare checks. . xt+2 would be written x(+2). Note that variable and parameter names used in the model block must be the same as those declared in the preamble. equations are entered one after the other. For example. this would count as two backward looking variables). each line of instruction ends with a semicolon (except when a line is too long and you want to break it across two lines. . Writing x(2) is also allowed. no matrix representation is necessary. as in the preamble and everywhere along the . xt−2 would be written x(−2) (incidentally. . • In the same way.5. For example. variables entering the system with a time t+n subscript are written with (+n) following them. Thus. the timing of each variable reﬂects when that variable is decided. SPECIFYING THE MODEL 19 3. For instance. in between.5.3 Notational conventions • Variables entering the system with a time t subscript are written plainly. is that the model block of the . . but this notation makes it slightly harder to count by hand the number of forward looking variables (a useful measure to check).4 Timing conventions • In Dynare. it will immediately let you know if there are any problems). more on this below . we would translate this equation into Dynare as k=i+(1delta)*k(1).
3.20 CHAPTER 3. Otherwise. the model in explogs. wages should appear with a one period lag. • A slightly more roundabout way to explain the same thing is that for stock variables. the second order diﬀerence equation would have two eigenvalues. consider that in some wage negociation models. a lot of papers use the “stock at the beginning of the period” convention. would look like: model (linear). If this condition is not met. • BlanchardKahn conditions are met only if the number of nonpredetermined variables equals the number of eigenvalues greater than one. It is investment during period t that sets stock at the end of period t. Thus. where repeating a letter for a variable means diﬀerence from steady state.ll.5. one needing to be greater and the other smaller than one for stability. you can write wage in period t (when they are set). end. all that is necessary is to write the term (linear) next to the command model. if you had one. in the equation for wages. yy l=yy . wages used during a period are set the period before. the linear model and second. SOLVING DSGE MODELS . but in the labor demand equation. In the ﬁrst case.5 Conventions specifying nonpredetermined variables • A (+1) next to a variable tells Dynare to count the occurrence of that variable as a jumper or forwardlooking or nonpredetermined variable. 3. Dynare will put up a warning. In this case. For instance. you must use a “stock at the end of the period” concept. • Note that a variable may occur both as predetermined and nonpredetermined. with just the equation for yl for illustration. this turns out to be a very useful option . consumption could appear with a lead in the Euler equation. First.6 Linear and loglinearized models There are two other variants of the system’s equations which Dynare accommodates. you may be interested to have Dynare take Taylor series expansions in logs rather than in levels. Our example. but also with a lag in a habit formation equation. Be careful.BASICS • As another example.5. as we did (on purpose to highlight this distinction!) in the setup of the example model above.
6 Specifying steady states and/or initial values Material in this section has created much confusion in the past. so that the level of a variable is given by exp(repeatedvariable). simply rewrite your equations by taking the exponential and logarithm of each variable. repeating a letter for a variable means log of that variable. or approximations of values. this time. Let’s see in more details how all this works. The ﬁrst two are instead covered in what follows. In passing. You can either enter exact steady state values into your . One of the functions of this section is indeed to provide these steady state values.6. recall that stochastic models need to be linearized. note that the relevant commands in this section are initval.1 Stochastic models and steady states In a stochastic setting. where. Thus. your model will need to be linearized before it is solved. This section is also useful to specify this starting value.mod ﬁle. though. are provided in section 3. they need to have a steady state. Dynare needs to know your model’s steady state (more details on ﬁnding a steady state. SPECIFYING STEADY STATES AND/OR INITIAL VALUES 21 when estimating models with unit roots.6. First. The Dynare input convention makes this very easy to do. In either case. 3. psi*exp(cc)/(1exp(ll)) = exp(ww).6.3. these values are entered in the initval block. endval or. But with some attention to the explanations below. more rarely. you should get through unscathed. as in the following fashion: initval.3 below). you may be interested to start your simulations or impulse response functions from either a steady state. .mod ﬁle. or another given point. 3. If so. end. or just approximations and let Dynare ﬁnd the exact steady state (which it will do using numerical methods based on your approximations). To do so. Second. Our example would need to be rewritten as follows (just shown for the ﬁrst two equations) model. as we will see in chapter 5. Let’s start by emphasizing the uses of this section of the . as well as tips to do so more eﬃciently. irrespective of whether you’re working with a stochastic or deterministic model. histval which is covered only in the Reference Manual. (1/exp(cc)) = beta*(1/exp(cc(+1)))*(1+exp(rr(+1))delta).
For the case in which you would like simulations and impulse response functions to begin at the steady state. e = 0. this means either using the command steady or entering exact steady state values. TIP! If you’re dealing with a stochastic model.3. k = 9. by using the command steady. steady. l = 0.3. or from the exact values you speciﬁed in the initval block. CHAPTER 3. remember that its linear approximation is good only in the vicinity of the steady state. l = 0. end. r = 0. if you don’t add the command steady. SOLVING DSGE MODELS .0. even if Dynare will have calculated your model’s exact steady state for the purpose of linearization. w = 2. end.7. c = 0. c = 0.7. z = 0. On the contrary. your simulations or impulse response functions will start from your initial values.BASICS Then. . the above block would be expanded to yield: initval. Adding steady just after your initval block will instruct Dynare to consider your initial values as mere approximations and start simulations or impulse response functions from the exact steady state. r = 0.22 k = 9. w = 2. e = 0. you can control whether you want to start your simulations or impulse response functions from the steady state. thus it is strongly recommended that you start your simulations from a steady state. z = 0.0.
is calculating actual steady state values. • solve algo = 3: uses the Sims solver. But even for simpler models. you would enter those values in your initval block and not use the steady command. Thus. 3. If you have trouble ﬁnding the steady state of your model. of course. and luck is unfortunately part of the equation. you do not need to provide a steady state for these model. your . most researchers are still interested to see how a model reacts to shocks when originally in steady state.6. the initval block serves very similar functions as described above. you can begin by playing with the options following the steady command. you would enter approximate (or exact) steady state values in the initval block. But it is usually only successful if the initial values you entered are close to the true steady state. followed by the command steady. you may still run into diﬃculties in ﬁnding your steady state.6. These are: • solve algo = 0: uses Matlab Optimization Toolbox FSOLVE • solve algo = 1: uses Dynare’s own nonlinear equation solver • solve algo = 2: splits the model into recursive blocks and solves each block in turn. Thus. Otherwise. But practically. As mentioned above. In this case. If you wanted to shock your model starting from a steady state value. Dynare can help in ﬁnding your model’s steady state by calling the appropriate Matlab functions.6. the following TIPS! may help. their initial values would all be zero. it is better to start with a smaller model and add new variables one by one. SPECIFYING STEADY STATES AND/OR INITIAL VALUES 23 3.3 Finding a steady state The diﬃculty in the above. another option is to enter your model in linear terms. Yet. An illustration of the initval block in the deterministic case appears further below. Often.2 Deterministic models and initial values Deterministic models do not need to be linearized in order to be solved. ﬁnding suitable initial values for the endogenous variables is the trickiest part of ﬁnding the equilibrium of that model.3. if any of your original (nonlinear) equations involve sums (a likely fact). This is the default option if none are speciﬁed. Doing so borders on a form of art. technically. In the deterministic case. Unfortunately. For complicated models. If so. variables would be expressed in percent deviations from steady state. if you wanted to begin your solution path from an arbitrary point.
SOLVING DSGE MODELS . which you would still need to calculate. If this condition is not . As mentioned earlier. For instance. You will instead need to write your steady state program as if you were solving for the steady state by hand. you may be left needing to calculate fewer steady state values than in the original.mod ﬁle to see if such a Matlab ﬁle exists. This computes and displays the eigenvalues of your system which are used in the solution method.mod ﬁle and ask Matlab to solve the system.mod ﬁle.mod ﬁle so that running loops with diﬀerent parameter values. followed by steadystate For instance. model. would be: (** example ﬁle to be added shortly) 3. your matlab (. it will use that ﬁle to ﬁnd steady state values regardless of whether you’ve provided initial values in your .4 Checking system stability TIP! A handy command that you can add after the initval or endval block (following the steady command if you decide to add one) is the check command. The alternative is to write a Matlab program to ﬁnd your model’s steady state.mod ﬁle.mod ﬁle. you could write an external Maple ﬁle and then enter the steady state solution by hand in Dynare. though (unless you’re working with a particular toolbox). For example. Alternatively.m and should be saved in the same directory as your .m) ﬁle should have the same name as your .6. That is. this procedure could be time consuming and bothersome. for instance. the steady state ﬁle corresponding to the above example. you could also use an external program to calculate exact steady state values. If so.mod ﬁle is called example. It is not enough to simply input the equations as you’ve written them in your . a necessary condition for the uniqueness of a stable equilibrium in the neighborhood of the steady state is that there are as many eigenvalues larger than one in modulus as there are forward looking variables in the system. But of course.BASICS linearized equations will include ratios of steady state values. your Matlab ﬁle should be called example steadystate. Because Matlab does not work with analytical expressions. nonlinear. if your . Dynare will automatically check the directory where you’ve saved your .mod. whereby each lefthand side variable is written in terms of known parameters or variables already solved in the lines above. Yet. NOTE! When doing so. in the stochastic case.24 CHAPTER 3. you need to do a little work to write your steady state program. especially if you want to alter parameter values (and thus steady states) to undertake robustness checks. Doing so has the clear advantages of being able to incorporate your Matlab program directly into your . becomes seamless. you need to input your expressions sequentially.
7.7. as explained in our original discussion on stochastic and deterministic models.permanent shocks To study the eﬀects of a permanent shock hitting the economy today. var z. The corresponding instructions would be: . In both cases. you are free to set the duration and level of the shock. Note that you can also use the mshocks command which multiplies the initial value of an exogenous variable by the mshocks value. values 0. The distinction is that under a temporary shock. you would write: shocks. If variables were in logs. the model eventually comes back to steady state. For instance. Finally. such as a structural change in your model. except for the value of technology which you may presume changes permanently. you have the choice of introducing both temporary and permanent shocks. while under a permanent shock. though.7. you would use the endval block following the usual initval block.temporary shocks When working with a deterministic model. ADDING SHOCKS 25 met. this would have corresponded to a 10% shock. you may specify all values to remain common between the two blocks. 3. but would simply tell the system to which (steady state) values you would like it to move and let Dynare calculate the transition path. the shocks are entirely expected. Given the above instructions. end. in order to study the anticipatory behavior of agents in response to future shocks.7 3. Dynare would replace the value of zt speciﬁed in the initval block with the value of 0.1 entered above. the model reaches a new steady state. To do so.1 Adding shocks Deterministic models .1. To specify a shock that lasts 9 periods on zt . for instance. you would not specify actual “shocks”.3. Dynare will tell you that the BlanchardKahn conditions are not satisﬁed (whether or not you insert the check command). periods 1:9. To work with a temporary shock. such as periods 5:10. note that we could have entered future periods in the shocks block. 3.2 Deterministic models .
r = 0. In this case.7. end. we make use of the second steady since the actual terminal steady state values are bound to be somewhat diﬀerent from those entered above. This is unusual. endval. k = 9.0. steady. var z.3. TIP! If you instead wanted to study the eﬀects of a permanent but future shock (anticipated as usual). steady. w = 2. w = 2.0. suppose you wanted the value of technology to move to 0. c = 0. the value of technology would move to 0. requires that the path of your endogenous variables pass through the steady state closest to your endval values). In the above example.3. z = 0. which are nothing but the initial values for all variables except for technology. end. z = 0. Then you would follow the above endval block with: shocks. l = 0.7. you would have to add a shocks block after the endval block to “undo” the ﬁrst several periods of the permanent shock. In our example. of telling Dynare to start and/ or end at a steady state close to the values you entered. If you do not use steady after endval. where steady can also be added to the endval block. when solved. r = 0. the other variables . you may impose on your system that it does not return to steady state.1.1. and serves the same functionality as described earlier (namely. but only in period 10.1 in period 1 (tomorrow) and thereafter. But of course.BASICS initval. l = 0. k = 9. and the latter does not list exact steady state values.will take longer to reach their new steady state values.26 CHAPTER 3. . SOLVING DSGE MODELS . For instance. your problem would become a socalled two boundary problem. which. c = 0.the endogenous variables .
end. var e = sigma^2. This can be particularly useful if you’re studying the eﬀects of anticipated shocks in a stochastic model. we have written an instructive . In that case. we would write: shocks.8. end. we would declare zt as an endogenous variable and et as an exogenous variable. Let’s see which are the appropriate commands to give to Dynare. shocks can only hit the system today.8 Selecting a computation So far. TIP! You can actually mix in deterministic shocks in stochastic models by using the commands varexo det and listing some shocks as lasting more than one period in the shocks block. but what should Dynare do with it? What are we interested in? In most cases. zt in our example. Furthermore. as we did in the preamble of the . SELECTING A COMPUTATION periods 1:9. that aﬀects the model’s true exogenous variable. though. Supposing we wanted to add a shock with variance σ 2 . Again. which is itself an AR(1). please see the Reference Manual. A permanent shock cannot be accommodated due to the need to stationarize the model around a steady state. you may be interested in what happens to your monetary model if agents began expecting higher inﬂation. 3. where σ is determined in the preamble block. values 0.mod ﬁle. we will distinguish between deterministic and stochastic models. 27 3. we can however make the eﬀect of the shock propagate slowly throughout the economy by introducing a “latent shock variable” such as et in our example. it will be impulse response functions (IRFs) due to the external shocks. or a depreciation of your currency. .7.3 Stochastic models Recall from our earlier description of stochastic models that shocks are only allowed to be temporary. as the expectation of future shocks must be zero. With that in mind.mod ﬁle in section 3.4. For instance.3. For information on how to do so. exactly as in the model we introduced from the outset.
and repeat this procedure a multitude of times in order to draw out an average response.1 For deterministic models In the deterministic case. If you instead linearize to a second order.8.8. so that the eﬀects of the shocks depend on the state of the system when the shocks hit. 1 . 3. impulse response functions are simply the algebraic forward iteration of your model’s policy or decision rule. it is impossible to get algebraic average values of all future shocks and their impact. the algorithm makes the simplifying assumption that the system is back to equilibrium after the speciﬁed number of periods. the trajectory will basicaly describe how the system gets back to equilibrium after being perturbed from the shocks you entered. impulse response functions and various descriptive statistics (moments. it uses a Newton method to solve simultaneously all the equations for every period (see Juillard (1996) for details). the variance decomposition depends upon the order of the variables in the varexo command. Note that unless you use the endval command. When the shocks are correlated. Thus. Thus. variance decomposition. That said. you will have cross terms involving the shocks. This command instructs Dynare to compute a Taylor approximation of the decision and transition functions for the model (the equations listing current values of the endogenous variables of the model as a function of the previous state of the model and current shocks). the command stoch simul is appropriate.mod ﬁle. To do so. Note that the command takes the option [ (periods=INTEGER) ] The command simul triggers the computation a numerical simulation of the trajectory of the model’s solution for the number of periods set in the option.TIP! If you linearize your model up to a ﬁrst order. This is because in second order linear equations.1 Impulse response functions are the expected future path of the endogenous variables conditional on a shock in period 1 of one standard deviation. correlation and autocorrelation coeﬃcients). The technique is instead to pull future shocks from their distribution and see how they impact your system. SOLVING DSGE MODELS . the variance decomposition is computed as in the VAR literature through a Cholesky decomposition of the covariance matrix of the exogenous variables. note that future shocks will not have a signiﬁcant impact For correlated shocks. you must specify a large enough number of periods such that increasing it further doesn’t change the simulation for all practical purpose.28 CHAPTER 3. for instance. In the case of a temporary shock.2 For stochastic models In the more common case of stochastic models. impulse response functions will be the result of actual Monte Carlo simulations of future shocks.BASICS 3. all you need to do is add the command simul at the bottom of your .
In both cases. . It may be necessary to increase it for highly autocorrelated processes (default = 512). given their mean of zero. please see the Reference Manual. Note that in the case of a second order approximation. Dynare will return the actual sample moments from the simulations. For a complete list of options. • nocorr: doesn’t print the correlation matrix (printing is the default). Dynare will instead report theoretical moments. For ﬁrst order linearizations. Details on implementing this appear below.8. the return to steady state is asymptotic. Setting IRF=0. have a look at Chapter 7. Here instead. Options following the stoch simul command: • ar = INTEGER: Order of autocorrelation coeﬃcients to compute and to print (default = 5). TIP! thus you should make sure to specify suﬃcient periods in your IRFs such that you actually see your graphs return to steady state. If you’re interested to peer a little further into what exactly is going on behind the scenes of Dynare’s computations. we focus on the application of the command and reproduce below the most common options that can be added to stoch simul.3. • drop = INTEGER: number of points dropped at the beginning of simulation before computing the summary statistics (default = 100). • hp ngrid = INTEGER: number of points in the grid for the discreet Inverse Fast Fourier Transform used in the HP ﬁlter computation. • irf = INTEGER: number of periods on which to compute the IRFs (default = 40). since they get averaged between each Monte Carlo trial and in the limit should sum to zero. • nofunctions: doesn’t print the coeﬃcients of the approximated solution (printing is the default). SELECTING A COMPUTATION 29 on your results. • hp ﬁlter = INTEGER: uses HP ﬁlter with lambda = INTEGER before computing moments (default: no ﬁlter). suppresses the plotting of IRF’s. • dr algo = 0 or 1: speciﬁes the algorithm used for computing the quadratic approximation of the decision rules: 0 uses a pure perturbation approach as in SchmittGrohe and Uribe (2004) (default) and 1 moves the point around which the Taylor expansion is computed toward the means of the distribution as in Collard and Juillard (2001b). • relative irf requests the computation of normalized IRFs in percentage of the standard error of each shock.
TIP! A simulation is similar to running impulse response functions with a model linearized to the second order. usefull for loops.mod ﬁle with the following command: stoch simul(periods=2100). Going back to our good old example. • order = 1 or 2 : order of Taylor approximation (default = 2).30 CHAPTER 3. you should at least try to run your program without using simul seed. Otherwise a diﬀerent sample is used for each run (default: seed not speciﬁed). • replic = INTEGER: number of simulated series used to compute the IRFs (default = 1 if order = 1. TIP! If you do decide to ﬁx the seed. but a simulation only repeats the process once. • simul seed = INTEGER or DOUBLE or (EXPRESSION): speciﬁes a seed for the random number generator so as to obtain the same random sample at each run of the program.BASICS • nomoments: doesn’t print moments of the endogenous variables (printing them is the default). • periods = INTEGER: speciﬁes the number of periods to use in simulations (default = 0). SOLVING DSGE MODELS . Dynare will actually undertake Monte Carlo simulations to generate moments of your variables.000001). in the way that both sample shocks from their distribution to see how the system reacts. unless you’re working with a linear model in which case the order is automatically set to 1. . whereas impulse response functions run a multitude of Monte Carlo trials in order to get an average response of your system. are basically happy with all default options and want to carry out simulations over a good number of periods. • noprint: cancel any printing. We would then end our . just to check the robustness of your results. except if you ﬁx the seed for the random number generator. • qz criterium = INTEGER or DOUBLE: value used to split stable from unstable eigenvalues in reordering the Generalized Schur decomposition used for solving 1st order problems (default = 1. suppose we were interested in printing all the various measures of moments of our variables. results are bound to be slightly diﬀerent each time you run your program. Because of the simulation. If you linearized to a second order. and 50 otherwise). want to see impulse response functions for all variables.
model. and for the pleasure of seeing our work bear its fruits. y = (k(1)^alpha)*(exp(z)*l)^(1alpha). 3. delta = 0.mod ﬁle For completion’s sake.75.MOD FILE 31 3. here are the complete . initval. parameters beta psi delta alpha rho gamma sigma epsilon.mod for stochastic models and RBC Monop Det. .03. end. You can ﬁnd the corresponding ﬁles in the models folder under UserGuide in your installation of Dynare. r = y*((epsilon1)/epsilon)*alpha/k(1).33. c = 0. psi = 1. epsilon = 10. l = 0. i = k(1delta)*k(1).023.007/(1alpha)).3.9 The complete .9. varexo e. z = rho*z(1)+e. w = y*((epsilon1)/epsilon)*(1alpha)/l.99.mod for deterministic models. sigma = (0. e = 0. w = 2.9. rho = 0. z = 0. r = 0. alpha = 0. THE COMPLETE . end. k = 9. y l = y/l. psi*c/(1l) = w. c+i = y.1 The stochastic model var y c k i l y l w r z.mod ﬁles corresponding to our example for the deterministic and stochastic case. beta = 0.07. (1/c) = beta*(1/c(+1))*(1+r(+1)delta).76. The ﬁles are called RBC Monop JFV.3.95.
r = 0.9. end. parameters beta psi delta alpha sigma epsilon. (1/c) = beta*(1/c(+1))*(1+r(+1)delta). c = 0.99. 3. y l = y/l.3. . delta = 0. check. c+i = y. beta = 0. i = k(1delta)*k(1). epsilon = 10. end. r = y*((epsilon1)/epsilon)*alpha/k(1). varexo z. model. var e = sigma^2. SOLVING DSGE MODELS .0.007/(1alpha)). shocks.7. w = 2.75.2 The deterministic model (case of temporary shock) var y c k i l y l w r . sigma = (0. psi = 1. steady.33. alpha = 0. end. w = y*((epsilon1)/epsilon)*(1alpha)/l. k = 9. z = 0. y = (k(1)^alpha)*(exp(z)*l)^(1alpha). l = 0.BASICS steady. initval.023.32 CHAPTER 3. stoch simul(periods=2100). psi*c/(1l) = w.
2. end. Model summary: a count of the various variable types in your model (endogenous. all you need to do is place your cursor in the Matlab command window and type. this may not be the case when testing the beta version of Matlab version 4) To run a .mod ﬁles are stored.3. which is conveniently installed by default in the Dynare “examples” directory (the . simul(periods=2100).tabular in the Matlab command window and graphical in one or more popup windows. and you should see a conﬁrmation of the BlanchardKahn conditions if you used the command check in your . for instance.mod and that corresponding to the deterministic model is called RBC Monop Det. dynare ExSolStoch. you should get two forms of output .mod ﬁle.1. As a result..mod ﬁle. 3.mod ﬁle corresponding to the stochastic model is called RBC Monop JFV.mod ﬁle.10.). periods 1:9. 33 3.10 File execution and results To see this all come to life.10. Let’s review these results. Once there. etc. Running these . or typing the path directly in the top white ﬁeld of Matlab. Eigenvalues should be displayed. to execute your .stochastic models The tabular results can be summarized as follows: 1. let’s run our . .mod ﬁles should take at most 30 seconds.mod). FILE EXECUTION AND RESULTS check. navigate within Matlab to the directory where the example .. (** note. shocks. var z. values 0.1 Results . You can do this by clicking in the “current directory” window of Matlab.mod ﬁle. jumpers.
show the actual impulse response functions for each of the endogenous variables. ut ) that could be plugged into the original model and satisfy the implied restrictions (the ﬁrst order conditions). The graphical results. yt−1 . 4. If not. your model will probably not converge. Correlation of simulated variables: these are the contemporaneous correlations. Details on the policy and transition function can be found in Chapter 6. as speciﬁed in the options of stoch simul. given that they actually moved. if all goes well! Finally.BASICS 3. 5.34 CHAPTER 3. TIP! If some variables do not return to their steady state. These can be especially useful in visualizing the shape of the transition functions and the extent to which each variable is aﬀected. eigenvalues will also be displayed and you should receive a statement that the rank condition has been satisﬁed. yt .10. you will get a list of your steady state results. ut )] = 0 . 3.2 Results . either check that you have included enough periods in your simulations. means ﬁnding an unkown function. If you entered check. you should detrend your variables and rewrite your model in terms of those variables. In other words. yt = g(yt−1 .mod ﬁle.e. and where gx is the ˆ ¯ ¯ partial derivative of the g function with respect to variable x. if not. Autocorrelation of simulated variables: up to the ﬁfth lag. you may want to try to increase the periods for the transition to the new steady state (the number of simulations . i. SOLVING DSGE MODELS . with ¯ ˆ yt = yt − y and y being the steadystate value of y. instead. If so. Et [f (yt+1 . TIP! You should see these errors decrease upon each iteration.deterministic models Automatically displayed results are much more scarce in the case of deterministic models. Matrix of covariance of exogenous shocks: this should square with the values of the shock variances and covariances you provided in the . you will see some intermediate output: the errors at each iteration of the Newton solver used to estimate the solution to your model. the table “Policy and Transition function” contains the elements of gy and gu . 7. presented in a table. 6. the function g is a time recursive (approximated) representation of the model that can generate timeseries that will approximatively satisfy the rational expectation hypothesis contained in the original model. that your steady state actually exists and is stable. or make sure that your model is stationary. Policy and transition functions: Solving the rational exectation model. A ﬁrst order approximation of this function can be written as yt = y + gy yt−1 + gu ut . Moments of simulated variables: up to the fourth moments. If you entered steady. In Dynare.
1. Of course. saving and viewing your output. To do so. FILE EXECUTION AND RESULTS 35 periods). But more often.3 of chapter 4 on ﬁnding. it may be a good idea to revise your equations. you should start by looking at section 4.10. . although Dynare does not display a rich set of statistics and graphs corresponding to the simulated output.3. it does not mean that you cannot create these by hand from Matlab.
.
the earlier chapter 3 on the basics of solving DSGE models. if you have read.1 4. don’t forget to check occasionally the Dynare contributions and examples forum to see if any other user has posted an example that could help 37 .not all related to each other . In particular.that you will probably ﬁnd interesting or at least understandable. ﬁnding and saving your output.1 Dynare features and functionality Other examples Other examples of . 4. but possibly also helpful for other work. using loops. and/ or feel comfortable with.mod ﬁles used to generate impulse response functions are available on the Dynare website.Chapter 4 Solving DSGE models advanced topics This chapter is a collection of topics . The second section overviews some of the inner workings of Dynare. The goal is to provide a brief explanation of the ﬁles that are created by Dynare to help you in troubleshooting or provide a starting point in case you actually want to customize the way Dynare works. Jesus FernandezVillaverde has provided a series of RBC model variants (from the most basic to some including variable capacity utilization.1. in the Oﬃcial Examples section of the Dynare website. The third section of the chapter focusses on modeling tips optimized for Dynare. along with helpful notes and explanations. To provide at least some consistency. indivisible labor and investment speciﬁc technological change). The ﬁrst section deals directly with features of Dynare. such as dealing with correlated shocks. this chapter is divided into three sections. Also. You can ﬁnd these. referring to external ﬁles and dealing with inﬁnite eigenvalues.
The economy is a real economy.38 CHAPTER 4. where part of output can be consumed and part invested to form physical capital. and these may be correlated to each other. Your model may have two or more shocks. the law of motion of capital is given by kt+1 = exp(bt )it + (1 − δ)kt with 0 < δ < 1. The example below illustrates how you would introduce this into Dynare. As is standard. It also aims to give you exposure to dealing with several correlated shocks. SOLVING DSGE MODELS . as usual. with 0 < α < 1. A social planner maximizes this utility function subject to the resource constraint ct + it = yt where it is investment and yt output. the disutility of labor θ > 0 and the labor supply elasticity ψ ≥ 0. complete example The following example aims to give you an alternative example to the one in chapter 3. We assume output is produced according to a standard constant returns to scale technology of the form α yt = exp(at )kt h1−α t with α being the capital elasticity in the production function. The model The model is a simpliﬁed standard RBC model taken from Collard and Juillard (2003) which served as the original User Guide for Dynare. and where at represents a stochastic technological shock (or Solow residual). .2 Alternative. the example provided is somewhat more complete than strictly necessary. This is to give you an alternative. Actually. Consumers are therefore also owners of the ﬁrms. The economy consists of an inﬁnitely living representative agent who values consumption ct and labor services ht according to the following utility function ∞ Et τ =t β τ −t h1+ψ log(ct ) − θ t 1+ψ where. or maybe you would like to post an example there yourself? 4.1.ADVANCED TOPICS you in your work. where δ is physical depreciation and bt a shock aﬀecting incorporated technological progress. the discount factor 0 < β < 1. to learn the workings of Dynare. fullblown example to the one described in chapter 3.
4. Et (νt ) = 0 and that the contemporaneous variancecovariance matrix of the innovations t and νt is given by σ2 ψσ σν and where corr( t νs ) = 0. note that to introduce shocks into Dynare. We will assume that you’re comfortable with these and simply present the ﬁnal . though. Note that the ﬁrst equation captures the labor supply function and the second the intertemporal consumption Euler equation.1. Fist. u = phi*0. stderr 0. we assume Et ( t ) = 0. stderr 0.yields the following ﬁrst order conditions (which are straightforward to reproduce in case you have doubts.mod ﬁle To “translate” the model into a language understandable by Dynare. ) and equilibrium conditions drawn from the description above.009*0. DYNARE FEATURES AND FUNCTIONALITY 39 Finally.009.mod ﬁle below. Either write: shocks. ct θh1+ψ = (1 − α)yt t 1 = βEt exp(bt )ct exp(bt+1 )ct+1 exp(bt+1 )α yt+1 +1−δ kt+1 α 1−α yt = exp(at )kt ht kt+1 = exp(bt )it + (1 − δ)kt at = ρat−1 + τ bt−1 + t bt = τ at−1 + ρbt−1 + νt The . .probably quite similar to standard RBC models you have run into .009. That is at bt = ρ τ τ ρ at−1 bt−1 + t νt where ρ + τ  < 1 and ρ − τ  < 1 to ensure stationarity (we call ρ the coeﬃcient of persistence and τ that of crosspersistence). we would follow the steps outlined in chapter 3. end. var u. we specify a shock structure that allows for shocks to display persistence across time and correlation in the current period. . corr( t s) ψσ σν 2 σν = 0 and corr(νt νs ) = 0 for all t = s. . var e. var e. we have two options (this was not discussed in the earlier chapter). This system . Furthermore.009.
end. here is the complete . c*theta*h^(1+psi)=(1alpha)*y. c.36.025. k = exp(b)*(yc)+(1delta)*k(1).95. So that you can gain experience by manipulating the entire model. end. You can ﬁnd the corresponding ﬁle in the models folder under UserGuide in your installation of Dynare. model. phi = 0. c = 0.40 CHAPTER 4. b = tau*a(1)+rho*b(1) + u. psi = 0. k. delta = 0.80359242014163.ADVANCED TOPICS where the last line speciﬁes the contemporaneous correlation between our two exogenous variables. u = phi*0. var y. a. rho = 0. varexo e. delta. parameters beta. Alternatively. . alpha.025.009^2. alpha = 0. The ﬁle is called Alt Ex1. tau = 0. initval. var u = 0. beta = 0.1.009. SOLVING DSGE MODELS . tau. k = beta*(((exp(b)*c)/(exp(b(+1))*c(+1))) *(exp(b(+1))*alpha*y(+1)+(1delta)*k)). theta. y = exp(a)*(k(1)^alpha)*(h^(1alpha)). var e = 0. psi. you can also write: shocks. u.95. rho. b. a = rho*a(1)+tau*b(1) + e.08068253095672. var e.mod. y = 1. theta = 2.99. h.009^2.mod ﬁle corresponding to the above example.009*0.
end.1. shocks. the matrix of autocorrelations that is automatically displayed in the results after running stoch simul has.29175631001732. 41 4. stderr 0. a = 0.009*0. while gamma y{2} represents autocovariances where variables on each column are lagged by one period and so on.1. the diagonal elements of each of the various autocorrelation matrices described here. running down each column. and columns correspond to the variables lagged 1. By default. u = 0. Thus.4. e = 0. b = 0. var u. stderr 0. k = 5.mod ﬁle): • steady state: the steady state of your variables • mean: the mean of your variables • var: the variance of your variables • autocorr: the various autocorrelation matrices of your variables.009.009. Each row of these matrices will correspond to a variables in time t. var e. stoch simul(periods=2100). gamma y{1} represents variances. for the ﬁrst matrix.009. The last matrix (gamma y{7} in .3 Finding. u = phi*0. end. DYNARE FEATURES AND FUNCTIONALITY h = 0. • gamma y: the matrices of autocovariances. or using the workspace interface. var e. then lagged 2 for the second matrix. saving and viewing your output Where is output stored? Most of the moments of interest are stored in global variable oo You can easily browse this global variable in Matlab by either calling it in the command line. and so on. In global variable oo you will ﬁnd the following (NOTE! variables will always appear in the order in which you declared them in the preamble block of your . Dynare will return autocovariances with a lag of 5.
To save your simulated variables. if you want to run loops.4 Referring to external ﬁles You may ﬁnd it convenient to refer to an external ﬁle. You could then retrieve the exogenous shocks from the oo ﬁle by saving them in a ﬁle called dataﬁle. since you could also use the builtin commands in Dynare to generate impulse response functions from estimated shocks. But you may also be interested in the second possibility described above. you will ﬁnd a global variable oo .42 CHAPTER 4.irfs comprising of vectors named endogenous variable exogenous variable. using the shocks(shocks file = datafile.ADVANCED TOPICS the default case) returns the variance decomposition. as described in chapter 5. say. namely of specifying shocks in an external ﬁle. . variables saved with the dynasave command can be retrieved by using the Matlab command load mat FILENAME. to simulate a model based on shocks from a prior estimation. Furthermore.6 of chapter 3 when discussing steady states. or when specifying shocks in an external ﬁle. where each column captures the independent contribution of each shock to the variance of each variable. SOLVING DSGE MODELS . The advantage of using Matlab. you should therefore do just as well to ask Dynare to compute your model’s steady state (except.mat) command. The former is described in section 3. Dynare will save all endogenous variables. you could simulate a deterministic model with the shocks saved from the estimation by specifying the source ﬁle for the shocks. 4. as a result of the independent impulse of each exogenous shock. to ﬁnd your model’s steady state was clear with respect to Dynare version 3. reporting the values of the endogenous variables corresponding to the impulse response functions. like y e.mod ﬁle: dynasave (FILENAME) [variable names separated by commas] If no variable names are speciﬁed in the optional ﬁeld. for instance. But of course. if you decide to run impulse response functions.1. But Dynare version 4 now uses the same analytical methods available in Matlab. in which case writing a Matlab program may be more handy). Finally. as the latter resorted to numerical approximations to ﬁnd steady state values. For most usage scenarios. maybe. you can add the following command at the end of your . this is a bit of a workaround. for instance.mat. In Matlab. either to compute the steady state of your model. to vary your parameter values.
(ii) symbolic derivation of the model equations. the preprocessor will output a C ﬁle (with . (iii) outputting of several ﬁles.these are are ﬁrmly grounded in the theory of generalized eigenvalues. As far as BlanchardKahn conditions are concerned inﬁnite eigenvalues are counted as explosive roots of modulus larger than one.mod”.out of curiosity or maybe to do some customization of your own.c extension) rather than a matlab ﬁle. At that point. parameters) are contained in a “y” (resp.m: a matlab ﬁle containing the model equations and their derivatives (ﬁrst. exogenous variables.m: a matlab ﬁle containing several instructions. Endogenous variables (resp. If the “use dll” option has been speciﬁed in the model declaration.2. The dynare preprocessors essentially does three successive tasks: (i) parsing of the mod ﬁle (it checks that the mod ﬁle is syntactically correct). Second (resp. don’t worry if you get inﬁnite eigenvalues . then the preprocessor creates the following ﬁles: • ﬁlename.1. you may get a message that there is an error in a ﬁle with a new name. . “params”) vector. lag) pairs in the declared model. 4. with an index number depending on the declaration order. If the mod ﬁle is “ﬁlename. which are used from matlab. up to the needed order (depending on the computing needs). You may therefore ﬁnd it helpful to get a brief overview of the internal ﬁles that Dynare generates and the function of each one. or you may want to have a closer look at how Dynare actually solves your model . “g3”). model equations are translated into expression trees). third) derivatives are in “g2” matrix (resp. FILES CREATED BY DYNARE 43 4.2 Files created by Dynare At times. Using a compiled C ﬁle is supposed to give better computing performance in model simulation/estimation. The model equations residuals are stored in a vector named “residuals”.5 Inﬁnite eigenvalues If you use the command check in your . They have no detrimental inﬂuence on the solution algorithm. “x”.mod ﬁle. The model jacobian is put in “g1” matrix. The “y” vector has as many entries as their are (variable. second and maybe third). notably the parameter initializations and the matlab calls corresponding to computing tasks • ﬁlename dynamic. Dynare will report your system’s eigenvalues and tell you if these meet the BlanchardKahn conditions. and its translation into internal machine representation (in particular. It is then compiled to create a library (DLL) ﬁle.4.
You can then reconstruct expost the nonstationary simulated variables after running impulse response functions. suppose your model included: ∞ β j xt+j = 0. with its jacobian. Note that in a stationary model.3. For example. Of course. Thus.3. For deterministic models. not just a variable. If you expect to see a growing curve for a variable.3. because of Jensen’s inequality.m: a matlab ﬁle containing the stationarized version of the model (i. 4.2 Expectations taken in the past For instance. The trick is to use a recursive representation of the sum. you cannot do this for terms that enter your equation in a nonlinear fashion. you can always write yt+1 as yt + dyt+1 . and needs particular care when working with Dynare. where lagged variables are replaced by current variables).ADVANCED TOPICS • ﬁlename static. If you do have nonlinear terms on which you want to take expectations in the past. you are thinking about a growth model. deﬁne st = Et [yt+1 ] and then use s(−1) in your . if you know the trend. to enter the term Et−1 yt . Again. 4. SOLVING DSGE MODELS . you must ﬁrst stationarize your model. Used to compute the steady state. More generally. where dyt = yt − yt−1 . it is expected that variables will eventually go back to steady state after the initial shock. if yt is I(1). Because growth models are nonstationary. j=0 . you can always add it back after the simulation of the stationary components of the variables.3 4. Same notations than the dynamic ﬁle. then linearize it. as if yt were an equation. it is easier to work with the stationarized version of such models. or by letting Dynare do the work.mod ﬁle. either by hand. you need to know the value of dyt at the ﬁnal equilibrium. 4. Replaced by a C ﬁle when “use dll” option is speciﬁed.44 CHAPTER 4.1 Modeling tips Stationarizing your model Models in Dynare must be stationary.e. Note that. such that you can linearize them around a steady state and return to steady state after a shock. you would need to apply the above manipulation to the entire equation.3 Inﬁnite sums Dealing with inﬁnite sums is tricky in general. the trick is to use only stationary variables in t + 1.
Finally. is how to input this inﬁnite sum into Dynare? It turns out that the Calvo price setting implies that the aggregate price follows the equation of motion pt = θpt−1 + (1 − θ)p∗ . β is a discount factor. The RBC model with monopolistic competition introduced in chapter 3 involved ﬂexible prices. exempliﬁed by papers such as Clarida. is instead typical of the new Keya nesian monetary literature. deﬁned as: ∞ St ≡ j=0 β j xt+j . of course. which can also be written in the following recursive manner: ∞ ∞ ∞ St ≡ j=0 β xt+j = xt + j=1 j β xt+j = xt + β j=0 j β j xt+1+j ≡ xt + St+1 This formulation turns out to be useful in problems of the following form: ∞ ∞ β xt+j = pt j=0 j=0 j γ j yt+j . The optimal price for a ﬁrm resetting its price in period t. we can also t rewrite the optimal price setting equation. The trouble. ` la Calvo for instance. i represents a ﬁrm of the continuum between 0 and 1. MODELING TIPS 45 Note that the above can also be written by using an auxiliary variable St .4. for instance. The extension with sticky prices. in a Calvo type setting. and Gertler (1999). given that it will be able to reset its price only with probability 1 − θ each period. and mct is marginal cost as described in the example in chapter 3. Gali. S1 = pt S2.3. S2t = yt + γS2t+1 . thus implying the t following inﬂation relationship πt = (1 − θ)(p∗ − pt−1 ). as ∞ ∞ p∗ − pt−1 = (1 − βθ) t k=0 (βθ)k Et [mct+k ] + k=0 (βθ)k Et [πt+k ] . as illustrated in the following brief example. which can be written as a recursive system of the form: S1t = xt + βS1t+1 . is ∞ p∗ (i) = µ + (1 − βθ) t k=0 (βθ)k Et [mcn (i)] t+k where µ is the markup. This is particularly helpful. after some algebraic manipulations.
by writing the right hand side as the ﬁrst term of the sum (with k = 0) plus the remainder of the sum. The trick now is to note that the above can be written recursively. .3. as the index of the expectations changes with each element of the sum. the best way to handle this is to write out the ﬁrst k terms explicitly and enter each one in Dynare. . 4. but for convenience. this yields: p∗ − pt−1 = (1 − βθ)mct+k + πt + β θEt [p∗ − pt ] t t+1 which has gotten rid of our inﬁnite sum! That would be enough for Dynare. which is the recognizable inﬂation equation in the new θ Keynesian (or new Neoclassical) monetary literature.+Et−k xt . as in the following example. such as: Et−1 xt +Et−2 xt +. SOLVING DSGE MODELS . Suppose your model included the following sum: ∞ yt = j=0 Et−j xt where yt and xt are endogenous variables. a diﬀerent approach than the one mentioned above is necessary.4 Inﬁnite sums with changing timing of expectations When you are not able to write an inﬁnite sum recursively.46 CHAPTER 4. . deﬁned as the marginal cost when prices are perfectly ﬂexible. we can go one step further and write the above as πt = βEt [πt+1 ] + λmct where λ ≡ (1−θ)(1−βθ) . Mathematically. In Dynare.ADVANCED TOPICS where mct+k = mct+k +µ is the deviation of the marginal cost from its natural rate. which can be written as the left hand side term scrolled forward one period and appropriately discounted.
Recall from chapter 3 that we are dealing with an RBC model with monopolistic competition. The goal of this chapter is to lead you through the basic functionality in Dynare to estimate models using Bayesian techniques. but not necessarily to duplicate a “real life scenario” you would face when writing a paper. First. though. the capital share of output. this chapter is structured around an example. We could then use Bayesian methods to estimate the parameters of the model: α. 47 . so that by the end of the chapter you should have the capacity to estimate a model of your own. there’s enough new material to keep you busy. featuring a nonstationary model. you can always move on to chapter 6 where you will ﬁnd a fullﬂedged replication of a recent academic paper. while more advanced topics of practical appeal are discussed in chapter 6. β. Instead. Suppose also that we thought our little RBC model did a good job of reproducing reality. the goal of the example in this chapter is really to explain features in context.1 Introducing an example The example introduced in this chapter is particularly basic. Suppose we had data on business cycle variations of output. This is for two reasons. Second.Chapter 5 Estimating DSGE models basics As in the chapter 3. we thought it would be easiest to simply continue working with the example introduced in chapter 3 with which you are probably already quite familiar. the discount factor. that subject is instead covered in some depth in chapter 8. we did not want to introduce yet another example in this section. Once you feel comfortable with the content of this chapter. This chapter is therefore very practicallyoriented and abstracts from the underlying computations that Dynare undertakes to estimate a model. 5.
w = y*((epsilon1)/epsilon)*(1alpha)/l. In the stationary case. The model’s variables would therefore be stationary and we can proceed without complications.but the Bayesian estimation procedure would still run successfully.mod ﬁle with: var y c k i l y l w r z. psi*c/(1l) = w. i = k(1delta)*k(1). ψ. Note that in Bayesian estimation. (1/c) = beta*(1/c(+1))*(1+r(+1)delta). ρ. The alternative scenario with nonstationary variables is more complicated and dealt with in chapter 6 (in the additional example). Let’s see how to go about doing this. ESTIMATING DSGE MODELS . 5. ρ. and . .yielding posterior distributions identical to prior distributions . c+i = y. our model block would look exactly as in chater 3: model.mod ﬁle. r = y*((epsilon1)/epsilon)*alpha/k(1). we must ﬁrst declare the model’s variables in the preamble of the . the degree of persistence in productivity. the markup parameter. the condition for undertaking estimation is that there be at least as many shocks as there are observables (a less stringent condition than for maximum likelihood estimation).3 Declaring the model Suppose that the equation of motion of technology is a stationary AR(1) with an autoregressive parameter. less than one. y = (k(1)^alpha)*(exp(z)*l)^(1alpha). z = rho*z(1)+e.BASICS δ. the depreciation rate. It may be that this does not allow you to identify all your parameters . parameters beta psi delta alpha rho epsilon.2 Declaring variables and parameters To input the above model into Dynare for estimation purposes. the weight on leisure in the utility function. y l = y/l. varexo e. This is done exactly as described in chapter 3 on solving DSGE models. 5. We thus begin the .48 CHAPTER 5.
mod ﬁle. TIP! During estimation. It is much more eﬃcient to write an external Matlab steady state ﬁle and let Dynare use that ﬁle to ﬁnd the steady state of your model by algebraic procedure. we write: varobs y.7 below. Dynare will end up spending 60 to 70% of the time recalculating steady states. 5. DECLARING OBSERVABLE VARIABLES end. For more details on writing an external Matlab ﬁle to ﬁnd your model’s steady state.5 Specifying the steady state Before Dynare estimates a model. This is just as explained in details and according to the same syntax outlined in chapter 3. Dynare must know which variables are observable for the estimation procedure.4. by providing approximate initial values and relying solely on the builtin Dynare algorithm to ﬁnd the steady state (a numerical procedure). The general syntax to introduce priors in Dynare is the following: . In fact.6. as this chapter uses the same model as that outlined in chapter 3. based on the newest round of parameters available.4 Declaring observable variables This should not come as a surprise. Priors. in ﬁnding the posterior mode. we must give it a hand by declaring approximate values for the steady state.3 of chapter 3. you will signiﬁcantly slow down the computation of the posterior mode. steady and check commands. are declared as a distribution. Thus. please refer to section 3. Dynare recalculates the steady state of the model at each iteration of the optimization routine (more on this later). For the moment. the steady state block will look exactly the same. NOTE! These variables must be available in the data ﬁle. Thus. in Bayesian estimation. 5. it ﬁrst linearizes it around a steady state.5. 49 5. a steady state must exist for the model and although Dynare can calculate it. covering the initval.6 Declaring priors Priors play an important role in Bayesian estimation and consequently deserve a central role in the speciﬁcation of the . as explained in section 5.
p3 ) B(µ. where the following table deﬁnes each term more clearly PRIOR SHAPE NORMAL PDF GAMMA PDF BETA PDF INV GAMMA PDF UNIFORM PDF Corresponding distribution N (µ. you would write alpha. Should it be symmetric? Skewed? If so. You can then pick the standard distribution that best ﬁts . PARAMETER NAME. 0. you should declare them as such. by using the parameters command and its related syntax.1. end. σ) G2 (µ. Some considerations may prove helpful. σ. PRIOR SHAPE. see the Reference Manual.BASICS estimated params. think about the domain of your prior over each parameter. p4 ] R+ [p3 . σ is the PRIOR STANDARD ERROR. and then specify parameters p3 and p4. as explained in chapter 3. yet very important endeavor.50 CHAPTER 5. PRIOR 3rd PARAMETER] [. as well as the syntax to declare priors for maximum likelihood estimation (not Bayesian). Should it be bounded? Should it be opened on either or both sides? Remember also that if you specify a probability of zero over a certain domain in your prior. and repeat the exercise by incrementally decreasing that value. p4 ) IG1 (µ. ESTIMATING DSGE MODELS . for instance. . p3 is the PRIOR 3rd PARAMETER (whose default is 0) and p4 is the PRIOR 4th PARAMETER (whose default is 1). . you will necessarily also ﬁnd a probability of zero in your posterior distribution. on which side? You may also go one step further and build a distribution for each of your parameters in your mind. It is worth spending time on your choice of priors and to test the robustness of your results to your priors. since the uniform distribution only takes p3 and p4 as arguments. p4 ) Range R [p3 . PRIOR STANDARD ERROR [. Then. For a more complete review of all possible options for declaring priors. p3 . First. +∞) [p3 . TIP! Choosing the appropriate prior for your parameters is a tricky. For instance. σ. say α. think about the shape of your prior distribution. uniform pdf. Ask yourself. PRIOR MEAN. σ) U (p3 . TIP! When specifying a uniform distribution between 0 and 1 as a prior for a parameter. Note also that if some parameters in a model are calibrated and are not to be estimated. what is the probability that your parameter is bigger than a certain value.PRIOR 4th PARAMETER] . p4 ] where µ is the PRIOR MEAN. you therefore have to put two empty spaces for parameters µ and σ.
bet. stderr e. end. 1. Coming back to our basic example. 0.003. psi.5.05.01. alpha.02. epsilon. than its inverse which often shows up in Euler equations. Thus you would write: model.02. we would write: estimated params. gamma pdf.35. This is especially useful when redoing an estimation .002. c = sig*c(+1)*mpk. TIP! Finally. you are instead encouraged to visualize these distributions yourself. DECLARING PRIORS 51 your perceived distribution.95.normal pdf.99. beta pdf. 0. it is possible to declare the parameter to be estimated in the parameters statement and to deﬁne the transformation at the top of the model section. TIP! It may at times be desirable to estimate a transformation of a parameter appearing in the model. β.0. beta pdf. . rather than the parameter itself.75.1. 0. For example. 0. inv gamma pdf.if the optimizer got stuck the ﬁrst time around. 10. end. you may ﬁnd it easier to deﬁne a prior over the discount factor. 0. delta. 0. estimated params. The Reference Manual gives more details as to the exact syntax of this command. 0. instead of describing here the shapes and properties of each standard distribution available in Dynare. 0. rho. In such a case.6. another useful command to use is the estimated params init command which declares numerical initial values for the optimizer when these are diﬀerent from the prior mean. beta. either in a statistics book or on the Web.003.and wanting to enter the posterior mode as initial values for the parameters instead of a prior. beta pdf. 0. beta pdf. end.025. as a Matlab expression. or needing a greater number of draws in the MetropolisHastings algorithm . # sig = 1/bet.05. gamma pdf. by adding a pound sign (#) at the beginning of the corresponding line. Finally. inf. 0. 0.
52 CHAPTER 5. and variable names would show up in the ﬁrst cell of each column. and therefore have zero mean. mh replic = INTEGER: number of replication for Metropolis Hasting algorithm. or instance. ﬁrst obs = INTEGER: the number of the ﬁrst observation to be used (default = 1). separated by commas. after the command estimation). 1. Below. . ESTIMATING DSGE MODELS . In Excel ﬁles. preﬁlter = 1: the estimation procedure demeans the data (default=0. Note that observations do not need to show up in any order.mat ﬁle. observations could be ordered in columns. nobs = INTEGER: the number of observations to be used (default: all observations in the ﬁle) 3. all that is necessary is to add the command estimation at the end of the . mh nblocks = INTEGER: number of parallel chains for Metropolis Hasting algorithm (default = 2).7 Launching the estimation To ask Dynare to estimate a model. and encourage you to view the Reference Manual for a complete list. dataﬁle = FILENAME: the dataﬁle (a .m ﬁle. conf sig = {INTEGER — DOUBLE }: the level for the conﬁdence intervals reported in the results (default = 0. a . But the real complexity comes from the options available for the command (to be entered in parentheses and sequentially. for instance. 4. to divide the observations into subperiods. 2.90) 7. This is useful if model variables are in deviations from steady state. or an . Demeaning the observations would also impose a zero mean on the observed variables. mh replic should be larger than 1200 (default = 20000) 8. This is useful when running loops. one of the key criteria to evaluate the eﬃciency of the MetropolisHastings to evaluate the posterior distribution.BASICS 5. for instance. we list the most common and useful options. Easy enough. More details on this subject appear in Chapter 6.xls ﬁle). but vectors of observations need to be named with the same names as those in var obs. it is advisable to work with a higher value. For the time being. nograph: no graphs should be plotted. 6.mod ﬁle. 5. no preﬁltering). This improves the computation of between group variance of the parameter means. such as 5 or more. Despite this low default value.
mode check: when mode check is set. which would defeat the purpose of running several blocks of MetropolisHastings chains. ideally. 0: the mode isn’t computed. 10. . The idea here is to draw initial values from a stretched out distribution in order to maximize the chances of these values not being too close together. an acceptance rate of 25% in the Metropolis. the literature has settled on a value of between 0. Note that the acceptance rate drops if you increase the scale used in the jumping distribution and vice a versa. mh drop = DOUBLE: the fraction of initially generated parameter vectors to be dropped before using posterior simulations (default = 0. mh jscale = DOUBLE: the scale to be used for the jumping distribution in MH algorithm. Dynare plots the minus of the posterior density for values around the computed mode for each estimated parameter in turn. 14. your MetropolisHastings iterations would never visit the tails of a distribution. 3: uses Matlab fminunc. This is helpful to diagnose problems with the optimizer.2). 4 (default): uses Chris Sim’s csminwel. mode ﬁle must be speciﬁed 1: uses Matlab fmincon (see the Reference Manual to set options for this command). while if it were too low. The default value is rarely satisfactory. the iterations would get stuck in a subspace of the parameter range. 12. mode ﬁle=FILENAME: name of the ﬁle containing previous value for the mode.7. mh init scale=DOUBLE: the scale to be used for drawing the initial value of the MetropolisHastings chain (default=2*mh jscale). mode compute=INTEGER: speciﬁes the optimizer for the mode computation. this means that the ﬁrst half of the draws from the MetropolisHastings are discarded). If the acceptance rate were too high. When computing the mode. LAUNCHING THE ESTIMATION 53 9. A clear indication of a problem would be that the mode is not at the trough (bottom of the minus) of the posterior distribution. This option must be tuned to obtain.5. 13.2 and 0. The idea is not to reject or accept too often a candidate parameter.5. 2: uses Lester Ingber’s Adaptive Simulated Annealing. 11.Hastings algorithm (default = 0.4. Dynare stores the mode (xparam1) and the hessian (hh) in a ﬁle called MODEL NAME mode. This is a particularly helpful option to speed up the estimation process if you have already undertaken initial estimations and have values of the posterior mode.
More details on these diagnostics are given in Chapter 6. The length of the IRFs are controlled by the irf option. ** will be implemented shortly . only the corresponding parameter estimates will be used to generate the IRFs. See the note below on the diﬀerence between ﬁltered and smoothed shocks.in Dynare version 4. load mh ﬁle: when load mh ﬁle is declared. not smoothed variables. or carry out maximum likelihood estimation.54 CHAPTER 5. are built only based on knowing past information. Filtered shocks. for each set of draws. 21. 17. separated by commas. as speciﬁed in chapter 3 when discussing the options for stoch simul. for instance. TIP! If you stop the estimation procedure after calculating the posterior mode. Dynare pulls parameter and shock values from the corresponding estimated distributions and. the IRFs will use the parameters the posterior distributions. nodiagnostic: doesn’t compute the convergence diagnostics for MetropolisHastings (default: diagnostics are computed and displayed). on the other hand. 18. Again.if not already . If you instead carry out a full MetropolisHastings estimation.if not already . ﬁltered vars: triggers the computation of the posterior distribution of ﬁltered endogenous variables and shocks. instead.8 of chapter 3.BASICS 15. To calculate one period ahead prediction errors. Repeating this process often enough generates a distribution of IRFs. Dynare adds to previous MetropolisHastings simulations instead of starting from scratch. ** will be implemented shortly . Actually seeing if the various blocks of MetropolisHastings runs converge is a powerful and useful option to build conﬁdence in your model estimation. . bayesian irf: triggers the computation of the posterior distribution of impulse response functions (IRFs). using all the information contained in the sample of observations.in Dynare version 4. either see the Reference Manual or section 3. 19. including the variance of the shocks. moments varendo: triggers the computation of the posterior distribution of the theoretical moments of the endogenous variables as in stoch simul (the posterior distribution of the variance decomposition is also included). 16. generates an IRF. To build the posterior distribution of the IRFs. you should use ﬁltered. All options available for stoch simul can simply be added to the above options. Smoothed shocks are a reconstruction of the values of unobserved shocks over the sample. smoother: triggers the computation of the posterior distribution of smoothed endogenous variables and shocks. 20. ESTIMATING DSGE MODELS . To view a list of these options. this is a useful option to speed up the process of estimation.
The corresponding graph includes one conﬁdence interval describing uncertainty due to parameters and one conﬁdence interval describing uncertainty due to parameters and future shocks. You need to run MetropolisHastings iterations before being able to run forecasts on an estimated model. Again. In this test ﬁle. make sure you have steady and possibly check following the initval block if you do not have exact steady state values and run a simulation using stoch simul at the end of your . 5.if not already . coming back to our example.5. except that the forecast does not begin at a steady state.mh jscale=0. as in bayesian irf.mod ﬁle. repeating this step often enough yields a posterior distribution of the forecast. as observation do not exist for all variables.mod ﬁle. Finally. .mod ﬁle To summarize and to get a complete perspective on our work so far. forecast = INTEGER: computes the posterior distribution of a forecast on INTEGER periods after the end of the sample used in estimation. here is the complete . You may therefore want to create a test version of your . TIP! Before launching estimation it is a good idea to make sure that your model is correctly declared. running a forecast is very similar to an IRF. THE COMPLETE .mod ﬁle for the estimation of our very basic model. those necessary are reconstructed by sampling out of the posterior distribution of parameters. Details on model solution and simulation can be found in Chapter 3.mh nblocks=2.mh drop=0.first obs=500. that a steady state exists and that it can be simulated for at least one set of parameter values. mh replic=2000. Note that Dynare cannot forecast out of the posterior mode.in Dynare version 4.mod ﬁle. add a shocks block.mod.45. ** will be implemented shortly . The goal of undertaking a forecast is to see how the system returns to steady state from this starting point. This ends our description of the .8. we could choose a standard option: estimation(datafile=simuldataRBC. remove the prior estimates for parameter values and replace them with actual parameter values in the preamble. You can ﬁnd the corresponding ﬁle in the models folder under UserGuide in your installation of Dynare.8.8 The complete . remove any nonstationary variables from your model.nobs=200. Finally. The ﬁle is called RBC Est. Of course. but simply at the point corresponding to the last set of observations. you would comment out or erase the commands related to estimation.MOD FILE 55 22. mode compute=6).
0. 0.02.BASICS var y c k i l y l w r z. 0. psi*c/(1l) = w. inf. beta pdf. parameters beta psi delta alpha rho epsilon. steady. z = 0.99. check. beta pdf. beta pdf. psi. rho.75. varexo e. y l = y/l. beta.first obs=500. w = y*((epsilon1)/epsilon)*(1alpha)/l. delta.95. end. initval. gamma pdf.002. r = y*((epsilon1)/epsilon)*alpha/k(1). end. c = 0. 0. e = 0.nobs=200. r = 0.02.mh jscale=0. l = 0.7.025. mh replic=2000. k = 9.mh drop=0. alpha. 10. (1/c) = beta*(1/c(+1))*(1+r(+1)delta). c+i = y. varobs y. stderr e.8. . epsilon. 1.56 CHAPTER 5. 0. z = rho*z(1)+e.003. 0. i = k(1delta)*k(1).003. inv gamma pdf. gamma pdf. estimation(datafile=simuldataRBC. mode compute=6). 0.35.01. estimated params. 0. end. w = 2. model. beta pdf. 0. ESTIMATING DSGE MODELS . 0.05. y = (k(1)^alpha)*(exp(z)*l)^(1alpha). 0.0.45.mh nblocks=2.3.
as explained in more details in Chapter 6.5. The table of eigenvalues is completed with a statement about the BlanchardKahn condition being met .9 Interpreting output As in the case of model solution and simulation.2 Graphical results ** corresponding graphs will be reproduced below. The ﬁgure returns a graphical representation of the priors for each parameter of interest.mod ﬁle above. It is followed by a second table summarizing the same results for the shocks. Dynare will display the following results. It may be entirely possible that you get an inﬁnite value for a standard deviation. presented in the order in which the endogenous variables are declared at the beginning of the . The improvement from one iteration to the next reaches zero. this is simply the limit case of the inverse gamma distribution. as opposed to standard maximum likelihood).9.mod ﬁle.1 Tabular results The ﬁrst results to be displayed (and calculated from a chronological standpoint) are the steady state results.9. 5. standard deviation and tstat of the mode (based on the assumption of a Normal. These results are followed by the eigenvalues of the system. On the basis of the options entered in the example .9. 5. Dynare returns both tabular and graphical output. The ﬁrst ﬁgure comes up soon after launching Dynare as little computation is necessary to generate it. Dynare give the value of the objective function (the posterior Kernel) at the mode and displays two important table summarizing results from posterior maximization.hopefully! The next set of results are for the numerical iterations necessary to ﬁnd the posterior mode. as well as the prior distribution and standard deviation (pstdev). Note the dummy values of 1 for the nonstationary variables Y obs and P obs. probably erroneous when undertaking Bayesian estimation. INTERPRETING OUTPUT 57 5. posterior mode. . The ﬁrst table summarizes results for parameter values. It includes: prior means.
you can also . Dynare plots a green line which represents the posterior mode. “m2”. corresponding to the measure at the initial value of the MetropolisHastings iterations. Dynare reports three measures: “interval”. Dynare reports both the within and the between chains measures. It is advisable to redo the estimation with diﬀerent priors. This is the idea of what the MCMC diagnostics track. grey lines). More speciﬁcally. the red and blue lines on the charts represent speciﬁc measures of the parameter vectors both within and between chains. Recall that Dynare completes several runs of MetropolisHastings simulations (as many as determined in the option mh nblocks. being a measure of the variance and “m3” based on third moments. where MCMC stands for Monte Carlo Markov Chains.displays the most interesting set of results. you can tell that indeed. This is the main source of feedback to gain conﬁdence. First. Note that the horizontal axis represents the number of MetropolisHastings iterations that have been undertaken. The ﬁgure entitled “multivariate diagnostic” presents results of the same nature. If you have trouble coming up with a new prior. ESTIMATING DSGE MODELS . and the vertical axis the measure of the parameter moments.ﬁgure 6 in our example . try starting with a uniform and relatively wide prior and see where the data leads the posterior distribution. For the results to be sensible. on the posterior distribution. TIP! If the plotted moments are highly unstable or do not converge. and the optimizer did not get stuck in an odd area of the parameter subspace.58 CHAPTER 5. the ﬁgure compares the posterior to the prior distribution (black vs. The ﬁrst to last ﬁgure .BASICS The second set of ﬁgures displays “MCMC univariate diagnostics”. or spot a problem. being constructed from an 80% conﬁdence interval around the parameter mean. we obtain convergence and relative stability in all measures of the parameter moments. you may have a problem of poor priors. And second. In addition. Another approach is to undertake a greater number of MetropolisHastings simulations. results within any of the however many iterations of MetropolisHastings simulation should be similar. towards which most of the computations undertaken by Dynare are directed: the posterior distribution. except that they reﬂect an aggregate measure based on the eigenvalues of the variancecovariance matrix of each parameter. In our example above. results between the various chains should be close. each time starting from a diﬀerent initial value). These allow you to make statements about your data other than simply concerning the mean and variance of the parameters. In each case. with the ﬁrst. If the results from one chain are sensible. two things should happen. In fact. with results. these should be relatively constant (although there will always be some variation) and they should converge.
5. the prior and the posterior distributions should not be excessively diﬀerent.9. If not. TIP! These graphs are of course especially relevant and present key results. the green mode (calculated from the numerical optimization of the posterior kernel) should not be too far away from the mode of the posterior distribution. INTERPRETING OUTPUT 59 discuss the probability that your parameter is larger or smaller than a certain value. First. Second. One thing to check is the fact that shocks should be centered around zero. That is indeed the case for our example. or at least not display a shape that is clearly nonnormal. it is advisable to undertake a greater number of MetropolisHastings simulations. in this case. the posterior distributions should be close to normal. The horizontal axis. . represents the number of periods in the sample. The last ﬁgure returns the smoothed estimated shocks in a useful illustration to eyeball the plausibility of the size and frequency of the shocks. but they can also serve as tools to detect problems or build additional conﬁdence in your results. Third.
.
This ﬁrst section introduces the model. The example shows how to use Dynare in a more realistic setting. So we thought you would enjoy (and continue learning from!) a more realistic example which reproduces the work in a recent . The chapter begins by presenting a more complex example than the one used for illustration purposes in chapter 5. while emphasizing techniques to deal with nonstationary observations and stochastic trends in dynamics. The chapter then follows with sections on comparing models to one another.1. The goal is to show how Dynare would be used in the more “realistic” setting of reproducing a recent academic paper. Note that the original paper by Schorfheide mainly focusses on estimation methodologies. 6. its basic intuitions and equations.academic paper. with a special interest in model comparison. and then to BVARs. while the mathematics and economic intuitions of the model it evaluates are drawn from Nason and Cogley (1994).1 Introducing the example The example is drawn from Schorfheide (2000). and ends with a table summarizing where output series are stored and how these can be retrieved. diﬃculties and solutions.Chapter 6 Estimating DSGE models advanced topics This chapter focusses on advanced topics and features of Dynare in the area of model estimation. We will then see in subsequent sections how to estimate it using Dynare.1 Alternative and nonstationary example The example provided in chapter 5 is really only useful for illustration purposes. That paper should serve as a helpful reference if anything is 61 . 6.and highly regarded .
You may want to refer back to the ﬁgure as you read through the following sections. So to clarify things. and a monetary authority which plays a minor role. ﬁgure 6. ﬁrms and banks (representing the ﬁnancial sector).62 CHAPTER 6. In essence.1. left unclear in the description below. Households maximize their utility function which depends on consumption. while deciding how much money to hold next period in cash.t − 1 interest. and hours worked. in order to earn RH. In the model. while observing only output and inﬂation. ESTIMATING DSGE MODELS . there are several markets and actors to keep track of.1 sketches the main dynamics of the model. Dt . while dashed lines show the ﬂow of real variables. Mt+1 and how much to deposit at the bank.1: Continuous lines show the circulation of nominal funds. the model studied by Schorfheide (2000) is one of cash in advance (CIA). Ct .ADVANCED TOPICS Figure 6. Households . Ht . The economy is made up of three central agents and one secondary agent: households. The goal of the paper is to estimate the model using Bayesian techniques.
wages. ﬁrms maximize the net present value of future dividends (discounted by the marginal utility of consumption. as well as dividends from ﬁrms. The third equation simply speciﬁes that bank loans are used to pay for wage costs. Xt from the central bank (which equals the net change in nominal money balances. Bt .Nt . receive cash deposits from households and a cash injection.t Dt + Ft + Bt where the second equation spells out the cash in advance constraint including wage revenues.Dt } 63 E0 ∞ t t=0 β [(1 − φ) ln Ct + φ ln(1 − Ht )] 0 ≤ Dt s. Note that it is the ﬁrms that engage in investment in this model. Ft . Its problem is summarized by max {Ft . It uses these funds to disburse loans to ﬁrms.t. and loans.t due to the equal risk proﬁles of the loans. bank dividends. To close the model.t = RF. on which they make a net return of RF.Ht . where It = Kt+1 − (1 − δ)Kt . Finally. since they are owned by households) by choosing dividends. Ht = Nt and Pt Ct = Mt + Xt .6. which in both cases are made up of net cash inﬂows deﬁned below.t Lt −RH. ALTERNATIVE AND NONSTATIONARY EXAMPLE therefore solve the problem max {Ct .Kt+1 .1.t Dt −Lt +Xt .t. Nt . as well as the condition that RH. and where δ is the rate of depreciation. Bt are simply equal to Dt +RF. Ft ≤ Lt + Pt Kt (At Nt )1−α − Kt+1 + (1 − δ)Kt − Wt Nt − Lt RF. next period’s capital stock. Pt Ct ≤ Mt − Dt + Wt Ht Mt+1 = (Mt − Dt + Wt Ht − Pt Ct ) + RH. Of course.Mt+1 . Mt+1 − Mt ). and from banks. labor demand. we add the usual labor and money market equilibrium equations.Lt } E0 ∞ Ft t+1 t=0 β Ct+1 Pt+1 α s. the third the inability to borrow from the bank and the fourth the intertemporal budget constraint emphasizing that households accumulate the money that remains after bank deposits and purchases on goods are deducted from total inﬂows made up of the money they receive from last period’s cash balances. Kt+1 . interests.t − 1. by trading oﬀ investment for dividends to consumers. Banks. banks are constrained in their loans by a credit market equilibrium condition Lt ≤ Xt + Dt . Finally. .t Wt Nt ≤ Lt α where the second equation makes use of the production function Yt = Kt (At Nt )1−α and the real aggregate accounting constraint (goods market equilibrium) Ct + It = Yt . Lt . on their end.
A. but an AR(2) with a unit root in the log of the level of money. Otherwise said. Note that theses expressions for trends are not written in the most straightforward manner nor very consistently. M. we obtain the following set of ﬁrst order and equilibrium conditions. and the second an autoregressive stationary process in the growth rate of money. aﬀecting the money stock. but encourage you to browse Nason and Cogley (1994) for additional details. As usual.1 When the above functions are maximized. The ﬁrst equation is therefore a unit root with drift in the log of technology. for details. we add a stochastic elements to the model.t . we could have written the AR(2) process in state space form and realized that the system has an eigenvalue of one.64 CHAPTER 6. see Hamilton (1994). the rate of growth of the series is a stationary stochastic process. ESTIMATING DSGE MODELS . to save space. aﬀecting technology and one nominal. σA ) and ln mt = (1 − ρ) ln m∗ + ρ ln mt−1 + M. one real. chapter 15.t 2 ∼ N (0. We will not dwell on the derivations here. These important equations are ln At = γ + ln At−1 + A.ADVANCED TOPICS More importantly. σM ) where mt ≡ MT +1 /Mt is the growth rate of the money stock. . if the logs of a variable are speciﬁed to follow a unit root process. This can be seen from the deﬁnition of mt which can be rewritten as ln Mt+1 = ln Mt + ln mt . But we reproduced them nevertheless to make it easier to compare this example to the original paper. The model allows for two sources of perturbations.t 2 ∼ N (0.t . one is a root of the second order autoregressive lag polynomial. We nonetheless give a brief intuitive explanation of each 1 Alternatively.
. We come back to this important topic in details in section 6. representing the tradeoﬀ to the economy of moving consumption goods across time. importantly. The system comes down to Et − Pt / Ct+1 Pt+1 mt = βe−α(γ+ A. labor demand. 2. we pause a moment to give some intuition for the above equations. linking labor supply. The equilibrium interest rate in which the marginal revenue product of labor equals the cost of borrowing to pay for that additional unit of labor. ALTERNATIVE AND NONSTATIONARY EXAMPLE equation. 3. hats over variables no longer mean deviations from steady state.t A. The Euler equation in the goods market.t ) Yt /Yt−1 A. these equations correspond to: 1.t ) ln(mt ) = (1 − ρ) ln(m ) + ρ ln(mt−1 ) + At ≡ dAt = exp(γ + At−1 Yt /Yt−1 = eγ+ A.t+1 ) 65 α−1 1−α Pt+1 αKt Nt+1 + (1 − δ) / ct+2 Pt+2 mt+1 Wt = Lt /Nt φ Ct Pt / (1 − Nt ) 1−φ −1 = Lt /Nt A. but instead represent variables that have been made stationary. In order. also aﬀecting labor demand. as ﬁrms use borrowed funds to pay for labor input.6. For now.t+1 ) −1 A.t ) Ct + Kt = e−α(γ+ Pt C = mt mt − 1 + Dt = Lt A.1. and the marginal rate of substitution between consumption and leisure. The ﬁrms’ borrowing constraint.1. The intertemporal labor market optimality condition.t M.t ) α Kt−1 N 1−α + (1 − δ)e−(γ+ Kt−1 α Yt = Kt−1 N 1−α e−α(γ+ A.3 below. 4.t+1 ) Rt = (1 − α)Pt e−α(γ+ Ct P t α Kt−1 Nt−α /Wt α Kt−1 Nt1−α = β (1 − α) Pt e−α(γ+ ×Et Lt mt Ct+1 Pt+1 A.t Pt /Pt−1 = (Pt /Pt−1 )(mt−1 /eγ+ ) where.
8. 11. 6. 7. which is stationary as there is no population growth). Nt .2 Declaring variables and parameters This block of the . The production function.ADVANCED TOPICS 5. The stochastic process for technology. 10. 9.mod ﬁle follows the usual conventions and would look like: var m P c e W R k d n l Y obs P obs y dA.1. and where the last ﬁve variables are deﬁned and explained in more details in the section below on declaring the model in Dynare. The nonstationarity comes out clearly when attempting to solve the model for a steady state and realizing it does not have one. ESTIMATING DSGE MODELS . 12. bet.1. Detrending therefore involves . where the choice of upper and lower case letters is not signiﬁcant. gam. The stochastic process for money growth. mst. The Euler equation in the credit market. up to l.3 The origin of nonstationarity The problem of nonstationarity comes from having stochastic trends in technology and money. varexo e a e m. parameters alp. 6.66 CHAPTER 6. are as speciﬁed in the model setup above. rho. the ﬁrst set of endogenous variables. It can be shown that when shocks are null. del. more details on these last two equations appear in the following section. nominal variables grow with Mt and prices with Mt /At . The money market equilibrium condition equating nominal consumption demand to money demand to money supply to current nominal balances plus money injection. The credit market equilibrium condition. The exogenous variables are as expected and concern the shocks to the evolution of technology and money balances. which ensures that giving up one unit of consumption today for additional savings equals the net present value of future consumption. psi. real variables grow with At (except for labor. The relationship between observable variables and stationary variables. 6. The aggregate resource constraint.
The above is the equation we retain for the . it . is to deﬁne dA = exp(gam+e a) by simply taking the exponential of both sides of the stochastic process of technology deﬁned in the model setup above. Pt = Pt ·At /Mt .mod ﬁle of Dynare into which we enter: y=k(1)^alp*n^(1alp)*exp(alp*(gam+e a)) The other equations are entered into the . ˆ variables. ct .mod ﬁle after transforming them in exactly the same way as the one above. that turns out to be useful since we often deal with the growth rate of technology. Q 6. if you wish (Nason and Cogley (1994) includes more details on the transformations of each equation). qt = qt /At .t )) where we go from the second to the third line by taking the exponential of both sides of the equation of motion of technology.4 Stationarizing variables Let’s illustrate this transformation on output. Thus. We deﬁne Yt to equal Yt /At and Kt as Kt /At . For nominal ˆ ˆ t = Qt /Mt . today’s capital stock is a result of yesterday’s decision). The resulting equation made up of stationary variables is Yt At = Kt−1 At−1 α 1−α At Nt1−α A−1 Aα t−1 t α Yt = Kt−1 Nt1−α At At−1 −α α = Kt−1 Nt1−α exp(−α(γ + A. in the output equation. lt .5 in chapter 3). and leave the transformations of the remaining equations as an exercise.6. We stationarize output by dividing its real variables (except for labor) by At .1. kt+1 ]. where Qt = [dt . where qt = [yt .1. And for prices. ALTERNATIVE AND NONSTATIONARY EXAMPLE 67 the following operations (where hats over variables represent stationary variables): for real variables. we should actually work with Kt−1 = Kt−1 /At−1 . A ﬁnal transformation to consider. Wt ]. that in Dynare variables take the time subscript of the period in which they are decided (in the case of the capital stock. . NOTE! Recall from section 3.
to our (stationary) model’s variables Yt by using the deﬁnition that Yt ≡ Yt /At . The goal is to add a line to the model block of our .mod ﬁle that relates the non stationary observables. we therefore follow this path in the remainder of the example. The details of the correct transformations for prices are left as an exercise and can be checked against the results below. W = l/n. the y of the . We could simply write Yobs = Yt At . R = P*(1alp)*exp(alp*(gam+e a))*k(1)^alp*n^(alp)/W. except that we use the relationship Pobs = Pt Mt /At as noted earlier. In the case of output. our other observable variable.mod ﬁle: gy obs = dA*y/y(1).68 CHAPTER 6. from what we just saw above.1.1. where. the observable variable would become Yt /Yt−1 .6 The resulting model block of the .mod ﬁle model. We would then have to relate this observable. call it gy obs. P/(c(+1)*P(+1)*m)+bet*P(+1)*(alp*exp(alp*(gam+log(e(+1))))*k^(alp1) *n(+1)^(1alp)+(1del)*exp((gam+log(e(+1)))))/(c(+2)*P(+2)*m(+1))=0.mod ﬁle. though. we could also work with nonstationary data in levels. we therefore add: Y obs/Y obs(1) = dA*y/y(1). But. but just a dAt . The result is not very diﬀerent. This complicates things somewhat. 6. ESTIMATING DSGE MODELS . (psi/(1psi))*(c*P/(1n))+l/n = 0. Thus. Yt . dA = exp(gam+e a). We could simply stationarize them by working with rates of growth (which we know are constant). to our stationary output.ADVANCED TOPICS 6. But since we don’t have an At variable. We of course do the same for prices. 1/(c*P)bet*P*(1alp)*exp(alp*(gam+e a))*k(1)^alp*n^(1alp)/ (m*l*c(+1)*P(+1)) = 0. but illustrates several features of Dynare worth highlighting. log(m) = (1rho)*log(mst) + rho*log(m(1))+e m. c+k = exp(alp*(gam+e a))*k(1)^alp*n^(1alp)+(1del) . To the . call them Yobs . we wewrite the above relationship in ratios. we add to the model block of the .mod ﬁle are the stationary Yt . we must make a decision as to our nonstationary observations.5 Linking stationary variables to the data And ﬁnally.
In our . to specify that our observable variables are indeed P obs and Y obs as noted in the section above. In the case of output. is that we cannot linearize the above system in levels.6. Y obs/Y obs(1) = dA*y/y(1). P obs/P obs(1) = (p/p(1))*m(1)/dA. Both output and prices exhibit stochastic trends.mod ﬁle we write varobs P obs Y obs. after all.1. 6. though. as the last two equations don’t have a steady state. such as ending lines with semicolons and indicating the timing of variables in parentheses. TIP! In the above model block. If we ﬁrst take logs. they become linear and it doesn’t matter anymore where we calculate their derivative when taking a Taylor expansion of all the equations in the system. end. you should add measurement shocks to your model where you deem most appropriate. nonstationary.8 Declaring trends in observable variables Recall that we decided to work with the nonstationary observable variables in levels. the input conventions.7 Declaring observable variables We begin by declaring which of our model’s variables are observables. 6. you must loglinearize your model (and not just linearize it).1.1. ALTERNATIVE AND NONSTATIONARY EXAMPLE *exp((gam+e a))*k(1). y = k(1)^alp*n^(1alp)*exp(alp*(gam+e a)). when dealing with nonstationary observations. concerning the observables which are. If this is not the case. of course. This can be seen explicitly by taking the diﬀerence of logs of output and prices to compute growth rates. are the same as those listed in chapter 3. Thus. The problem that arises. though. we make use of the usual (by now!) . this is a point to which we will return later. m1+d = l. 69 where. e = exp(e a). P*c = m. notice that what we have done is in fact relegated the nonstationarity of the model to just the last two equations. NOTE! Recall from earlier that the number of observed variables must be smaller or equal to the number of shocks such that the model be estimated.
mod unit root vars P obs Y obs. whereas we know ∆ ln Yt is stationary in steady state.2 In Dynare. 6.t Note from the original equation of motion of ln mt that in steady state. end. In the case of prices. The algorithm to compute a true diﬀuse prior is taken from Durbin and Koopman (2001). the command observation trends speciﬁes linear trends as a function of model parameters for the observed variables in the model.t where all terms on the right hand side are constant. Note that for stationary variables. whether deterministic or stochastic (the drift term) must be declared up front. we therefore write (in a somewhat cumbersome manner) observation trends. technology and money.t emphasizing clearly the drift term γ.t − γ − A. ESTIMATING DSGE MODELS . Unit roots are only related to stochastic trends. we write in the . NOTE! You don’t need to declare unit roots for any nonstationary model. You don’t need to use a diﬀuse This can also be see from substituting for ln mt−1 in the above equation with the equation of motion of ln mt to yield: ∆ ln Pt = ∆ ln Pt + ln m∗ + ρ(ln mt−2 − ln m∗ ) + M. we ﬁnd: ∆ ln Yt = ∆ ln Yt + γ + A. In the case of our example. so that the drift terms in the above equation are ln m∗ − γ.9 Declaring unit roots in observable variables And ﬁnally. To give these instructions to Dynare. ln mt = ln m∗ .ADVANCED TOPICS relationship Yt = Yt · At . P obs (log(mst)gam). we apply the same manipulations to show that: ∆ ln Pt = ∆ ln Pt + ln mt−1 − γ − A. except for ln m∗ and γ. since P obs and Y obs inherit the unit root characteristics of their driving variables. Y obs (gam). 2 . Taking logs of both sides and subtracting the same equation scrolled back one period.70 CHAPTER 6. we must tell Dynare to use a diﬀuse prior (inﬁnite variance) for their initialization in the Kalman ﬁlter. any trends. the unconditional covariance matrix of these variables is used for initialization. In general.1.
0. alp. end. 0. 0. 1. 0.008862.10 Specifying the steady state Declaring the steady state is just as explained in details and according to the same syntax explained in chapter 3.m.5..12 Launching the estimation We add the following commands to ask Dynare to run a basic estimation of our model: . normal pdf. This is Dynare convention and must be the case for all your nonstationary variables. normal pdf.11 Declaring priors We expand our . nothing has changed with respect to the equivalent ﬁle of chapter 5. 0. 0. intuitive. deﬁning each variable in terms of parameters or variables solved in the lines above.65. inf. 0.035449. beta pdf. ALTERNATIVE AND NONSTATIONARY EXAMPLE 71 initial condition in the case of a deterministic trend. Second. 0. In this case.129. So far.223. The check=0 limits steady state values to real numbers.05. covering the initval. 0.1. beta pdf. 0. mst. beta pdf. P obs and Y obs which take the value of 1. 0.1. the output of the function is the endogenous variables at steady state. There are some things to notice. but tedious..993.02. 0. inv gamma pdf. The only novelty is the declaration of the nonstationary variables. 6.1.007. This functionality may be updated in later versions of Dynare. bet.0002.003.0085. inv gamma pdf. rho. we also discussed the usefulness of providing an external Matlab ﬁle to solve for the steady state. psi. steady and check commands. inf.6.01. The ﬁle is called fs2000ns steadystate. 0.002. gam.mod ﬁle with the following information: estimated params. section 5. 6. del. notice the declaration of parameters at the beginning. beta pdf. stderr e a. In chapter 5. since the variance is ﬁnite. Third. First. 6. you can ﬁnd the corresponding steady state ﬁle in the models folder under UserGuide.005.356. 0. the ys vector.1. stderr e m. beta pdf. 0. note that the ﬁle is really only a sequential set of equalities.
psi. . c+k = exp(alp*(gam+e a))*k(1)^alp*n^(1alp)+(1del) exp((gam+e a))*k(1).mod. mode compute=6. A simple linearization would fail as these variables do not have a steady state.mh nblocks=2. P/(c(+1)*P(+1)*m)+bet*P(+1)*(alp*exp(alp*(gam+log(e(+1)))) *k^(alp1)*n(+1)^(1alp)+(1del) *exp((gam+log(e(+1)))))/(c(+2)*P(+2)*m(+1))=0. P obs/P obs(1) = (p/p(1))*m(1)/dA. del.65). mst. For convenience.mod separately. The ﬁle is called fs2000ns. m1+d = l. rho.mod ﬁle We have seen each part of the . W = l/n. varobs P obs Y obs.45.ADVANCED TOPICS estimation(datafile=fsdat. Fortunately.72 CHAPTER 6. varexo e a e m. ESTIMATING DSGE MODELS . var m P c e W R k d n l Y obs P obs y dA. since it contains nonlinear equations in nonstationary variables.1. we need to instruct Dynare to loglinearize our model. log(m) = (1rho)*log(mst) + rho*log(m(1))+e m. dA = exp(gam+e a). bet.nobs=192. it’s now time to get a picture of what the complete ﬁle looks like. gam. R = P*(1alp)*exp(alp*(gam+e a))*k(1)^alp*n^(alp)/W. 1/(c*P)bet*P*(1alp)*exp(alp*(gam+e a))*k(1)^alp*n^(1alp)/ (m*l*c(+1)*P(+1)) = 0. NOTE! As mentioned earlier. P*c = m.loglinear.mh replic=2000. parameters alp.13 The complete .mh drop=0. end. Y obs/Y obs(1) = dA*y/y(1). y = k(1)^alp*n^(1alp)*exp(alp*(gam+e a)). 6. (psi/(1psi))*(c*P/(1n))+l/n = 0. e = exp(e a).mh jscale=0. model. the ﬁle also appears in the models folder under UserGuide in your Dynare installation. taking the log of the equations involving nonstationary variables does the job of linearizing them.
bet.003.035449. normal pdf. dA = exp(gam). e = 1. beta pdf.0085. beta pdf.mh nblocks=2. P obs (log(mst)gam).002.356. 0. gam. d = 0. end. stderr e m. inf. del.008862. 0. 0.1. end. c = 0. 1.01. . 0. 0. y = 0. 0.02. 0. 0.mh drop=0. 0.02. steady.loglinear.05. since the estimation will use the Matlab // steady state file also provided and discussed above. mode compute=6.007. alp. 0.005.993. P = 2. l = 0. m = mst. ALTERNATIVE AND NONSTATIONARY EXAMPLE observation trends. n = 0.6. 0. unit root vars P obs Y obs. k = 6. end. normal pdf.6.45. beta pdf. beta pdf.25. stderr e a.19.mh replic=2000. W = 4.0002.86. 73 // the above is really only useful if you want to do a stoch simul // of your model. 0. 0. inf. mst. Y obs (gam).85.mh jscale=0. rho. 0. inv gamma pdf. 0. estimated params. estimation(datafile=fsdat.223.65.65). beta pdf.129. inv gamma pdf. R = 1.45.nobs=192. initval. psi.
This is done in ﬁgure 6.74 CHAPTER 6. to gain a “bird’s eyeview” on what we have just accomplished.14.14 Summing it up The explanations given above of each step necessary to translate the Schorfheide (2000) example into language that Dynare can understand and process was quite lengthy and involved a slew of new commands and information. and summarize the most important steps at a high level. 6.ADVANCED TOPICS Figure 6.2: At a high level. 6. It may therefore be useful.1.1. there are ﬁve basic steps to translate a model into Dynare for successful estimation. ESTIMATING DSGE MODELS .2 Comparing models based on their posterior distributions ** TBD .
3. . The Reference Manual overviews the complete set of potential output ﬁles and describes where you can ﬁnd each one.6. depending on the instructions you give Dynare.3 Where is your output stored? The output from estimation can be extremely varied. WHERE IS YOUR OUTPUT STORED? 75 6.
.
note that in this chapter we will focus on stochastic models . they disappear. 77 . or read Collard and Juillard (2001a) or SchmittGrohe and Uribe (2004) which gives a good overview of the most recent solution techniques based on perturbation methods. you may want to go through Michel Juillard’s presentation on solving DSGE models to a ﬁrst and second order (available on Michel Juillard’s website).2 What is the advantage of a second order approximation? As noted in chapter 3 and as will become clear in the section below. see Juillard (1996). and when expectations are taken. or will spur your curiosity to have a look at more detailed material.1 of chapter 3. Going into details would be beyond the scope of this User Guide which will instead remain at a high level. to get an idea of the methodologies and algorithms used in its computations.1. What you will ﬁnd below will either comfort you in realizing that Dynare does what you expected of it and what you would have also done if you had had to code it all yourself (with a little extra time on your hands!).Chapter 7 Solving DSGE models Behind the scenes of Dynare 7. Thus. For more details on the NewtonRaphson algorithm used in Dynare to solve deterministic models. 7. This is because only the ﬁrst moments of the shocks enter the linearized equations.1 Introduction The aim of this chapter is to peer behind the scenes of Dynare. as explained in section 3. or under its hood.which is where the major complication lies. Finally. If so. linearizing a system of equations to the ﬁrst order raises the issue of certainty equivalence.
For instance. it may instead be quite misleading. It is therefore very convenient that Dynare allows you to choose between a ﬁrst or second order linearization of your model in the option of the stoch simul command. you also need second order approximations of the policy function.3 How does dynare solve stochastic DSGE models? In this section. in the case of asset pricing models. ut )} = 0 E(ut ) = 0 E(ut ut ) = Σu and where: y : vector of endogenous variables of any dimension u : vector of exogenous stochastic shocks of any dimension The solution to this system is a set of equations relating variables in the current period to the past state of the system and current shocks. yt−1 . SOLVING DSGE MODELS . 7.BEHIND THE SCENES OF 78 DYNARE unconditional expectations of the endogenous variables are equal to their nonstochastic steady state values. To summarize. a DSGE model is a collection of ﬁrst order and equilibrium conditions that take the general form: Et {f (yt+1 . Yet more clearly. although at a higher level of complexity. yt . But depending on the context. The summary below is taken mainly from Michel Juillard’s presentation “Computing ﬁrst order approximations of DSGE models with Dynare”. that satisfy the original system. This is what we call the policy function. when using second order welfare functions to compare policies. The second order follows very much the same approach. This may be an acceptable simpliﬁcation to make. we shall brieﬂy overview the perturbation methods employed by Dynare to solve DSGE models to a ﬁrst order approximation. ut ) .CHAPTER 7. especially regarding second order approximations (available on Michel Juillard’s website). which you should read if interested in particular details. linearizing to the second order enables you to take risk (or the variance of shocks) into consideration . we can write this function as: yt = g(yt−1 . Sticking to the above notation.a highly desirable modeling feature.
This is technically why certainty equivalence holds . yt−1 . ut . 0) = 0 y ¯ ¯ having the property that: y = g(¯. y . they drop out when taking expectations of the linearized equations. ut+1 ) 79 We can then deﬁne a new function F . y ) + fy+ gy (gy y + gu u) + gu u y ¯ ¯ ˆ +fy0 (gy y + gu u) + fy− y + fu u ˆ ˆ =0 with y = yt−1 − y . ut+1 ) = Et f (¯. ∂f ∂yt+1 . ut . y . ut ). fy− = ∂f ∂yt−1 . 0) ¯ y The ﬁrst order Taylor expansion around y yields: ¯ Et F (1) (yt−1 . y . fy+ = ˆ ¯ fu = ∂f ∂ut . u = ut+1 . ut+1 ) = f (¯. ut+1 ). ut . fy0 = ∂f ∂yt . it is straightforward to rewrite yt+1 as yt+1 = g(yt . u = ut . y . ut ). gu = ∂g ∂ut . since future shocks only enter with their ﬁrst moments (which are zero in expectations). ut+1 ) = g(g(yt−1 . ut ). such that: F (yt−1 . y ) + fy+ (gy (gy y + gu u)) y ¯ ¯ ˆ +fy0 (gy y + gu u) + fy− y + fu u ˆ ˆ = fy+ gy gy + fy0 gy + fy− y + fy+ gy gu + fy0 gu + fu u ˆ = 0 As you can see. ut+1 )] = 0 We then venture to linearize this model around a steady state deﬁned as: f (¯.3. Taking expectations (we’re almost there!): Et F (1) (yt−1 . g(yt−1 . gy = ∂g ∂yt−1 . ut ) which enables us to rewrite our system in terms of past variables. ut . HOW DOES DYNARE SOLVE STOCHASTIC DSGE MODELS? Then. ut+1 ) = f (g(g(yt−1 .7. and current and future shocks: Et [F (yt−1 .
one of the conditions that comes out of the solution of this equation is the BlanchardKahn condition: there must be as many roots larger than one in modulus as there are forwardlooking variables in the model. we have solved for the (approximate) policy (or decision) function and have succeeded in solving our DSGE model. yet applies more complex algebraic techniques to recover the various partial derivatives of the policy function. Since the above equation holds for any y and any u. SOLVING DSGE MODELS . we would simply iterate the policy function starting from an initial value given by the steady state. Note that in the case of a second order approximation of a DSGE model. Finally.CHAPTER 7. recovering gu is then straightforward from the second parenthesis. for instance. The second order solution uses the same “perturbation methods” as above (the notion of starting from a function you can solve . Having recovered gy .like a steady state and iterating forward).BEHIND THE SCENES OF 80 DYNARE in a system linearized to its ﬁrst order. which we can solve with a series of algebraic trics that are not all immediately apparent (but detailed in Michel Juillard’s presentation). The second thing to note is that we have two unknown variables in the above equation: gy and gu each of which will help us recover the policy function g. each parenthesis must ˆ be null and we can solve each at a time. the variance of future shocks remains after taking expectations of the linearized equations and therefore aﬀects the level of the resulting policy function. Incidentally. But the general approach is perfectly isomorphic. yields a quadratic equation in gy . If we were interested in impulse response functions. The ﬁrst. . notice that a ﬁrst order linearization of the function g yields: yt = y + gy y + gu u ¯ ˆ And now that we have gy and gu .
Recent papers have attracted signiﬁcant attention.Chapter 8 Estimating DSGE models Behind the scenes of Dynare This chapter focuses on the theory of Bayesian estimation. It begins by motivating Bayesian estimation by suggesting some arguments in favor of it as opposed to other forms of model estimation. some of these include: Schorfheide (2000) which uses Bayesian methods to compare the ﬁt of two competing DSGE models of consumption. estimating the posterior function thanks to the MetropolisHastings algorithm. FernandezVillaverde and RubioRamirez (2004) which reviews the econometric properties of Bayesian estimators and compare estimation results with maximum likelihood and BVAR methodologies. Ireland (2004) which emphasizes instead maximum likelihood estimation. To do so.1 Advantages of Bayesian estimation Bayesian estimation is becoming increasingly popular in the ﬁeld of macroeconomics. It then attempts to shed some light on what goes on in Dynare’s machinery when it estimates DSGE models. 8. Lubik and Schorfheide (2005) which applies Bayesian estimation methods to an open macro model focussing on issues of misspeciﬁcation and identiﬁcation. including deﬁning what are prior and posterior distributions. and ﬁnally Rabanal and RubioRamirez (2005) which compares the ﬁt. Smets and Wouters (2003) which applies Bayesian estimation techniques to a model of the Eurozone. based on posterior distributions. 81 . of four competing speciﬁcations of New Keynesian monetary models with nominal rigidities. using the Kalman ﬁlter to ﬁnd the likelihood function. and comparing models based on posterior distributions. Lubik and Schorfheide (2003) which investigates whether central banks in small open economies respond to exchange rate movements. this section surveys the methodologies adopted for Bayesian estimation.
in the structural equations. as opposed to GMM estimation which is based on particular equilibrium relationships such as the Euler equation in consumption. and discussions with. Bayesian estimation ﬁts the complete. ESTIMATING DSGE MODELS . Bayesian estimation explicitly addresses model misspeciﬁcation by including shocks. St´phane Adjemian. estimation in the Bayesian case is based on the likelihood generated by the DSGE system. if your model is entirely misspeciﬁed. is discussed more technically in the subsection below. More technically. the problem arises when the posterior distribution is ﬂat over a subspace of parameter values. rather than the more indirect discrepancy between the implied DSGE and VAR impulse response functions. the posterior distribution corresponding to competing models can easily be used to determine which model best ﬁts the data. when estimating a model. Fourth. due to the stylized and often misspeciﬁed nature of DSGE models. Third. It can be summarized by diﬀerent values of structural parameters leading to the same joint distribution for observables. Of course. a member of the Dynare development e . 8. Indeed. Second. estimating it using Bayesian techniques could be a disadvantage. First. This procedure. Sixth. solved DSGE model. Unfortunately. as other topics mentioned above.2 The basic mechanics of Bayesian estimation This and the following subsections are based in great part on work by. the problem of identiﬁcation often arises. Bayesian techniques allow the consideration of priors which work as weights in the estimation process so that the posterior distribution avoids peaking at strange points where the likelihood peaks. Bayesian estimation naturally leads to the comparison of models based on ﬁt. but ﬁve of these stand out as particularly important and general enough to mention here. which can be interpreted as observation errors. But the weighting of the likelihood with prior densities often leads to adding just enough curvature in the posterior distribution to facilitate numerical maximization. Likewise.BEHIND THE SCENES 82 OF DYNARE There are a multitude of advantages of using Bayesian methods to estimate a model.CHAPTER 8. Indeed. the inclusion of priors also helps identifying parameters. leading to the “dilemma of absurd parameter estimates”. the likelihood often peaks in regions of the parameter space that are contradictory with common observations.
θ A . introduction to the basic mechanics of Bayesian estimation. remember to also check the open online examples of the Dynare website for examples of . these two building blocks . At its most basic level. And the maximum likelihood approach enters through the estimation process based on confronting the model with data.priors and likelihood functions . which includes a clear and quite complete introduction to Bayesian estimation. In the end. Let’s see how. First. although summarized in presentation format. Other helpful material includes An and Schorfheide (2006). or uniform function. A) ≡ p(YT θ A . given the model and its parameters: L(θ A YT . and on the other. we have a prior density p(θ) on the one hand. Bayesian estimation is a bridge between calibration and maximum likelihood. Also.mod ﬁles touching on Bayesian estimation. a likelihood p(YT θ). the appendix of Schorfheide (2000) contains details as to the exact methodology and possible diﬃculties encountered in Bayesian estimation. The tradition of calibrating models is inherited through the speciﬁcation of priors.8. the likelihood function describes the density of the observed data.2. You may also want to take a glance at Hamilton (1994). although somewhat outdated. Using the Bayes theorem . illustrated by the application of a simple DSGE model. Finally. A) t=1 p(yt Yt−1 . Generally speaking. generalized beta. gamma. More technically. we are interested in p(θYT ).are tied together by Bayes’ rule. A) where YT are the observations until period T . which provides a very clear. Finally. A) = p(y0 θ A . from example ﬁles to lecture notes to related papers. Second. chapter 12. THE BASIC MECHANICS OF BAYESIAN ESTIMATION 83 team. beta. and where in our case the likelihood is recursive and can be written as: T p(YT θ A . priors can be seen as weights on the likelihood function in order to give more importance to certain areas of the parameter subspace. shifted gamma. A) We now take a step back. Some of this work. the websites of Frank Schorfheide and Jesus FernandezVillaverde contain a wide variety of very helpful material. the posterior density. θ A represents the parameters of model A. Together. inverse gamma. p(•) stands for a probability density function (pdf) such as a normal. priors are described by a density function of the form p(θ A A) where A stands for a speciﬁc model. is available in the “Events” page of the Dynare website.
an example Suppose a data generating process yt = µ + εt for t = 1. YT ) ⇔ p(θ. corresponds to the numerator of the posterior density: p(θ A YT . A)p(θ A A) p(YT A) where p(YT A) is the marginal density of the data conditional on the model: p(YT A) = ΘA p(θ A .BEHIND THE SCENES 84 OF DYNARE twice we obtain this density of parameters knowing the data. where εt ∼ N (0.2. 1) is gaussian white noise. Then. we can combine the prior density and the likelihood function discussed above to get the posterior density: p(θ A YT . A)p(θ A A) ≡ K(θ A YT . we have p(θ. A) = p(YT θ A . Generally. The trick will be to estimate the likelihood function with the help of the Kalman ﬁlter and then simulate the posterior kernel using a samplinglike or Monte Carlo method such as the MetropolisHastings. though.T ] = 1 T.. the subsection below gives a simple example based on the above reasoning of what we mean when we say that Bayesian estimation is “somewhere in between calibration and maximum likelihood estimation”. Before moving on. These topics are covered in more details below. A) ∝ p(YT θ A .T = 1 T T 1 T 2 t=1 (yt −µ) T t=1 yt ≡ y and that V[µM L. ESTIMATING DSGE MODELS . . the likelihood is given by p(YT µ) = (2π)− 2 e− 2 We know from the above that µM L. although other similar examples can be found in Hamilton (1994)..CHAPTER 8.. T . YT ) = p(YT θ) × p(θ) p(θ) By using these identities. chapter 12. . YT A)dθ A Finally. YT ) p(θYT ) = p(YT ) We also know that p(YT θ) = p(θ. A) This is the fundamental equation that will allow us to rebuild all posterior moments of interest.1 Bayesian estimation: somewhere between calibration and maximum likelihood estimation . 8. The example is drawn from Zellner (1971). given that the marginal density above is a constant or equal for any parameter). the posterior kernel (or unnormalized posterior density.
ut )} = 0. In more appropriate terms for what follows. In particular. yt−1 . Only yt is observable. by: p (µYT ) ∝ 1 1 − 2 (2πσµ )− 2 e 2 (µ−µ0 )2 2 σµ × (2π)− 2 e− 2 T 1 T 2 t=1 (yt −µ) Or equivalently. up to a constant. so we just estimate the model) then E[µ] → µM L. if σµ → ∞ (ie. we have no prior information. which is really a collection of ﬁrst order and equilibrium conditions. ut ).3. thus leaving no room for estimation) then E[µ] → µ0 . But if σµ → 0 (ie. DSGE MODELS AND BAYESIAN ESTIMATION 85 In addition. Other variables are described below. Then. which we call the decision rule.3 8. p (µYT ) ∝ e − (µ−E[µ])2 V[µ] . let our prior be a gaussian distribution with expectation µ0 2 and variance σµ . we can rewrite the solution to a DSGE model as a system in the following manner: ∗ yt = M y (θ) + M yt + N (θ)xt + ηt ¯ ˆ yt = gy (θ)ˆt−1 + gu (θ)ut ˆ y E(ηt ηt ) = V (θ) E(ut ut ) = Q(θ) where yt are variables in deviations from steady state. with 1 −2 + σµ V[µ] = and E[µ] = 1 −1 T 1 −1 −2 µM L. yt .1 DSGE models and Bayesian estimation Rewriting the solution to the DSGE model Recall from chapter 7 that any DSGE model. and it is related to the true . can be written in the form Et {f (yt+1 .T .T + σµ µ0 T −2 1 −1 + σµ T From this. Most of the time. we’re sure of ourselves and we calibrate the parameter of interest. y is the vector of steady ˆ ¯ state values and θ the vector of deep (or structural) parameters to be estimated. we’re somewhere in the middle of these two extremes. But the equation expresses a relationship among true endogenous variables that ∗ are not directly observed. the 2 maximum likelihood estimator. the prior mean. taking as a solution equations of the type yt = g(yt−1 .3. we can tell that the posterior mean is a convex combination of 2 the prior mean and the ML estimate. the posterior density is deﬁned. The second equation is the familiar decision rule mentioned above. 8.8.
see Hamilton (1994). The ﬁrst apparent problem. which is captured with N (θ)xt to allow for the most general case in which the trend depends on the deep parameters. the recursion follows ∗ vt = yt − y ∗ − M yt − N xt ¯ ˆ Ft = M Pt M + V Kt = gy Pt gy Ft−1 yt+1 = gy yt + Kt vt ˆ ˆ Pt+1 = gy Pt (gy − Kt M ) + gu Qgu For more details on the Kalman ﬁlter.BEHIND THE SCENES 86 OF DYNARE variables with an error ηt . . Furthermore. From the Kalman ﬁlter recursion. they are linear in the endogenous and exogenous variables so that the likelihood may be evaluated with a linear prediction error algorithm like the Kalman ﬁlter. The ﬁrst and second equations above therefore naturally make up a system of measurement and transition or state equations. chapter 13. V (θ) and ∗ ∗ Q(θ) and where YT expresses the set of observable endogenous variables yt found in the measurement equation. The loglikelihood above gets us one step closer to our goal of ﬁnding the posterior distribution of our parameters. respectively. ESTIMATING DSGE MODELS .3. it’s not a coincidence!).2 Estimating the likelihood function of the DSGE model The next logical step is to estimate the likelihood of the DSGE solution system mentioned above. Yet. here’s what the Kalman ﬁlter recursion does. it is possible to derive the loglikelihood given by T Tk 1 1 ∗ ln L (θYT ) = − ln(2π) − Ft  − vt Ft−1 vt 2 2 2 t=1 where the vector θ contains the parameters we have to estimate: θ. . 8. . and are also known. This is exactly what Dynare does.CHAPTER 8. recall. T and with initial values y1 and P1 given. For t = 1. the log posterior kernel can be expressed as ∗ ∗ ln K(θYT ) = ln L (θYT ) + ln p(θ) where the ﬁrst term on the right hand side is now known after carrying out the Kalman ﬁlter recursion. though. as is typical for a Kalman ﬁlter (you guessed it. . As a reminder. are the priors. it may have a trend. The second. is that the equations are non linear in the deep parameters. Indeed. .
8. The general idea of the MetropolisHastings algorithm is to simulate the posterior distribution. but fortunately doable with modern computers.3. Draw a proposal θ ∗ from a jumping distribution J(θ ∗ θ t−1 ) = N (θ t−1 . and run a loop over 234. The distribution will be given by the kernel equation above.3 Finding the mode of the posterior distribution Next. 2. to samplinglike methods. where this is typically the posterior mode.3. cΣm ) where Σm is the inverse of the Hessian computed at the posterior mode.4 Estimating the posterior distribution Finally. This is indeed the method adopted by Dynare. this maximization problem is not completely straightforward. Thus. DSGE MODELS AND BAYESIAN ESTIMATION 87 8. we are instead more often interested in the mean and variance of the estimators of θ. To do so. the algorithm builds on the fact that under general conditions the distribution of the deep parameters will be asymptotically normal. It is a “rejection sampling algorithm” used to generate a sequence of samples (also known as a “Markov Chain” for reasons that will become apparent later) from a distribution that is unknown at the outset. . the MetropolisHastings algorithm implements the following steps: 1. instead. we are now in a position to ﬁnd the posterior distribution of our parameters. of which the MetropolisHastings has been retained in the literature as particularly eﬃcient. to ﬁnd the mode of the posterior distribution . The algorithm.a key parameter and an important output of Dynare . We resort. More precisely. in the words of An and Shorfheide. 8. 18). Recall that the likelihood function is not Gaussian with respect to θ but to functions of θ as they appear in the state equation. Choose a starting point θ ◦ . we cannot obtain an explicit form for it. but again. Thus. Remember that all we have is the posterior mode. it is a nonlinear and complicated function of the deep parameters θ. This allows for an eﬃcient exploration of the posterior distribution at least in the neighborhood of the mode” (An and Schorfheide (2006). p.we simply maximize the above log posterior kernel with respect to θ. “constructs a Gaussian approximation around the posterior mode and uses a scaled version of the asymptotic covariance matrix as the covariance matrix for the proposal distribution.3. This is done in Dynare using numerical methods.
If the acceptance ratio is greater than one. decide whether or not to hold on to your candidate parameter. Update the mean of your drawing distribution. choose a candidate paramter. the point is for each “bucket” of the histogram to shrink to zero. We should not be too quick to simply throw out the candidate giving a lower value of the posterior kernel.3. whose mean has been set to θt−1 (this will become clear in just a moment). 1) otherwise. the acceptance rate (the fraction of candidate parameters that are accepted in a window of time) will be too high and the Markov Chain of candidate parameters will “mix slowly”. θ∗ from a Normal distribution. the acceptance rate will be very low (as the candidates are likely to land in regions of low probability density) and the chain will spend too much . Then. Of course. the idea is to allow the search to turn away from taking a small step up. Otherwise. Finally accept or discard the proposal θ ∗ according to the following rule. if necessary.4 tries to clarify the above.BEHIND THE SCENES 88 OF DYNARE 3. This “smoothed histogram” will eventually be the posterior distribution after suﬃcient iterations of the above steps. Of course. then deﬁnitely keep your candidate.CHAPTER 8. build a histogram of those retained values. But why have such a complicated acceptance rule? The point is to be able to visit the entire domain of the posterior distribution. After having repeated these steps often enough. On the other hand. In step 3. Figure 8. meaning that the distribution will take a long time to converge to the posterior distribution since the chain is likely to get “stuck” around a local maximum. and compare it to the value of the kernel from the mean of the drawing distribution. and update. Compute the acceptance ratio r= K(θ ∗ YT ) p(θ ∗ YT ) = p(θ t−1 YT ) K(θ t−1 YT ) 4. Metaphorically. do two things. and note the value of the parameter your retain. and instead take a few small steps down in the hope of being able to take a big step up in the near future. In step 2. If the scale factor is too small. an important parameter in this searching procedure is the variance of the jumping distribution and in particular the scale factor. compute the value of the posterior kernel for that candidate parameter. go back to the candidate of last period (this is true in very coarse terms. the jumping distribution: θt = θ∗ θ t−1 with probability min(r. in the ﬁnal step. just in case using that candidate for the mean of the drawing distribution allows us to to leave a local maximum and travel towards the global maximum. In step 1. ESTIMATING DSGE MODELS . notice that in fact you would keep your candidate only with a probability less than one). if the scale factor is too large.
. Imagine repeating these steps a large number of times. These include: How should we choose the scale factor c (variance of the jumping distribution)? What is a satisfactory acceptance rate? How many draws are ideal? How is convergence of the MetropolisHastings iterations assessed? These are all important questions that will come up in your usage of Dynare. time in the tails of the posterior distribution.1: The above sketches the MetropolisHastings algorithm. at least to a machine needing to undertake the above calculations. used to build the posterior distribution function. and smoothing the “histogram” such that each “bucket” has inﬁnitely small width.7 of Chapter 5. several practical questions arise when carrying out Bayesian estimation. DSGE MODELS AND BAYESIAN ESTIMATION 89 Figure 8. While these steps are mathematically clear.3. They are addressed as clearly as possible in section 5.8.
p(YT I). especially for large samples. I)p(θ m I) I I I where θ m is the posterior mode. the comparison of the two models is done very naturally through the ratio of the posterior model distributions. specifying the basic mechanics of Bayesian estimation. is the Gaussian (called a Laplace approximation). while touting the advantages of Bayesian estimation. We call this the posterior odds ratio: p(AYT ) p(A) p(YT A) = p(BYT ) p(B) p(YT B) The only complication is ﬁnding the magrinal density of the data conditional on the model. The advantage of this technique is its comI putational eﬃciency: time consuming MetropolisHastings iterations are not k 1 . B we can evaluate. This requires some detailed explanations of their own. The ﬁrst is to assume a functional form of the posterior kernel that we can integrate. For each model I = A. The most straightforward and the best approximation. there are two options. In this case. the marginal density of the data conditional on the model by integrating out the deep parameters θ I from the posterior kernel: p(YT I) = ΘI p(θ I . you may want to glance back at the ﬁrst subsection above.CHAPTER 8. Using Bayes’ rule. where I = A. Suppose we have a prior distribution over two competing models: p(A) and p(B). we would have the following estimator: p(YT I) = (2π) 2 Σθm  2 p(θ m YT . Then.B p(I)p(YT I) where this formula may easily be generalized to a collection of N models. B p(IYT ) = p(I)p(YT I) I=A. To remind you of this. ESTIMATING DSGE MODELS . I)dθ I Note that the expression inside the integral sign is exactly the posterior kernel. which is also the denominator of the posterior density p(θYT ) discussed earlier. the posterior distribution oﬀers a particularly natural method of comparing models. we can compute the posterior distribution over models. To obtain the marginal density of the data conditional on the model. I)dθ I = ΘI p(θ I I) × p(YT θ I . YT θ I .4 Comparing models using posterior distributions As mentioned earlier.BEHIND THE SCENES 90 OF DYNARE 8. at least theoretically. Let’s look at an illustration.
8.4. COMPARING MODELS USING POSTERIOR DISTRIBUTIONS 91 necessary, only the numerically calculated posterior mode is required. The second option is instead to use information from the MetropolisHastings runs and is typically referred to as the Harmonic Mean Estimator. The idea is to simulate the marginal density of interest and to simply take an average of these simulated values. To start, note that p(YT I) = E f (θ I ) θI , I p(θ I I)p(YT θ I , I)
−1
where f is a probability density function, since E f (θ I ) θI , I = p(θ I I)p(YT θ I , I)
ΘI ΘI
f (θ)dθ
p(θ I I)p(YT θ I , I)dθ I
and the numerator integrates out to one (seeGeweke (1999) for more details). This suggests the following estimator of the marginal density 1 p(YT I) = B
(b) B (b) b=1
f (θ I ) p(θ I I)p(YT θ I , I)
(b)
(b)
−1
where each drawn vector θ I comes from the MetropolisHastings iterations and where the probability density function f can be viewed as a weights on the posterior kernel in order to downplay the importance of extreme values of θ. Geweke (1999) suggests to use a truncated Gaussian function, leading to what is typically referred to as the Modiﬁed Harmonic Mean Estimator.
Chapter 9
Optimal policy under commitment
93
.
Thanks for your precious help! 95 .Chapter 10 Troubleshooting To make sure this section is as user friendly as possible. how Dynare tells you about it and how you solve it. the best is to compile what users have to say! Please let me know what your most common problem is with Dynare.
.
and S. J. J.” Econometric Review.Bibliography An. 97 . and M. F. 1–126. Geweke. (1994): Time Series Analysis. NJ. Juillard (2001a): “Accuracy of stochastic perturbation methods: The case of asset pricing models.” Journal of Economic Dynamics and Control. J. and J. Ireland. 1661–1707.. Forthcoming. Hamilton.” Journal of Economic Literature. 979–999.. (1999): “Using Simulation Methods for Bayesian Econometric Models: In.” Journal of Economic Dynamics and Control. S. Princeton. Koopman (2001): Time Series Analysis by State Space Methods. (2003): “Stochastic simulations with DYNARE.ference.” CEPREMAP working papers 9602. (1996): “Dynare : a program for the resolution and simulation of dynamic models with forward variables through the use of a relaxation algorithm. FernandezVillaverde. A practical guide. Collard. 123(1). Gertler (1999): “The Science of Monetary Policy: A New Keynesian Perspective. R. Oxford University Press. P. Clarida. RubioRamirez (2004): “Comparing dynamic equilibrium models to data: a Bayesian approach. J. 1205–1226. Durbin.. Juillard. XXXVII. 153–187. 18(1). 25(67). M. 125–39. D.K. J..” Econometric Review. 28(6). Gali.” Journal of Econometrics.. CEPREMAP. (2001b): “A HigherOrder Taylor Expansion Approach to Simulation of Stochastic ForwardLooking Models with an Application to a Nonlinear Phillips Curve Model. U. Development and Communication. and F.. and M. N. (2004): “A method for taking models to the data. Schorfheide (2006): “Bayesian Analysis of DSGE Models. Oxford. 17(23). F. Princeton University Press.” Computational Economics.” CEPREMAP mimeo.
Nason.” Journal of Applied Econometrics. F.Department of Economics. New York.. Schorfheide (2003): “Do Central Banks Respond to Exchange Rate Movements? A Structural Investigation. M.” Journal of Applied Econometrics.. F. S.. and R. Wouters (2003): “An Estimated Dynamic Stochastic General Equilibrium Model of the Euro Area. Uribe (2004): “Solving dynamic general equilibrium models using a secondorder approximation to the policy function. 28(4). 1123–1175. Schorfheide. 9(S). and J. SchmittGrohe.. (2000): “Loss functionbased evaluation of DSGE models. 645–670. (2005): “A Bayesian Look at New Open Economy Macroeconomics. 755–775. Smets. The Johns Hopkins University. Inc. Cogley (1994): “Testing the Implications of LongRun Neutrality for Monetary Business Cycle Models. and T.98 BIBLIOGRAPHY Lubik. 52(6). J. 1(5). . F.” Journal of Monetary Economics. and M.” Journal of Economic Dynamics and Control. John Wiley & Sons.. The Johns Hopkins University. A. Zellner. P. Rabanal.. 1151–1166.” Journal of the European Economic Association. (1971): An Introduction to Bayesian Inference in Econometrics.” Economics Working Paper Archive 505. T. and F.” Economics Working Paper Archive 521. S37–70.Department of Economics. 15(6). RubioRamirez (2005): “Comparing New Keynesian models of the business cycle: A Bayesian approach.