Practical DSGE

Modelling
Alina Barnett
Martin Ellison

Bank of England, December 2005

1
 Practical DSGE modelling

Bank of England
December 2005

Overview

This is an introductory course for individuals with little or no previous knowledge of

the practicalities involved in dynamic stochastic general equilibrium modelling. The
intention of the course is not to train participants to become specialist model-builders
but instead for them to be fully conversant in the main techniques involved in building
and using dynamics models in their work. The aim is to make participants
sophisticated consumers of dynamic models and to provide a deeper framework and
knowledge within which to frame their discussions. Although no prior knowledge is
assumed it is a fast paced course and additional reading is provided to help
participants to consolidate material.

Course objectives

By the end of the course participants should:

Understand the economic foundations of dynamic models and how they relate
to current policy issues
Be able to write dynamic models in a general form suitable for solution and
simulation by computer
Implement basic solution and simulation techniques to analyse dynamic
models, showing how the model economies behave and how they respond to
different shocks
Be aware of how to extend simple models to answer more complex questions
associated with monetary policy
Feel confident in discussing model-based analysis, evaluating the strength of
models used and engaging in conversation with people building economic
models

Organisation

The course runs over a series of five mornings. Each session will be a mixture of
small-group teaching and practical exercises using MATLAB software to solve and
simulate dynamic models. Afternoons are therefore free, allowing participants time to
reflect and consolidate the morning work before the next session. The instructors will
be available throughout each afternoon to answer any questions or queries that arise.

Pre-readings

Some pre-course readings will be assigned for those participants who wish to prepare
in advance. The issues involved in the pre-reading will be dealt with in the course so
there is no need to understand everything. The pre-reading is as much about getting
participants to gauge what they do and dont understand as it is to get ideas and issues
across.

2
Ouline

Session 1 Introduction to DSGE modelling.

In this session we describe the key ingredients in a dynamic model, using the
simplest available dynamic model with monetary policy.

Exercises are designed to familiarise participants with the MATLAB

computer language environment.
Session 2 Taking a model to the computer

We show how to rewrite the model in a form suitable for solving and
simulating on a computer. There are two parts to this: (i) approximating the
nonlinear model by a linear model and (ii) collecting the equations together
into the so-called state-space form.

In exercises we practice the techniques of log-linearisation and define state-

space forms.
Session 3 Solution techniques

After a short discussion of the range of possibilities, we explain the standard

Blanchard-Kahn technique for solving linear models. We also discuss the role
of the Blanchard-Kahn conditions in ensuring determinacy of the solution.

The computer exercises take participants from the state-space form to the
solution to the model and show the relevance of determinacy conditions.
Session 4 Simulation techniques

Having obtained the solution in the previous session, this session is devoted
to developing a range of simulation techniques through which to analyse the
model. Amongst others, we discuss stylised facts of volatilities and
correlations at different leads and lags, impulse response analysis, forecast
error variance decomposition.

Computing exercises develop participants skills at analysing dynamic

models.
Session 5 Advanced dynamic models

In the final session we survey some of the more advanced dynamic models
that are useful for policy analysis. In particular, we focus on how to handle
models with simple monetary policy rules such as those famously associated
with Taylor.

The exercises in the final session consolidate the work of the previous
sessions. The participants will work through a series of exercises which
simulate the behaviour of a dynamic economy under Taylor rules. Time
permitting, participants will be able to analyse the response of the economy
for different coefficients in the monetary policy rules.

3
Instructors

The course architects are Martin Ellison and Alina Barnett.

Martin Ellison, Assistant Professor, Warwick University

Martin has an MPhil from Oxford University and a PhD from the European
University Institute, Florence. He is an affiliate of the Centre for Economic Policy
Research and a consultant at the Bank of England. His research focuses on monetary
theory and monetary policy and he has just received a three-year ESRC Research
Fellowship for his project Improving Monetary Policy for Macroeconomic Stability
in the 21st Century. His research is regularly published in the top academic journals,
as well as presented at leading conferences and central banks.

Alina Barnett, Postgraduate Research Fellow, Warwick University

Alina holds an MA in Economics from the Central European University in Budapest

and a BA in Finance and Banking from the Academy of Economic Studies in
Bucharest. Her research interests are in macroeconomics, time series econometrics
and developing economies. Her professional experience includes financial analysis,
banking and teaching economics at university and secondary school level.

4
 Practical DSGE
modelling
Alina Barnett
Martin Ellison

Bank of England, December 2005

Objective

To make participants sophisticated consumers

of dynamic stochastic general equilibrium
models, and to provide a deeper framework
and knowledge within which to frame
discussions of economic policy issues.

5
Aims

Understanding of simple DSGE models

Ability to solve and simulate simple DSGE

models using MATLAB

Organisation

Five mornings

Each morning is a mixture of small-group teaching and

practical exercises using MATLAB

Share of teaching is higher in first couple of days

Course organisers are available each afternoon to give

extra help and answer questions

6
Outline

Day Topics

1 Introduction to DSGE models / Introduction to

MATLAB
2 Writing models in a form suitable for computer

3 Solution techniques

4 Simulation techniques

5 Advanced dynamic models

Introduction to
DSGE modelling
Martin Ellison
University of Warwick and CEPR

Bank of England, December 2004

7
 Dynamic
Stochastic
General Equilibrium

Dynamic

t-1 t t+1

expectations

8
Stochastic
Frisch-Slutsky paradigm

Impulses

Propagation

Fluctuations

General equilibrium

Monetary
authority

Firms Households

9
Households

Maximise present discounted value of expected

utility from now until infinite future, subject to
budget constraint

Households characterised by
utility maximisation
consumption smoothing

Households
We show household consumption behaviour in a
simple two-period deterministic example with no
uncertainty

initial wealth W0
consumption C0 and C1
prices p0 and p1
nominal interest at rate i0 on savings from t0 to t1

Result generalises to infinite horizon stochastic

problem with uncertainty

10
Household utility
C
1 max U (C ) + U (C )
C0 0 1

dU = 0 = U ' (C )dC + U ' (C )dC

0 0 1 1

dC1 U ' (C )
= 0
U dC0 U ' (C )
1
C
0

Household budget constraint

C
1
p1C1 = (W0 p0C0 )(1 + i0 )

p1dC1 = p0 dC0 (1 + i0 )

dC1 1 + i0
=
dC0 1 + 1

C
0

11
Household utility maximisation
C
1

U ' (C )
0 = 1 + i0
U ' (C ) 1 + 1
1

C
0

Households

General solution for stochastic -horizon case

1 + it
U ' (C ) = Et U ' (C
t + 1 1 + t +1
)
t

Known as the dynamic IS curve

Known as the Euler equation for consumption

12
Households - intuition

1 + it
U ' (C ) = Et U ' (C
t + 1 1 + t +1
)
t

it U(Ct) Ct Higher interest rates

reduce consumption
Ett+1 U(Ct) Ct Higher expected future
inflation increases
consumption

Firms

Maximise present discounted value of expected

profit from now until infinite future, subject to
demand curve, nominal price rigidity and labour
supply curve.

Firms characterised by
profit maximisation
subject to nominal price rigidity

13
Firms

Firm problem is mathematically complicated

(see Walsh chapter 5)

We present heuristic derivation of the results

Nominal price rigidity

Calvo model of price rigidity

Proportion of firms able to

change their price in a period
1

Proportion of firms unable to

change their price in a period

14
Aggregate price level

p t = (1 ) p it + p t 1

price price
setters non-setters

Do not worry about the hat (^) notation. We will

explain it later

Derivation
p t = (1 ) p it + p t 1

15
Optimal price setting

p it = (1 ) p t* + Et p it +1

price myopic desired

set at t price price at t+1

=0 perfect price flexibility p it = p t*

p it Et p it +1
1 price inflexibility

Derivation
p t = (1 ) p it + p t 1

p it = (1 ) p t* + Et p it +1

16
Myopic price
Approximate myopic price with price that would
prevail in flexible price equilibrium
pt* = k pt mct

Price is constant mark-up k over marginal cost

In our hat (^) notation to be explained later
the myopic price is given by

p t* = p t + m ct

Full derivation
p t = (1 ) p it + p t 1

p it = (1 ) p t* + Et p it +1

p t* = p t + m ct

17
Marginal cost

No capital in model all marginal costs

due to wages

Assume linearity between wages and marginal

cost

m ct = w t

Derivation
p t = (1 ) p it + p t 1

p it = (1 ) p t* + Et p it +1

p t* = p t + m ct

m ct = w t

18
Wages
Assume a labour supply function

wages rise when wages rise

output is above trend with output gap

1/ is elasticity of wage w.r.t output gap

1
w t = xt

Full derivation
p t = (1 ) p it + p t 1

p it = (1 ) p t* + Et p it +1

p t* = p t + m ct

m ct = w t

1
w t = xt

19
Firms

Full solution

xt = ( t Et t +1 )
(1 )(1 )

Known as the New Keynesian Phillips curve

Known as the forward-looking Phillips curve

Firms - intuition

xt = ( t Et t +1 )
(1 )(1 )

( t - Ett+1) < 0 xt < 0 Inflation expected to

rise in future, firms set
high prices now,
choking supply
Ett+1 pit xt Higher expected future
inflation chokes supply

20
Monetary authority

Sets the interest rate

Simplest case is simple rule

Interest rate reacts to inflation, with shocks

it = t + vt

Baseline DSGE model

Monetary it = t + vt
authority

Firms Households

1 + it
xt = ( t Et t +1 ) U ' (C ) = Et U ' (C )
(1 )(1 ) t t + 1 1 + t +1

21
Next steps
Introduction to MATLAB

How to write the DSGE model in a format

suitable for solution

How to solve the DSGE model

Solving the DSGE model in MATLAB

22
 Introduction to
MATLAB

Martin Ellison
University of Warwick and CEPR

Bank of England, December 2005

What is MATLAB?

MATLAB is a tool for doing numerical

computations with matrices and vectors. It is very
powerful and easy to use. It integrates
computation, graphics and programming in the
same environment.

MATLAB stands for Matrix Laboratory.

23
Matrix

MATLAB works with essentially only one kind

of object a rectangular numerical matrix with
possible complex entries.

Entering a matrix

Matrices can be

Entered manually
Generated by built-in functions

24
An example

A = [1, 2, 3; 7, 8, 9]

Use ; to indicate the end of each row

Use comma to separate elements of a row

Matrix operations
+ addition
- subtraction
* multiplication
^ power
transpose

To make * and ^ operate element-by-element,

we write .* and .^

25
Example

A= [1, 2; 3, 4]
B = [0.5, 0.6; 1, 1.5]

C = A*B
C = A.*B

Subscripts

The element in row i and column j of A is

denoted by A(i, j).

Example: A = zeros(2,2);
A(1,1) + A(1,2) + A(2,2)

26
The colon operator
The colon : is one of MATLAB s most
important operators. It has many uses.

3:-2:-11 is a row vector containing integers

from 3 to -11 with a increment of -2.

Subscript expressions involving colons refer to

portions of a matrix. A(1:3, 2) is the first to the
third elements of the second column of A.

Working with matrices

MATLAB provides four functions that generate
basic matrices.

zeros: all zeros. A = zeros(1,3)

ones: all ones. A = ones(2,4)
rand: uniformly distributed random
numbers. A = rand(3,5)
randn: normally distributed random
numbers. A = randn(2,2)

27
Working with matrices

Concatenation: join small (compatible) matrices

to make bigger ones. B = [A A-2; A*2 A/4]

Deleting rows and columns. B(:,2) = [ ]

Functions
MATLAB provides a large range of standard
elementary mathematical functions, including abs,
sqrt, exp, and sin.

For help on functions, type

help elfun (elementary mathematical functions)

help specfun (advanced mathematical functions)
help elmat (advanced matrix functions)
help datafun (data analysis functions)

28
Suppressing output

If you simply type a statement and press Enter,

MATLAB automatically displays the results on
screen. If you end the line with a semicolon ;
MATLAB performs the computation but does
not display any result.

Example: C = randn(5,1)
C = randn(5,1);

Programming with MATLAB

Files that contain code in the MATLAB

language are called M-files. You create M-files
using a text editor, then use them as you would
any other MATLAB functions or command.

29
Flow Control

MATLAB has many flow controls. The most

basic are

if statement
for loops
while loops

if elseif else end

if A > B
greater
elseif A < B
less
elseif A = = B
equal
end

30
for end
for i = 1:m
for j = 1:n
H(i,j) = 1/(i+j)
end
end

while end
i = 0;
while (i<10000)
s = s + i;
i = i + 1;
end

31
Graphics
x = 0 : 0.01 : 100;
y = x^2;
plot(x,y)

Adding plots to an existing graph: hold on

Multiple plots in one figure: subplot

32
 Taking a model
to the computer
Martin Ellison
University of Warwick and CEPR

Bank of England, December 2005

Baseline DSGE model

Monetary it = t + vt
authority

Firms Households

1 + it
xt = ( t Et t +1 ) U ' (C ) = Et U ' (C )
(1 )(1 ) t t + 1 1 + t +1

33
Households

Two simplifying assumptions:

Ct1
CRRA utility function U (Ct ) = U ' (Ct ) = Ct
1

No capital Ct = Yt

Dynamic IS curve

1 + it
Yt = Et Yt +1
1 + t +1

Non-linear relationship

Difficult for the computer to handle

We need a simpler expression

34
Log-linear approximation

Begin by taking logarithms of dynamic IS curve

1 + it
ln Yt = ln + ln Et Yt +1
1 + t +1

Problem is last term on right hand side

Properties of logarithms
Taylor series expansion of logarithmic function
z 2 z3 z 4
ln(1 + z ) = z + + ... 1 < z < 1
2 3 4

To a first order (linear) approximation

ln E (1 + z ) E ( z ) E ln(1 + z )

Applied to dynamic IS curve

1 + it 1 + it
ln Et Yt +1 Et ln Yt +1
1 + t +1 1 + t +1

35
Log-linearisation
Log-linear expansion of dynamic IS curve
ln Yt = ln Et ln Yt +1 + ln(1 + it ) Et ln(1 + t +1 ) (1)

Steady-state values (more later)

ln Y = ln ln Y + ln(1 + i ) ln (2)
(1) (2)
(ln Yt ln Y ) = Et (ln Yt +1 ln Y ) + ln(1 + it ) Et (ln(1 + t +1 ))

Deviations from steady state

What is ln Yt ln Y ?

Zt Zt Z Z
ln Z t ln Z = ln 1 = t = Z t
Z Z Z

percentage deviation of
Zt from steady state Z

In case of output, ln Yt ln Y is output gap, xt

36
 Log-linearised IS curve

it Et t +1
xt = Et xt +1 1 (it Et t +1 )
Et xt +1

Slope = -

xt

Advanced log-linearisation
The dynamic IS curve was relatively easy to
log-linearise

For more complicated equations, need to apply

following formula

f x ( x, y ) x , y f y ( x, y )
ln f ( x, y ) ln f ( x , y ) + (x x) + ( y y)
x,y

f ( x, y) f (x, y)

37
Firms

Previously solved for firm behaviour directly in

log-linearised form. Original model is in Walsh
(chapter 5).

Aggregate price level

Original equation Log-linearised

version

pt1 = (1 ) pit1 + pt11 p t = (1 ) p it + p t 1

38
Optimal price setting

Original equation Log-linearised

version


P
Et C p t +i t +i
i i 1
t +i
*

Pit i =0 Pt p it = (1 ) p t* + Et p it +1
= 1
Pt 1
i i 1 Pt + i
Et Ct +i
i =0 Pt

Myopic price

Original equation Log-linearised

version

pt* = pt mct p t* = p t + m ct
1

39
Marginal cost

Original equation Log-linearised

version

Wt
= mct m ct = w t
Pt

Wages

Original equation Log-linearised

version

Wt N t
= 1
Pt Ct w t = xt

Ct = N t

40
Monetary authority

We assumed it = t + vt

Equivalent to ln it ln i = (ln t 1 ln ) + vt

Very similar to linear rule if it small

Log-linearised DSGE model

Monetary it = t + vt
authority

Firms Households

xt = ( t Et t +1 ) xt = Et xt +1 1 (it Et t +1 )
(1 )(1 )

41
Steady state

Need to return to original equations to calculate

steady-state
1
Assume i = 0 for monetary authority

1+ i
From household Y = Y
1 +
=0

Steady state calculation

From firm P = Pi = P *
W 1
mc = =
P
1
1 +
Y =

42
Full DSGE model

xt = Et xt +1 1 (it Et t +1 )
t = Et t +1 + xt
it = t + vt

(1 )(1 )
=

Alternative representation

Et xt +1 + 1 Et t +1 = xt + 1 t + 1vt
Et t +1 = xt + t

1 1 Et xt +1 1 1 xt 1
+
0 E = 0 vt
t t +1 1 t

43
State-space form

Generalised state-space form

A0 Et X t +1 = A1 X t + B0 vt +1
Models of this form (generalised linear rational
expectations models) can be solved relatively
easily by computer

Next steps

Derive a solution for log-linearised models

Blanchard-Kahn technique

44
 Solution techniques
Martin Ellison
University of Warwick and CEPR

Bank of England, December 2005

State-space form
Generalised state-space form

A0 Et X t +1 = A1 X t + B0 vt +1

Many techniques available to solve this class of

models

We use industry standard: Blanchard-Kahn

45
Alternative state-space form

Et X t +1 = A01 A1 X t + A01 B0 vt +1

A B

Et X t +1 = AX t + Bvt +1

Partitioning of model
backward-looking variables
predetermined variables
w
Xt = t
yt forward-looking variables
control variables

wt +1 w
E y = A t + Bvt +1
t t +1 yt

46
Jordan decomposition of A

wt +1 w
E y = A t + Bvt +1
t t +1 yt

A = PP 1

eigenvectors diagonal matrix

of eigenvalues

Blanchard-Kahn condition

The solution of the rational expectations model

is unique if the number of unstable eigenvectors
of the system is exactly equal to the number of
forward-looking (control) variables.

i.e., number of eigenvalues in greater than 1

in magnitude must be equal to number of
forward-looking variables

47
Too many stable roots

yt multiple solutions

equilibrium path
not unique

need alternative
w0 wt techniques

Too many unstable roots

yt no solution

all paths are

explosive

transversality
w0 wt conditions violated

48
Blanchard-Kahn satisfied

yt one solution

equilibrium path
is unique

system has saddle

w0 wt path stability

Rearrangement of Jordan form

wt +1 1 t
w
E y = P P y + Bvt +1
t t +1 t
wt +1
1 1 t
w
P = P + P 1 Bvt +1
Et yt +1 yt

49
Partition of model

w w
P 1 t +1 = P 1 t + Rvt +1
Et yt +1 yt

stable

0 P* P12* R
= 1 P 1 = 11* R = 1
0 2 P21 P22* R2

unstable

Transformed problem

P11* P12* wt +1 1 0 P11* P12* wt R1

* = + vt +1
P
21 P22* Et yt +1 0 2 P21* P22* yt R2

~ ~
w
w t
Et ~t +1 ~
yt +1 t
y

~
P11* wt + P12* yt = w ~
w 0 w ~ R
t +1
= 1 ~t + 1 vt +1
t
E ~ 2 yt R2
P w +P y = ~
*
21 t
*
22 y t t t yt +1 0

50
Decoupled equations
~
w 0 w ~ R
t +1
E ~ = 1 ~t + 1 vt +1
t yt +1 0 2 yt R2

~ = w ~
wt +1 1 t + R1vt +1 stable

Et ~
yt +1 = 2 ~
yt + R2 vt +1 unstable

Decoupled equations can be solved separately

Solution strategy

Solve unstable ~
yt
transformed equation

Solve stable ~
transformed equation wt

Translate back into wt

original problem y
t

51
Solution of unstable equation
Solve unstable equation forward to time t+j
yt + j = ( 2 ) ~
Et ~
j
yt

As 2 > 1 , only stable solution is ~yt = 0 t

~
yt = P21* wt + P22* yt = 0 yt = P22*1 P21* wt

Forward-looking (control) variables are function

of backward-looking (predetermined) variables

Solution of stable equation

Solve stable equation forward to time t+j
~ = ( ) j w
Et w ~
t+ j 1 t

As 1 < 1 , no problems with instability

~ = P* w + P* y
wt 11 t 12 t ~ = ( P * P* P *1 P * ) w
w
yt = P22*1 P21* wt
t 11 12 22 21 t

52
Solution of stable equation
~ = w ~
wt +1 1 t + R1vt +1
~
( P11* P12* P22*1 P21* ) wt +1 = w ~ = ( P * P * P *1 P * ) w
w
t +1 t 11 12 22 21 t

wt +1 = ( P11* P12* P22*1 P21* ) 1 1 ( P11* P12* P22*1 P21* ) wt

+ ( P11* P12* P22*1 P21* ) 1 R1vt +1

Future backward-looking (predetermined)

variables are function of current backward-
looking (predetermined) variables

Full solution

yt = P22*1 P21* wt

wt +1 = ( P11* P12* P22*1 P21* ) 1 1 ( P11* P12* P22*1 P21* ) wt

+ ( P11* P12* P22*1 P21* ) 1 R1vt +1

All variables are function of backward-looking

(predetermined) variables: recursive structure

53
Baseline DSGE model
State space form

1 1 Et xt +1 1 1 xt 1
+
0 E = 0 vt
t t +1 1 t

To make model more interesting, assume

policy shocks vt follow an AR(1) process

vt +1 = vt + t +1

New state-space form

1 0 0 vt +1 0 0 vt 1
1
0 1 Et xt +1 = 1 1 xt + 0 t +1
1

0 0 E 0 1 t 0
t t +1

One backward-looking variable vt

Two forward-looking variables xt , t

54
Blanchard-Khan conditions

Require one stable root and two unstable roots

wt = vt
Partition model according to x
yt = t
t

Next steps

Exercise to check Blanchard-Kahn conditions

numerically in MATLAB

Numerical solution of model

Simulation techniques

55
Simulation techniques
Martin Ellison
University of Warwick and CEPR

Bank of England, December 2005

Baseline DSGE model

yt = P22*1 P21* wt

wt +1 = ( P11* P12* P22*1 P21* ) 1 1 ( P11* P12* P22*1 P21* ) wt

+ ( P11* P12* P22*1 P21* ) 1 R1vt +1

Recursive structure makes model easy to

simulate

56
Numerical simulations

Stylised facts

Impulse response functions

Forecast error variance decomposition

Stylised facts
Variances

Covariances/correlations

Autocovariances/autocorrelations

Cross-correlations at leads and lags

57
Recursive simulation
1. Start from steady-state value w0 = 0

2. Draw shocks {vt} from normal distribution

3. Simulate {wt} from {vt} recursively using

wt +1 = ( P11* P12* P22*1 P21* ) 1 1 ( P11* P12* P22*1 P21* ) wt
+ ( P11* P12* P22*1 P21* ) 1 R1vt +1

Recursive simulation

4. Calculate {yt} from {wt} using yt = P22*1 P21* wt

5. Calculate desired stylised facts, ignoring first

few observations

58
Variances

Standard
deviation
Interest rate 0.46

Output gap 1.39

Inflation 0.46

Correlations

Interest Output Inflation

rate gap
Interest rate 1 -1 -1
Output gap -1 1 1
Inflation -1 1 1

59
Autocorrelations

t,t-1 t,t-2 t,t-3 t,t-4

Interest rate 0.50 0.25 0.12 0.06

Output gap 0.50 0.25 0.12 0.06

Inflation 0.50 0.25 0.12 0.06

Cross-correlations

Correlation with output gap at time t

t-2 t-1 t t+1 t+2
Output gap 0.25 0.50 1 0.50 0.25
Inflation 0.25 0.50 1 0.50 0.25
Interest rate -0.25 -0.50 -1 -0.50 -0.25

60
Impulse response functions
What is effect of 1 standard deviation shock in
any element of vt on variables wt and yt?

1. Start from steady-state value w0 = 0

0 0 0 0 L
M M M M
2. Define shock of interest
{vt } = 1 0 0 0 L
M M M M

0 0 0 0 L

Impulse response functions

3. Simulate {wt} from {vt} recursively using
wt +1 = ( P11* P12* P22*1 P21* ) 1 1 ( P11* P12* P22*1 P21* ) wt
+ ( P11* P12* P22*1 P21* ) 1 R1vt +1

4. Calculate impulse response {yt} from {wt}

using yt = P22*1 P21* wt

61
Response to vt shock
0.6
0.4
0.2
0
-0.2 0 1 2 3 4 5 6 7 8 9 1011 12
-0.4 t
-0.6
-0.8
-1
i nt er e st r a t e
-1.2 o ut p ut g ap
i nf l a t i o n
-1.4

Forecast error variance

decomposition (FEVD)
Imagine you make a forecast for the output gap
for next h periods

Because of shocks, you will make forecast errors

What proportion of errors are due to each shock

at different horizons?

FEVD is a simple transform of impulse response

functions

62
FEVD calculation
Define impulse response function of output gap
to each shocks v1 and v2
1.5
1
0.5
0 t

-0.5 0 1 2 3 4 5 6 7 8

-1
-1.5

response to v1
11 21 31 41 51 61 71 81
12 22 32 42 52 62 72 82
response to v2
response at horizons 1 to 8

FEVD at horizon h = 1
At horizon h = 1, two sources of forecast errors

Shock
vt1 vt2
Impulse
response at 11 12
horizon 1
Contribution to
variance at (11 )2 21 ( )
1
2 2 2
v2
horizon 1

63
FEVD at horizon h = 1

Contribution of v1

(11 ) 2 v21
(11 ) 2 v21 + (12 ) 2 v22

FEVD at horizon h = 2
At horizon h = 2, four sources of forecast errors
Shock
vt1 vt2 vt1+1 vt2+1
Impulse
response at 21 22 11 12
horizon 2
Contribution to
variance at ( )
(21 )2 21 22 2 v22 ( ) ( )
1 2
1
2
1 1
2 2 2
v2
horizon 2

64
FEVD at horizon h = 2

Contribution of v1

(11 ) 2 v21 + (21 ) 2 v21

(11 ) 2 v21 + (21 ) 2 v21 + (12 ) 2 v22 + (22 ) 2 v22

FEVD at horizon h
At horizon h, 2h sources of forecast errors

Contribution of v1

( )
i =1
1 2
i
2
v1

h h

( )
i =1
1 2
i
2
v1
+ (i2 ) 2 v22
i =1

65
FEVD for output gap
1
0.9
0.8
0.7
0.6 i nt e r est r at e
sho c k
0.5 co st - p us h
sho c k
0.4
0.3
0.2
0.1
0 h
0 1 2 3 4 5 6 7 8 9 10

FEVD for inflation

1
0.9
0.8
0.7
0.6 i nt e r est r at e
sho c k
0.5
co st - p us h
0.4 sho c k

0.3
0.2
0.1
0 h
0 1 2 3 4 5 6 7 8 9 10

66
FEVD for interest rates
1
0.9
0.8
0.7
0.6 i nt e r est r at e
sho c k
0.5
co st - p us h
0.4 sho c k

0.3
0.2
0.1
0 h
0 1 2 3 4 5 6 7 8 9 10

Next steps

Models with multiple shocks

Taylor rules

Optimal Taylor rules

67
 Advanced dynamic
models
Martin Ellison
University of Warwick and CEPR

Bank of England, December 2005

More complex models

Frisch-Slutsky paradigm

Impulses

Propagation

Fluctuations

68
Impulses
Can add extra shocks to the model

xt = Et xt +1 1 (it Et t +1 ) + g t
t = Et t +1 + xt + ut
it = t + vt

Shocks may be correlated

vt +1 11 12 13 vt 11 12 13 tv+1

ut +1 = 21 22 23 ut + 21 22 23 tu+1
g
t +1 31 32 33 g t 31 32 33 tg+1

Propagation
Add lags to match dynamics of data
(Del Negro-Schorfeide, Smets-Wouters)

h 1 h
xt = xt 1 + Et xt +1 1 (it Et t +1 ) 0.35
1+ h 1+ h 1+ h
p p
t = t 1 + Et t +1 + xt 0.29
1 + p 1 + p 1 + p

Taylor rule it = t + x xt + vt

69
Solution of complex models
Blanchard-Kahn technique relies on invertibility
of A0 in state-space form.

A0 E t X t +1 = A1 X t + B 0 v t +1
E t X t +1 = A01 A1 X t + A01 B 0 v t +1

A B

E t X t +1 = AX t + Bv t +1

QZ decomposition
For models where A0 is not invertible

A0 Et X t +1 = A1 X t + B0 vt +1

QZ decomposition: Q, , Z , s.t. Q' Z ' = A0

Q ' Z ' = A1
upper
triangular

70
Recursive equations
11 12 w ~ 12 w~
t +1
~ = 11
~t + (Q' ) 1 B0 vt +1
0 22 Et yt +1 0 22 yt

~ + E ~
11w ~ ~
t +1 12 t yt +1 = 11 wt + 12 yt + R1vt +1 stable

22Et ~
yt +1 = 22 ~
yt + R2vt +1 unstable

Recursive structure means unstable equation

can be solved first

Solution strategy

Solve unstable ~
yt
transformed equation

Substitute into stable ~

transformed equation wt

Translate back into wt

original problem y
t

71
Simulation possibilities

Stylised facts

Impulse response functions

Forecast error variance decomposition

Optimised Taylor rule

What are best values for parameters in Taylor

rule it = t + x xt + vt ?

Introduce an (ad hoc) objective function for policy

min i ( t2 + x xt2 + i it 2 )
i =0

72
Brute force approach
Try all possible combinations of Taylor rule
parameters

Check whether Blanchard-Kahn conditions are

satisfied for each combination

For each combination satisfying B-K condition,

simulate and calculate variances

Brute force method

Calculate simulated loss for each combination

Best (optimal) coefficients are those satisfying

B-K conditions and leading to smallest
simulated loss

73
Grid search
2 For each point
check B-K
conditions
x 1

Find lowest loss

amongst points
0 1 2
satisfying B-K
condition

Next steps

Ex 14: Analysis of model with 3 shocks

Ex 15: Analysis of model with lags

Ex 16: Optimisation of Taylor rule coefficients

74