Chapter 1

Expectations and Economic

Dynamics
Expectations lie at the core of economic dynamics as they usually determine,
not only the behavior of the agents, but also the main properties of the economy under study. Although having been soon recognized, the question of
expectations has been neglected for a while, as this is a pretty difficult issue to deal with. In this course, we will mainly be interested by rational
expectations

1.1

The rational expectations hypothesis

The term rational expectations is most closely associated with Nobel Laureate Robert Lucas of the University of Chicago, but the question of rationality
of expectations came into the place before Lucas investigated the issue (see
Muth [1960] or Muth [1961]). The most basic interpretation of rational expectations is usually summarized by the following statement:
Individuals do not make systematic errors in forming their expectations; expectations errors are corrected immediately, so that
on average expectations are correct.
But rational expectation is a bit more subtil concept that may be defined in
3 ways.
Definition 1 (Broad definition) Rational expectations are such that individuals formulate their expectations in an optimal way, which is actually com1

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

parable to economic optimization.

This definition actually states that individuals collect information about the
economic environment and use it in an optimal way to specify their expectations. For example, assume that an individual wants to make forecasts on
an asset price, she needs to know the series of future dividends and therefore needs to make predictions about these dividends. She will then collect all
available information about the environment of the firm (expected demand, investments, state of the market. . . ) and use this information in an optimal way
to make expectations. But two key issues emerge then: (i) the cost of collecting information and (ii) the definition of the objective function. Hence, despite
its general formulation, this definition remains weakly operative. Therefore, a
second definition was proposed in the literature.
Definition 2 (middefinition) Agents do not waste any available piece of
information and use it to make the best possible fit of the real world.
This definition has the great advantage of avoiding to deal with the problem
of the cost of collecting information we only need to know that agents do
not waste information but it remains weakly operative in the sense it is
not mathematically specified. Hence, the following weak definition is most
commonly used.
Definition 3 (weak definition) Agents formulate expectations in such a way
that their subjective probability distribution of economic variables (conditional
on the available information) coincides with the objective probability distribution of the same variable (the state of Nature) in an equilibrium:
xet = E(xt |)
where denote the information set
When the model satisfies a markovian property, essentially consists of past
realizations of the stochastic variables from t=0 on. For instance, if we go back
to our individual who wants to predict the price of an asset in period t, will
essentially consist of all past realizations of this asset price: = {pti ; i =
1 . . . t}. Beyond, this definition assumes that agents know the model and the

1.1. THE RATIONAL EXPECTATIONS HYPOTHESIS

probability distributions of the shocks that hit the economy that is what is
needed to compute all the moments (average, standard deviations, covariances
. . . ) which are needed to compute expectations. In other words, and this is
precisely what makes rational expectations so attractive:
Expectations should be consistent with the model
= Solving the model is finding an expectation function.
Notation: Hereafter, we will essentially deal with markovian models, and will
work with the following notation:
Eti (xt ) = E(xt |ti )
where ti = {xk ; k = 0 . . . t i}.
The weak definition of rational expectations satisfies two vary important properties.
Proposition 1 Rational Expectations do not exhibit any bias: Let x
bt = xt xet
denote the expectation error:

Et1 (b
xt ) = 0
which essentially corresponds to the fact that individuals do not make systematic errors in forming their expectations.
Proposition 2 Expectation errors do not exhibit any serial correlation:
Covt1 (b
xt , x
bt1 ) = Et1 (b
xt x
bt1 ) Et1 (b
xt )Et1 (b
xt1 )
= Et1 (b
xt )b
xt1 Et1 (b
xt )b
xt1

= 0
Example 1 Lets consider the following AR(2) process
xt = 1 xt1 + 2 xt2 + t
such that the roots lies outside the unit circle and t is the innovation of the
process.

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

1. Lets now specify = {xk ; k = 0, . . . , t 1}, then
E(xt |) = E(1 xt1 + 2 xt2 + t |)
= E(1 xt1 |) + E(2 xt2 |) + E(t |)
Note that by construction, we have xt1 and xt2 , therefore,
E(xt1 |) = xt1 and E(xt2 |) = xt2 . Since, t is an innovation,
it is orthogonal to any past realization of the process, t such that
E(t |) = 0. Hence
E(xt |) = 1 xt1 + 2 xt2
2. Lets now specify = {xk ; k = 0, . . . , t 2}, then
E(xt |) = E(1 xt1 + 2 xt2 + t |)
= E(1 xt1 |) + E(2 xt2 |) + E(t |)
Note that by construction, we have xt2 , such that as before E(xt2 |) =
xt2 . Further, we still have t such that E(t |) = 0. But now
xt1
/ such that
E(xt |) = 1 E(xt1 |) + 2 xt2
and we shall compute E(xt1 |):
E(xt1 |) = E(1 xt2 + 2 xt3 + t1 |)
= E(1 xt2 |) + E(2 xt3 |) + E(t1 |)
Note that xt2 , xt3 and t1 , such that
E(xt1 |) = 1 xt2 + 2 xt3
Hence
E(xt |) = (21 + 2 )xt2 + 2 xt3

This example illustrates the so called law of iterated projection.

Proposition 3 (Law of Iterated Projection) Lets consider two information sets t and t1 , such that t t1 , then
E(xt |t1 ) = E(E(xt |t )|t1 )

1.1. THE RATIONAL EXPECTATIONS HYPOTHESIS

Beyond, the example reveals a very important property of rational expectations: a rational expectation model is not a model in which the individual knows everything. Everything depends on the information structure. Lets consider some simple examples.
Example 2 (signal extraction) In this example, we will deal with a situation where the agents know the model but do not perfectly observe the shocks
they face. Information is therefore incomplete because the agents do not know
perfectly the distribution of the true shocks.
Assume that a firm wants to predict the demand, d, it will be addressed, but
only observes a random variable x that is related to d as
x=d+

(1.1)

where E(d) = 0, E(d2 ) = d < , E( 2 ) = < , E(d) = , and

E() = 0. This assumption amounts to state that x differs from d by a measurement error, . Note that in this example, we assume that there is a noisy
information, but the firm still knows the overall structure of the model (namely
it knows 1.1). The problem of the firm is then to formulate an expectation for
d only observing x: = {1, x}. In this case, the problem of the entrepreneur
is to determine E(d|). Since the entrepreneur knows the linear structure of
the model, it can guess that
E(d|) = 0 + 1 x
From proposition 1, we know that the expectation error exhibits no bias so that
E(d E(d|)|) = 0
which amounts to
E(d 0 1 x|) = 0
or

E(d 0 1 x|1) = 0
E(d 0 1 x|x) = 0

These are the two normal equation associated with an OLS estimate, hence we
have
1 =

2
Cov(x, d)
Cov(d + , d)
=
= 2 d 2
V(x)
V(d + )
d +

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

and
0 =

d2 + 2

Example 3 (bounded memory) In this example, we deal with a situation

where the agents know the model but have a bounded memory in the sense they
forget past realization of the shocks.
Lets consider the problem of a firm which demand depends on expected aggregate demand and the price level. In order to keep things as simple as possible, we will assume that the price is an exogenous i.i.d process with mean p
and variance p2 ) and that aggregate demand is driven by the following simple
AR(1) process
Yt = Yt1 + (1 )Y + t
where || < 1 and t is the innovation of the process. The demand then takes
the following form
dt = E(Yt+1 |) pt
But rather than being defined as = {Yti , pti , ti ; i = 0 . . . }, now takes
the form = {Yti , pti , ti ; i = 0 . . . k, k < }. Computing the rational
expectation is now a bit more tricky. We first have to write down the Wold
decomposition of the process of Y
Yt = Y +

i ti

i=0

Then E(Yt+1 |) can be computed as

E(Yt+1 |) = E

Y +

X
i=0

 !

i t+1i 

Since Y is a deterministic constant, E Y = Y , such that
E(Yt+1 |) = Y +

i E(t+1i |)

i=0

Since = {Yti , pti , ti ; i = 0 . . . k, k < }, we have ti i > k, such

that, in this case E(ti |) = 0. Hence,
E(Yt+1 |) = Y +

k
X
i=0

i+1 ti

1.2. A PROTOTYPICAL MODEL OF RATIONAL EXPECTATIONS

hence
dt = Y +

k
X
i=0

i+1 ti

pt

which may be reexpressed in terms of observable variables as





dt = Y + Yt Y k+1 Yt(k+1) Y
pt

1.2

1.2.1

A prototypical model of rational expectations

Sketching up the model

In this section we try to characterize the behavior of an endogenous variable

y that obeys the following expectational difference equation
yt = aEt yt+1 + bxt

(1.2)

where Et yt+1 E(yt+1 |) where = {yti , xti , i = 0, . . . , }.

Equation (1.2) may be given different interpretations. We now provide you
with a number of models that suit this type of expectational difference equation.
Assetpricing model:

Let pt be the price of a stock, dt be the dividend,

and r be the rate of return on a riskless asset, assumed to be held constant

over time. Standard theory of finance teaches us that if agents are risk neutral,
then the arbitrage between holding stocks and the riskless asset should be such
that the expected return on the stock given by the expected rate of capital
gain plus the dividend/price ratio should equal the riskless interest rate:
Et pt+1 pt dt
+
=r
pt
pt
or equivalently
pt = aEt pt+1 + adt where a
The Cagan Model:

1
<1
1+r

The Cagan model is a macro model that was designed

to furnish an explanation to the hyperinflation problem. Cagan assumes that

the demand for real balances takes the following form


Mtd
e
= exp t+1
Pt

(1.3)

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

e
where t+1
denotes expected inflation
e
t+1

Et (Pt+1 ) Pt
Pt

In an equilibrium, money demand equals money supply, such that

Mtd = Mts = Mt
hence in an equilibrium, equation (1.3) reduces to


Mt
Et (Pt+1 ) Pt
= exp
Pt
Pt

(1.4)

Taking logs lowercases will denote logged variables using the approximation log(1 + x) x and reorganizing, we end up with
pt = aEt (pt+1 ) + (1 a)mt where a =
Monopolistic competition

1+

Consider a monopolist that faces the following

demand
pt = yt Et yt+1

(1.5)

the term in yt accounts for the fact that the greater the greater the price
is, the lower the demand is. The term in Et yt+1 accounts for the fact that
greater expected sells tend to lower the price.1 The firm acts as a monopolist
maximizing its profit
max pt yt ct yt
yt

taking the demand (1.5) into account. ct is the marginal cost, which is assumed to follow an exogenous stochastic process. Note that we assume, for the
moment, that the firm adopts a purely static behavior. Profit maximization
taking (1.5) into account yields
2yt Et yt+1 ct = 0
which may be rewritten as
yt = aEt (pt+1 ) + bct + d where a =
1

,b=
and d =
2
2
2

If < 0, the model may be given an alternative interpretation. Greater expected sells
lead the firm to raise its price (you may think of goods such as tobacco, alcohol, . . . , each
good that may create addiction).

1.2. A PROTOTYPICAL MODEL OF RATIONAL EXPECTATIONS

At this point we are left with the expectational difference equation (1.2),
which may either be solved forward or backward looking depending on
the value of a. When |a| < 1 the solution should be forward looking, as it
will become clear in a moment, conversely, when |a| > 1 the model should be
solved backward. The next section investigates this issue.

1.2.2

Forward looking solutions: |a| < 1

The problem that arises with the case |a| < 1 may be understood by looking
at figure 1.1, which reports the dynamics of equation
Et yt+1 =

b
1
yt xt
a
a

Holding xt constant and therefore eliminating the expectation. As can be

seen from the figure, the path is fundamentally unstable as soon as we look at
it in the usual backward looking way. Starting from an initial condition that
differs from y, say y0 , the dynamics of y diverges. The system then displays
a bubble.2 A more interesting situation arises when the variable yt represents
a variable such as a price or consumption in any case a variable that shifts
following a shock and that does not have an initial condition but a terminal
condition of the form
lim |yt | <

(1.6)

In fact such a terminal condition which is often related to the socalled

transversality condition arising in dynamic optimization models bounds
the sequence of {yt }
t=0 and therefore imposes stationarity.

Solving this

system then amounts to find a sequence of stochastic variable that satisfies

(1.2). This may be achieved in different ways and we now present 3 possible
methods.
Forward substitution
This method proceeds by iterating forward on the system, making use of the
law of iterated projection (proposition 3). Let us first recall the expectational
difference equation at hand:
yt = aEt yt+1 + bxt
2

We will come back to this point later on.

10

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Figure 1.1: The regular case
Et yt+1 6
-

45

yt

y y0

Iterating one step forward that is plugging the value of yt evaluated in t + 1

in the expectation, we get
yt = aEt (Et+1 (ayt+2 + bxt+1 )) + bxt
The law of iterated projection implies that Et (Et+1 (yt+2 )) = Et yt+2 , so that
yt = a2 Et (yt+2 ) + abEt (xt+1 ) + bxt
Iterating one step forward, we get
yt = a2 Et (Et+2 (ayt+3 + bxt+2 )) + abEt (xt+1 ) + bxt
Once again making use of the law of iterated projection, we get
yt = a3 Et (yt+3 ) + a2 bEt (xt+2 ) + abEt (xt+1 ) + bxt
Continuing the process, we get
yt = b lim

k
X
i=0

ai Et (xt+i ) + lim ak+1 Et (yt+k+1 )

k

1.2. A PROTOTYPICAL MODEL OF RATIONAL EXPECTATIONS

11

For the first term to converge, we need the expectation Et (xt+k ) not to increase
at a too fast pace. Then provided that |a| < 1, a sufficient condition for the
first term to converge is that the expectation explodes at a rate lower than
|1/a 1|.3 In the sequel we will assume that this condition holds.
Finally, since |a| < 1, imposing that lim |yt | < holds, we have
t

k+1

lim a

Et (yt+k+1 ) = 0

and the solution is given by

yt = b

ai Et (xt+i )

(1.7)

i=0

In other words, yt is given by the discounted sum of all future expected values
of xt . In order to get further insight on the form of the solution, we may be
willing to specify a particular process for xt . We shall assume that it takes
the following AR(1) form:
xt = xt1 + (1 )x + t
where || < 1 for sake of stationarity and t is the innovation of the process.
Note that
Et xt+1 = xt + (1 )x
Et xt+2 = Et xt+1 + (1 )x = 2 xt + (1 )(1 + )x
Et xt+3 = Et xt+2 + (1 )x = 3 xt + (1 )(1 + + 2 )x
..
.
Et xt+i = i xt + (1 )(1 + + 2 + . . . + i )x = i xt + (1 i+1 )x
Therefore, the solution takes the form

X
ai (i xt + (1 i )x)
yt = b
= b

i=0

X
i=0

(a) (xt x) +

X
i=0

ax


xt x
x
+
= b
1 a 1 a
b
ab(1 )
=
x
xt +
1 a
(1 a)(1 a)


This will actually be the case with a stationary process.

12

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Figure 1.2 provides an example of the process generated by such a solution, in

the deterministic case and in the stochastic case. In the deterministic case, the
economy always lies on its longrun value y , which is the only stable point.
We then talk about steady state that is a situation where yt = yt+k = y .
In the stochastic case, the economy fluctuates around the mean of the process,
and it is noteworthy that any change in xt instantaneously translates into a
change in yt . Therefore, the persistence of yt is given by that of xt .
Figure 1.2: Forward Solution
Deterministic Case

Stochastic Case

9
8

5.5
7
5

6
5

4.5
4
4
0

50

100
Time

150

200

3
0

50

100
Time

150

200

Note: This example was generated using a = 0.8, b = 1, = 0.95, = 0.1 and x = 1.
Matlab Code: Forward Solution
\simple
%
% Forward solution
%
lg = 100;
T
= [1:long];
a
= 0.8;
b
= 1;
rho = 0.95;
sx = 0.1;
xb = 1;
%
% Deterministic case
%
y=a*b*xb/(1-a);
%
% Stochastic case
%
%
% 1) Simulate the exogenous process
%
x
= zeros(lg,1);
randn(state,1234567890);

1.2. A PROTOTYPICAL MODEL OF RATIONAL EXPECTATIONS

13

e
= randn(lg,1)*sx;
x(1) = xb;
for i=2:long;
x(i) = rho*x(i-1)+(1-rho)*xb+e(i);
end
%
% 2) Compute the solution
%
y
= b*x/(1-a*rho)+a*b*(1-rho)*xb/((1-a)*(1-a*rho));

Factorization
The method of factorization was introduced by Sargent [1979]. It amounts to
make use of the forward operator F , introduced in the first chapter.4 In a first
step, equation (1.2) is rewritten in terms of F
yt = aEt yt+1 +bxt Et (yt ) = aEt (yt+1 )+bEt (xt ) (1aF )Et yt = bEt xt
which rewrites as
E t yt = b

Et xt
1 aF

since |a| < 1, we have

X
1
ai F i
=
1 aF
i=0

Therefore, we have

E t y t = yt = b

ai F i Et xt = b

ai Et xt+i

i=0

i=0

Note that although we get, obviously, the same solution, this method is not
as transparent as the previous one since the terminal condition (1.6) does not
appear explicitly.
Method of undetermined coefficients
This method proceeds by making an initial guess on the form of the solution.
An educated guess for the problem at hand would be
yt =

i Et xt+i

i=0

Recall that the forward operator is such that F i Et (xt ) = Et (xt+i ).

14

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Plugging the guess in (1.2) leads to

i Et xt+i = aEt

i=0

i Et+1 xt+1+i

i=0

+ bxt

Solving the model then amounts to find the sequence of i , i = 0, . . . , such

that the guess satisfies the equation. We then proceed by identification.
i = 0 0 = b
i = 1 1 = a0
i = 2 2 = a1
..
.
such that i = ai1 , with 0 = b. Note that since |a| < 1, this sequence
converges toward 0 as i tends toward infinity. Therefore, the solution writes
yt = b

ai Et xt+i

i=0

The problem with such an approach is the we need to make the right guess
from the very beginning. Assume for a while that we had specified the following guess
yt = xt
Then
xt = aEt xt+1 + bxt
Identifying term by terms we would have obtained = b or = 0, which is
obviously a mistake.
As a simple example, let us assume that the process for xt is given by the same
AR(1) process as before. We therefore have to solve the following dynamic
system

yt = aEt yt+1 + bxt

xt = xt1 + (1 )x + t

Since the system is linear and that xt exhibits a constant term, we guess a
solution of the form
yt = 0 + 1 xt
Plugging this guess in the expectational difference equation, we get
0 + 1 xt = aEt (0 + 1 xt+1 ) + bxt

1.2. A PROTOTYPICAL MODEL OF RATIONAL EXPECTATIONS

15

which rewrites, computing the expectation5

0 + 1 xt = a0 + a1 xt + a1 (1 )x + bxt
Identifying term by term, we end up with the following system of equations
0 = a0 + a1 (1 )x1 = a1 + b
The second equation yields
1 =

b
1 a

the first one gives

0 =

ab(1 )
x
(1 a)(1 a)

One advantage of this method is that it is particularly simple, and it requires

the user to know enough on the economic problem to formulate the right guess.
This latter property precisely constitutes the major drawback of the method
as if formulating a guess is simple for linear economies it may be particularly
tricky even impossible in all other cases.

1.2.3

Backward looking solutions: |a| > 1

Until now, we have only considered the case of a regular economy in which
|a| < 1, which provided we are ready to impose a nonexplosion condition
yields a unique solution that only involves fundamental shocks. In this
section we investigate what happens when we relax the condition |a| < 1
and consider the case |a| > 1. This fundamentally changes the nature of the
solution, as can be seen from figure 1.3. More precisely, any initial condition
y0 for y is admissible as any leads the economy back to its longrun solution
y. The equilibrium is then said to be indeterminate.
From a mathematical point of view, the sum involved in the forward solution
is unlikely to converge. Therefore, the solution should be computed in an
alternative way. Let us recall the expectational difference equation
yt = aEt yt+1 + bxt
5
Note that this is here that we make use of the assumptions on the process for the
exogenous shock.

16

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Figure 1.3: The irregular case
Et yt+1 6

45
y

y0

yt

Note that, by construction, we have

yt+1 = Et (yt+1 ) + t+1
where t+1 is the expectational error, uncorrelated by construction with
the information set, such that Et t+1 = 0. The expectational difference equation then rewrites
yt = a(yt+1 t+1 ) + bxt
which may be restated as
yt+1 =

b
1
yt + xt + t+1
a
a

Since |a| > 1 this equation is stable and the system is fundamentally backward
looking. Note that t+1 is serially uncorrelated, and not necessarily correlated
with the innovations of xt . In other words, this shock may not be a fundamental shock and is alike a sunspot. For example, I wake up in the morning,
look at the weather and decides to consume more. Why? I dont know! This
is purely extrinsic to the economy!

1.2. A PROTOTYPICAL MODEL OF RATIONAL EXPECTATIONS

17

Figure 1.4 reports an example of such an economy. We have drawn the solution
to the model for different values of the volatility of the sunspot, using the
same draw. As can be seen, although each solution is perfectly admissible,
the properties of the economy are rather different depending on the volatility
of the sunspot variable. Besides, one may compute the volatility and the first
Figure 1.4: Backward Solution
=0.1

Without sunspot
2.5

2.5

1.5

1.5

0.5

0.5

0
0

50

100
Time
=0.5

150

200

0
0

50

50

100
Time

150

200

100
Time

150

200

0
0

100
Time
=1

150

200

2
0

50

Note: This example was generated using a = 1.8, b = 1, = 0.95, = 0.1 and x = 1.

order autocorrelation of yt :6
y2 =
y (1) =

a2
b2 ( + a)
2

+
2
x
(a2 1)(a )
a2 1
"
#
b2 (a2 1)x2
1
1+ 2
a
b (a + )x2 + a2 (a )2

Therefore, as should be expected, the overall volatility of y is an increasing

function of the volatility of the sunspot, but more important is the fact that
its persistence is lower the greater the volatility of the sunspot. Hence, there
6

We leave it to you as an exercize.

18

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

may be many candidates to the solution of such a backward looking equation,

each displaying totally different properties.
Matlab Code: Backward Solution
%
% Backward solution
%
lg = 200;
T
= [1:lg];
a
= 1.8;
b
= 1;
rho = 0.95;
sx = 0.1;
xb = 1;
se = 0.1;
%
% 1) Simulate the exogenous process
%
x
= zeros(lg,1);
randn(state,1234567890);
e
= randn(lg,1)*sx;
x(1) = xb;
for i=2:lg;
x(i) = rho*x(i-1)+(1-rho)*xb+e(i);
end
%
% 2) Compute the solution
%
randn(state,1234567891);
es = randn(lg,1);
y1 = zeros(lg,1); % without sunspot
y2 = zeros(lg,1); % with sunspot (se=0.1)
y3 = zeros(lg,1); % with sunspot (se=0.5)
y4 = zeros(lg,1); % with sunspot (se=1)
y1(1) = 0;
y2(1) = es(1)*0.1;
y3(1) = es(1)*0.5;
y4(1) = es(1);
for i=2:lg;
y1(i) = y1(i-1)/a+b*x(i-1)/a;
y2(i) = y2(i-1)/a+b*x(i-1)/a+0.1*es(i);
y3(i) = y3(i-1)/a+b*x(i-1)/a+0.5*es(i);
y4(i) = y4(i-1)/a+b*x(i-1)/a+es(i);
end

1.2.4

One step backward: bubbles

Lets now go back to the forward looking solution. The ways we dealt with it
led us to eliminate any bubble that is we imposed condition (1.6) to bound
the sequence. By doing so, we restricted ourselves to a particular class of

1.2. A PROTOTYPICAL MODEL OF RATIONAL EXPECTATIONS

19

solution, but there may exist a wider class of admissible solution that satisfy
(1.2) without being bounded.
Let us now assume that such an alternative solution of the form does exist
yet = yt + bt

where yt is the solution (1.7) and bt is a bubble. In order for yet to be a solution
to (1.2), we need to place some additional assumption on its behavior.

If yet = yt + bt it has to be the case that Et yet+1 = Et yt+1 + Et bt+1 , such that

plugging this in (1.2), we get

yt + bt = aEt yt+1 + aEt bt+1 + bxt

Since yt is a solution to (1.2), we have that yt = aEt yt+1 + bxt such that the
latter equation reduces to
bt = aEt bt+1 Et bt+1 = a1 bt
Therefore, any bt that satisfies the latter restriction will be such that yet is a

solution to (1.2). Note that since |a| < 1 in the case of a forward solution,

bt explodes in expected values therefore referring directly to the common

sense of a speculative bubble. Up to this point we have not specified any
particular functional form for the bubble. Blanchard and Fisher [1989] provide
two examples of such bubbles:
1. The everexpanding bubble: bt then simply follows a deterministic trend
of the form:
bt = b0 at
It is then straightforward to verify that bt = aEt bt+1 . How should we
interpret such a behavior for the bubble? In order to provide with some
insights, lets consider the case of the assetpricing equation:
Et pt+1 pt dt
+
=r
pt
pt
where dt = d (for simplicity). It is straightforward to check that the
nobubble solution (the fundamental solution) takes the form:
pt = p =

d
r

20

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

which sticks to the standard solution that states that the price of an asset
should be the discounted sum of expected dividends (you may check that
P
i
d /r =
i=0 (1+r) d ). If we now add a bubble of the kind we consider

that is bt = b0 at = b0 (1 + r)t provided b0 > 0 the price of the

asset will increase exponentially though the dividends are constant. The

explanation for such a result is simple: individuals are ready to pay a

price for the asset greater than expected dividends because they expect
the price to be higher in future periods, which implies that expected
capital gains will be able to compensate for the low price to dividend
ratio. This kind of anticipation is said to be selffulfilling. Figure 1.5
reports an example of such a bubble.
Figure 1.5: Deterministic Bubble
Asset price

Bubble

28

2.5
Bubble solution
Fundamental solution

27
2
26
1.5
25
24

10
Time

15

20

1
0

dividend/price

10
Time

15

20

Capital Gain (%)

0.039

0.35

0.0385

0.3

0.038
0.25
0.0375
0.2

0.037
0.0365
0

10
Time

15

20

0.15
0

10
Time

Note: This example was generated using d = 1, r = 0.04.

Matlab Code: Deterministic Bubble

%
% Example of a deterministic bubble
% The case of asset pricing (constant dividends)
%
d_star = 1;

15

20

1.2. A PROTOTYPICAL MODEL OF RATIONAL EXPECTATIONS

21

r
= 0.04;
%
% Fundamental solution p*
%
p_star = d_star/r;
%
% bubble
%
long
= 20;
T
= [0:long];
b
= (1+r).^T;
p
= p_star+b;

2. The burstingbubble: A problem with the previous example is that the

bubble is everexpanding whereas observation and common sense suggests that sometimes the bubble bursts. We may therefore define the
following bubble:
bt+1 =

(a)1 bt + t+1
t+1

with probability
with probability 1

with Et t+1 = 0. So defined, the bubble keeps on inflating with probability and bursts with probability (1 ). Lets check that bt = aEt bt+1
bt =
=
=
=

aEt (((a)1 bt + t+1 ) + (1 )t+1 )

aEt ((a)1 bt ) + t+1 )
aEt (a1 bt )
bt

taking bursting into account

grouping terms in t+1
since Et t+1 = 0
since bt is known in t

Figure 1.6 reports an example of such a bubble (the vertical lines in

the upper right panel of the figure corresponds to time when the bubble
bursts). The intuition for the result is the same as before: individuals are
ready to pay a higher price for the asset than the expected discounted
dividends because they expect with a sufficiently high probability that
the price will be high enough in subsequent periods to generate sufficient
capital gains to compensate for the lower price to dividend ratio. The
main difference with the previous case is that this bubble is now driven
by a stochastic variable, labelled as sunspot in the literature.

22

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Figure 1.6: Bursting Bubble

Asset price

Bubble

200

150
Bubble solution
Fundamental solution

150

100

100

50

50

0
0

50

100
Time

150

200

50
0

50

dividend/price

100
Time

150

200

Capital Gain (%)

0.06

50

0.05
0.04

0.03
0.02

50

0.01
0
0

50

100
Time

150

200

100
0

50

100
Time

150

200

Note: This example was generated using d = 1, = 0.95, r = 0.04.

1.3. A STEP TOWARD MULTIVARIATE MODELS

23

Matlab Code: Bursting Bubble

%
% Example of a bursting bubble
% The case of asset pricing (constant dividends)
%
d_star = 1;
r
= 0.04;
%
% Fundamental solution p*
%
p_star = d_star/r;
%
% bubble
%
long
= 200;
prob
= 0.95;
randn(state,1234567890);
e
= randn(long,1);
rand(state,1234567890);
ind
= rand(long,1);
b
= zeros(long,1);
dum
= zeros(long,1);
b(1)
= 0;
for i = 1:long-1;
dum(i)= ind(i)<prob;
b(i+1)= dum(i)*(b(i)*(1+r)/prob+e(i+1))+(1-dum(i))*e(i+1);
end;
p
= p_star+b;

Up to this point we have been dealing with very simple situations where the
problem is either backward looking or forward looking. Unfortunately, such
a case is rather scarce, and most of economic problems such as investment
decisions, pricing decisions . . . are both backward and forward looking. We
examine such situations in the next section.

1.3

A step toward multivariate Models

We are now interested in solving a slightly more complicated problem involving

one lag (for the moment!) of the endogenous variable:
yt = aEt yt+1 + byt1 + cxt

(1.8)

This equation may be encountered in many different models, either in macro,

micro, IO. . . as we will see later on. For the moment, let us assume that this
is obtained from whatever model we may think of and let us take it as given.

24

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

We are now willing to solve this expectational equation. As before, there exist
many methods.

1.3.1

The method of undetermined coefficients

Let us recall that solving the equation using undetermined coefficients amounts
to formulate a guess for the solution and find some restrictions on the coefficients of the guess such that equation (1.8) is satisfied. An educated guess in
this case is given by
yt = yt1 +

i Et xt+i

i=0

Where does this guess come from? Experience! and this is precisely why
the method of undetermined coefficients, although it may appear particularly
practical in a number of (simple) problems, is not always appealing.
Plugging this guess in equation (1.8) yields
#
"

X
X
i Et+1 xt+1+i + byt1 + cxt
i Et xt+i = aEt yt +
yt1 +
i=0

i=0

= a yt1 +

i Et xt+i

i=0

+ aEt

"

+byt1 + cxt

= (a2 + b)yt1 + a

i Et xt+i + a

i=0

i Et+1 xt+1+i

i Et xt+1+i + cxt

i=0

i=0

Everything is then a matter of identification (term by term):

= a2 + b

(1.9)

0 = a0 + c

(1.10)

i = ai + ai1

i > 1

Solving (1.9) for amounts to solve the second order polynomial

1
b
2 + = 0
a
a
which admits two solutions such that

1 + 2 =
1 2 = ab

1
a

Three configurations may emerge from the above equation

(1.11)

1.3. A STEP TOWARD MULTIVARIATE MODELS

25

1. the two solutions lie outside the unit circle: the model is said to be a
source and only one particular point the steady state is a solution
to the equation.
2. One solution lie outside the unit circle and the other one inside: the
model exhibits the saddle path property.
3. The two solutions lie inside the unit circle: the model is said to be a sink
and there is indeterminacy.
Here, we will restrict ourselves to the situation where an extended version of
the condition |a| < 1 we were dealing with in the preceding section holds,
namely one root will be of modulus greater than one and the other less than
one. The model will therefore exhibit the socalled saddle point property, for
which we will provide a geometrical interpretation in a moment. To sum up,
we consider a situation where |1 | < 1 and |2 | > 1. Since we restrict ourselves
to the stationary solution, we necessarily have || < 1 so that = 1 .
Once has been obtained, we can solve for i , i = 0, . . .. 0 is obtained from
(1.10) and takes the value
0 =

c
1 a1

We then get i , i > 1, from (1.11) as

i =

a
i1 =
1 a1

1
a

1
i1
1

Since 1 + 2 = 1/a, the latter equation rewrites

i = 1
2 i1
where |2 | > 1, such that this sequence converges toward zero. Therefore the
solution is given by
yt = 1 yt1 +

X
c
i
2 Et xt+i
1 a1
i=0

Example 4 In the case of an AR(1) process for xt , the solution is straightforward, as all the process may be simplified. Indeed, let us consider the following
problem

yt = aEt yt+1 + byt1 + cxt

xt = xt1 + t

26

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

with t ; N (0, ). An educated guess for the solution of this equation would
be
yt = yt1 + xt
Let us then compute the solution of the problem, that is let us find and
. Plugging the guess for the solution in the expectational difference equation
leads to
yt1 + xt = aEt (yt + xt+1 ) + byt1 + cxt
= a2 yt1 + axt + axt + byt1 + cxt
= (a2 + b)yt1 + (c + a( + ))xt
Therefore, we have to solve the system

= a2 + b
= c + a( + )
Like in the general case, we select the stable root of the first equation 1 , such
that |1 | < 1, and =

c
1a(1 +)

Figure (1.7) reports an example of such an

economy for two different parameterizations.

Matlab Code: BackwardForward Solution
%
% Solve for
%
% y(t)=a E y(t+1) + b y(t-1) + c x(t)
% x(t)= rho x(t-1)+e(t)
e iid(0,se)
%
% and simulate the economy!
%
a
= 0.25;
b
= 0.7;
c
= 1;
rho
= 0.95;
se
= 0.1;
mu
[m,i]
mu1
[m,i]
mu2

=
=
=
=
=

roots([a -1 b]);
min(mu);
mu(i);
max(mu);
mu(i);

alpha = b/(1-a*(mu1+rho));
%
% Simulation
%

1.3. A STEP TOWARD MULTIVARIATE MODELS

27

Figure 1.7: Backwardforward solution

x

6
4

0.5

2
0

2
0.5
0

50

100
Time

150

200

4
0

50

100
Time

150

200

150

200

Note: a = 0.25, b = 0.7, = 0.95, = 0.1

3
2

0.5

1
0

1
0.5
0

50

100
Time

150

200

2
0

50

Note: a = 0.7, b = 0.25, = 0.95, = 0.1

100
Time

28

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

lg
= 200;
randn(state,1234567890);
e
= randn(lg,1)*se;
x
= zeros(lg,1);
y
= zeros(lg,1);
x(1) = 0;
y(1) = alpha*x(1);
for i = 2:lg;
x(i) = rho*x(i-1)+e(i);
y(i) = mu1*y(i-1)+alpha*x(i);
end

Note that contrary to the simple case we considered in the previous section, the
solution does not only inherit the persistence of the shock, but also generates its
own persistence through 1 as can be seen from the first order autocorrelation
(1) =

1.3.2

1 +
1 + 1

Factorization

The method of factorization proceeds into 2 steps.

1. Factor the model (1.8) making use of the leading operator F :
(aF 2 F + b)Et yt1 = cEt xt
which may be rewritten as


1
b
c
F2 F +
Et yt1 = Et xt
a
a
a
which may also be rewritten as
c
(F 1 )(F 2 )Et yt1 = Et xt
a
Note that 1 and 2 are the same as the ones obtained using the method
of undetermined coefficients, therefore the same discussion about their
size applies. We restrict ourselves to the case |1 | < 1 (backward part)
and |2 | > 1 (forward part) that is to saddle path solutions.
2. Derive a solution for yt : Starting from the last equation, we can rewrite
it as

c
(F 1 )Et yt1 = (F 2 )1 Et xt
a

1.3. A STEP TOWARD MULTIVARIATE MODELS

or
(F 1 )Et yt1 =

29

c
1
(1 1
2 F ) Et xt
a2

Since |2 | > 1, we know that

1
=
(1 1
2 F)

i
i
2 F

i=0

so that
(F 1 )Et yt1 =

c X i i
c X i
2 F Et xt =
2 Et xt+i
a2
a2
i=0

i=0

Now, applying the leading operator on the left hand side of the equation
and acknowledging that 2 = 1/a 1 , we have

X
c
i
yt = 1 yt1 +
2 Et xt+i
1 a1
i=0

1.3.3

A matricial approach

In this section, we would like to provide you with some geometrical intuition of
what is actually going on when the saddle path property applies in the model.
To do so, we will rely on a matricial approach. First of all, let us recall the
problem we have in hands:
yt = aEt yt+1 + byt1 + cxt
Introducing the technical variable zt defined as
zt+1 = yt
the model may be rewritten as7

  1


 
Et yt+1
yt
c
ab
a
=
xt

zt+1
zt
1
1 0
Remember that Et yt+1 = yt+1 t+1 where t+1 is an iid process which represents the expectation error, therefore, the system rewrites

 

 

  1
yt
c
1
yt+1
ab
a

xt
t+1
=
zt
1
0
zt+1
1 0
7
In the next section we will actually pool all the equations in a single system, but for
pedagogical purposes let us separate exogenous variables from the rest for a while.

30

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

In order to understand the saddle path property let us focus on the homogenous part of the equation

  1




b
yt+1
yt
yt
a a
=
=W
zt+1
zt
zt
1 0
Provided b 6= 0 the matrix W can be diagonalized and may be rewritten as
W = P DP 1
where D contains the two eigenvalues of W and P the associated eigenvectors.
Figure 1.8 provides a way of thinking about eigenvectors and eigenvalues in
dynamical systems. The figure reports the two eigenvectors, P1 and P2 , associated with the two eigenvalues 1 and 2 of W . 1 is the stable root and
2 is the unstable root. As can be seen from the graph, displacements along
Figure 1.8: Geometrical interpretation of eigenvalues/eigenvectors
z 6

P2

P1
x2

6
x1 R

x2

x1

z
x3
x4


I
x4

x3 )

P1 are convergent, in the sense they shift either x1 or x4 toward the center of
the graph (x1 and x4 ), while displacements along P2 are divergent (shift of x2
and x3 to x2 and x3 ). In fact the eigenvector determines the direction along

1.3. A STEP TOWARD MULTIVARIATE MODELS

31

which the system will evolve and the eigenvalue the speed at which the shift
will take place.
The characteristic equation that gives the eigenvalues, in the case we are studying, is given by
1
b
det(W I) = 0 2 + = 0
a
a
which exactly corresponds to the equations we were dealing with in the previous sections. We will not enter the formal resolution of the model right now,
as we will undertake an extensive treatment in the next section. However, we
will just try to understand what may be going on using a phase diagram like
approach to understand the dynamics. Figures 1.91.11 report the different
possible configuration we may encounter solving this type of model. The first
one is a source (figure 1.9), which is such that no matter the initial condition
we feed the system with except y0 = y , z0 = z the system will explode.
Both y and z will not be bounded. The second one is a sink (figure 1.10), all
trajectories converge back to the steady state of the economy, one is then free
to choose whatever trajectory it wants to go back to the steady state. The
equilibrium is therefore indeterminate.
Figure 1.9: A source
yt

P2

yt+1 = 0

zt+1 = 0

6


I
i

P1
y

R



zt

32

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Figure 1.10: A sink: indeterminacy
yt

P1
y

P2

yt+1 = 0

zt+1 = 0

z z
-



y

i
i




zt

In the last situation (figure 1.11) this corresponds to the most commonly
encountered situation in economic theory the economy lies on a saddle: one
branch of the saddle converges to the steady state, the other one diverges. The
problem is then to select where to start from. It should be clear to you that in
t, zt is perfectly known as zt = yt1 which was selected in the earlier period.
zt is then said to be predetermined: the agents is endowed with its value when
she enters the period. This is part of the information set. Solving the system
therefore amounts to select a value for yt , given that for zt and the structure
of the model. How to proceed then? Let us assume for a while that at time
0, the economy is endowed with z0 , and assume that we impose the value y01
as a starting value for y. In such a case, the economy will explode: in other
words a solution including a bubble has been selected. If, alternatively, y02 is
selected, then the economy will converge to the steady state (z , y ) and all
the variables will be bounded. In other words, we have selected a trajectory
such that
lim |yt | <

holds. Otherwise stated, bubbles have been eliminated by imposing a terminal

condition. In the sequel, we will be mostly interested by situation were the

1.4. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (THE SIMPLE CASE)33

Figure 1.11: The saddle path
yt

P2

yt+1 = 0

P1

zt+1 = 0

y2
0

1
y0

y



?

z0

zt

economy either lies on a saddle path or is indeterminate. In the next section,

we will show you how to solve an expectational multivariate system of the
kind we were considering up to now.

1.4
1.4.1

Multivariate Rational Expectations Models (The

simple case)
Representation

Let us assume that the model writes

Mcc Yt = Mcs Mcs St
Mss0 Et St+1 + Mss1 St = Msc0 Et Yt+1 + Msc1 Yt + Mse Et+1

(1.12)
(1.13)

where Yt is a ny 1 vector of endogenous variables, t is a 1 vector of

exogenous serially uncorrelated random disturbances. A fairly natural interpretation of this dynamic system may be found in the statespace form literature: equation (1.17) corresponds to the standard measurement equation. It
relates variables of interest Yt to state variables St . (1.13) is the state equa-

34

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

tion that actually drives the dynamics of the economy under consideration:8 it
relates future values of states St+1 to current and expected values of variables
of interest, current state variables and shocks to fundamentals Et+1 . In other
words, (1.13) furnishes the transition from one state of the system to another
one. Our problem is then to solve this system.
As a first step, it would be great if we were able to eliminate all variables
defined by the measurement equation and restrict ourselves to a state equation,
as it would bring us back to our initial problem. To do so, we use (1.17) to
eliminate Yt .
1
Yt = Mcc
Mcs St

Plugging this expression in (1.13), we obtain:

Et St+1 = WS St + WE Et+1
where
1 M
WS = Mss0 Msc0 Mcc
cs

WE =

1 M
Mss0 Msc0 Mcc
cs

1

1

1 M
Mss1 Msc1 Mcc
cs

M se

We are then back to our expectational difference equation. But it needs additional work. Indeed, Farmer proposes a method that enables us to forget
about expectations when solving for the system. He proposes to replace the
expectation by the actual variable minus the expectation error
Et St+1 = St+1 Zt+1
where Et Zt+1 = 0. Then the system rewrites
St+1 = WS St + WE Et+1 + Zt+1

(1.14)

This is the system we will be dealing with.

8
Let us accept that statement for the moment, things will become clear as we will move
to examples.

1.4. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (THE SIMPLE CASE)35

1.4.2

Solving the system

? have shown that the existence and uniqueness of a solution depends fundamentally on the position of the eigenvalues of WS relative to the unit circle.
Denoting by NB and NF the number of, respectively, predetermined and jump
variables, and by NI and NO the number of eigenvalues that lie inside and
outside the unit circle, we have the following proposition.
Proposition 4
(i) If NI = NB and NO = NF , then there exists a unique solution path for
the rational expectation model that converges to the steady state;
(ii) If NI > NB (and NO < NF ), then the system displays indeterminacy;
(iii) If NI > NB (and NO > NF ), then the system is a source.
Hereafter we will deal with the two first situations, the last one being never
studied in economics.
The diagonalization of WS leads to
WS = P D P 1
where D is the matrix that contains the eigenvalues of WS on its diagonal and
P is the matrix that contains the associated eigenvectors. For convenience,
we assume that both D and P are such that eigenvalues are sorted in the
ascending order. We shall then consider two cases
1. The model satisfies the saddle path property (NI = NB and NO = NF )
2. The model exhibit indeterminacy (NI > NB and NO < NF )
The saddle path
In this section, we consider the case were the model satisfies the saddle path
property (NI = NB and NO = NF ). For convenience, we consider the following
partitioning of the matrices







DB 0
PBB PBF
PBB PBF
1
D=
, P =
, P =
0 DF
PF B PF F
PF B PF F

36

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

This partition conforms the position of the eigenvalues relative to the unit
circle. For instance, a B stands for the set of eigenvalues that lie within the
unit circle, whereas B stands for the set of eigenvalues that lie out of it.
We then apply the following modification to the system in order to make it
diagonal:
Set = P 1 St

so that

P 1 St+1 = P 1 WS P P 1 St + P 1 WE Et+1 + P 1 Zt+1

or
Set+1 = D Set + R Et+1 + P 1 Zt+1

The same partitioning is applied to R



RB.
R=
RF.
and the state vector
Set =

The system then rewrites as

! 

SeB,t+1
DB 0
=
0 DF
SeF,t+1

SeB,t
SeF,t

SeB,t
SeF,t
!

!


RB.
RF.

Et+1 +

PB.

PF.

Zt+1

Therefore, the law of motion of forward variables is given by

SeF,t+1 = DF SeF,t + RF. Et+1 + PF.

Zt+1

Taking expectations on both side of the equation

Et SeF,t+1 = DF SeF,t SeF,t = DF1 Et SeF,t+1

since DF is a diagonal matrix, forward iteration yields

SeF,t = lim DFj Et SeF,t+j
j

Provided Et SeF,t+j is bounded which amounts to eliminate bubbles we

have

lim DFj Et SeF,t+j = 0 SeF,t = 0

1.4. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (THE SIMPLE CASE)37

Then by construction, we have
SeF,t = PF B SB,t + PF F SF,t

which furnishes a restriction on SB,t and SF,t

PF B SB,t + PF F SF,t = 0
This condition expresses the relationship that relates the jump variables to the
predetermined variables, and therefore defined the initial condition SF,t which
is compatible with (i) the initial conditions on the predetermined variables
and (ii) the stationarity of the solution:
SF,t = (PF F )1 PF B SB,t = SB,t
Plugging this result in the law of motion of backward variables we have
SB,t+1 = (WBB + WBF )SB,t + RB Et+1 + ZB,t+1
but by definition, no expectation error may be done when predicting a predetermined variable, such that ZBt+1 = 0. Hence, the solution of the problem is
given by
SB,t+1 = MSS SB,t + MSE Et+1

(1.15)

where MSS = (WBB + WBF ) and MSE = RB .

As far as the measurement equation is concerned, thing are then rather simple.
.
Let us define = M 1 M = ( .. ), we have
cs

cc

Yt = B SB,t + F SF,t = SB,t

where = (B + F ).
The system is therefore solved and may be represented as
SB,t+1 = MSS SB,t + MSE Et+1

(1.16)

Yt = SB,t

(1.17)

SF,t = SB,t

(1.18)

38

1.5

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Multivariate Rational Expectations Models (II)

In this section we present a method to solve for multivariate rational expectations models, a because there are many of them (almost as many as authors
that deal with this problem).9 The one we present was introduced by Sims
[2000] and recently revisited by Lubik and Schorfheide [2003]. It has the advantage of being general and explicitly dealing with expectation errors. This
latter property makes it particularly suitable for solving sunspot equilibria.

1.5.1

Preliminary Linear Algebra

Generalized Schur Decomposition:

This is a method to obtain eigenval-

ues from a system which is not invertible. One way to think of this approach is
to remember that when we compute the eigenvalues of a diagonalizable matrix
A, we want to find a number and an associated eigenvector V such that
(A I)V = 0
The generalized Schur decomposition of two matrices A and B attempts to
compute something similar, but rather than considering (AI), the problem
considers (A B). A more formal, and above all a more rigorous
statement of the Schur decomposition is given by the following definitions and
theorem.
Definition 4 Let P C Cnn be a matrixvalued function of a complex
variable (a matrix pencil). Then the set of its generalized eigenvalues (P ) is
defined as
(P ) = {z C : |P (z) = 0}
When P (z) writes as Az B, we denote this set as (A, B). Then there exists
a vector V such that BV = AV .
Definition 5 Let P (z) be a matrix pencil, P is said to be regular if there
exists z C such that |P (z)| =
6 0 i.e. if (P ) 6= C.
9
In the appendix we present an alternative method that enables you to solve for singular
systems.

1.5. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (II)

39

Theorem 1 (The complex generalized Schur form) Let A and B belong

to Cnn and be such that P (z) = Az B is a regular matrix pencil. Then
there exist unitary n n matrices of complex numbers Q and Z such that
1. S = Q AZ is upper triangular,
2. T = Q BZ is upper triangular,
3. For each i, Sii and Tii are not both zero,
4. (A, B) = {Tii /Sii : Sii 6= 0}
5. The pairs (Tii , Sii ), i = 1 . . . n can be arranged in any order.
A formal proof of this theorem may be found in Golub and Van Loan [1996].
Singular Value Decomposition:

The singular value decomposition is used

for nonsquare matrices and is the most general form of diagonalization. Any
complex matrix A(n m) can be factored into the form
A = U DV
where U (n n), D(n m) and V (m m), with U and V unitary matrices
(U U = V V = I(nn) ). D is a diagonal matrix with positive values dii ,
i = 1 . . . r and 0 elsewhere. r is the rank of the matrix. dii are called the
singular values of A.

1.5.2

Representation

Let us assume that the model writes

A0 Yt = A1 Yt1 + Bt + Ct

(1.19)

where Yt is a n 1 vector of endogenous variables, t is a 1 vector of

exogenous serially uncorrelated random disturbances, and t is a k 1 vector
of expectation errors satisfying Et1 t = 0 for all t. A0 and A1 are both n n
coefficient matrices, while B is n and C is n k.

40

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

As an example of a model, let us consider the simple macro model

Et yt+1 + Et t+1 = yt + Rt
Et t+1 = t yt
Rt = t + gt
gt = gt1 + t
Let us then recall that by definition of an expectation error, we have
t = Et1 t + t
yt = Et1 yt + ty
Plugging the definition of Rt into the first two equations, and making use of
the definition of expectation errors, the system rewrites
yt = Et1 yt + ty
t = Et1 t + t
Et yt+1 + Et t+1 yt t gt = 0
Et t+1 t + yt = 0
gt = gt1 + t
Now defining

1
0
0
1

1
0
0

Yt = (yt , t , Et yt+1 , Et t+1 , gt ) ,

0 0 0
1 0 0 0
0 1 0 0
0 0 0

0 0 0 0
1 1
Y
=
t

0 0 0 0
0 0
0 0 1
0 0 0 0

the system may be writte10

0
0
1
0
0
0

0 t + 0
0
Y
+
t1

0
0
0

1
0

0
1
0
0
0

A nice feature of this representation is that it makes full use of expectation

errors and therefore may be given a fully interpretable economic meaning.

1.5.3

Solving the system

We now turn to the resolution of the system (1.19). Since, A0 is not necessarily
invertible, we will make full use of the generalized Schur decomposition of
(A0 , A1 ). There therefore exist matrices Q, Z, T and S such that
Q T Z = A0 , Q SZ = A1 , QQ = ZZ = Inn
10

Note that Yt1 = (yt1 , t1 , Et1 yt , Et1 t , gt1 )

 y 
t

1.5. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (II)

41

and T and S are upper triangular. Let us then define Xt = Z Yt and pre
multiply (1.19) by Q to get


 

 

T11 T12
W1,t
S11 S12
W1,t1
Q1
=
+
(Bt + Ct )
0 T22
W2,t
0 S22
W2,t1
Q2
(1.20)
Let us assume, without loss of generality that the system is ordered and partitioned such that the m 1 vector W2,t is purely explosive. Accordingly, the
remaining n m 1vector W1,t is stable. Let us first focus on the explosive
part of the system
T22 W2,t = S22 W2,t1 + Q2 (Bt + Ct )
For this particular block, the diagonal elements of T22 can be null, while S22
is necessarily full rank, as its diagonal elements must be different from zero if
the model is not degenerate. Therefore, the model may be written
1
1
W2,t = M W2,t+1 S22
Q2 (Bt+1 + Ct+1 ) where M S22
T22

Iterating forward, we get

s

W2,t = lim M W2,t+s

t

1
M s1 S22
Q2 (Bt+s + Ct+s )

s=1

In order to get rid of bubbles, we have to impose limt M s W2,t+s = 0, such

that
W2,t =

1
M s1 S22
Q2 (Bt+s + Ct+s )

s=1

Note that by definition of the vector Yt which does not involve any variable
which do not belong to the information set available in t, we should have
Et W2,t = W2,t . But,
Et W2,t = Et

1
M s1 S22
Q2 (Bt+s + Ct+s ) = 0

s=1

This therefore imposes a restriction on t and t . Indeed, if we go back to the

recursive formulation of W2,t and take into account that W2,t = 0 for all t, this
imposes
t
+
Q2 C
t
=
0
Q2 B
|{z}
|{z}
| {z }
|{z}
| {z }
(m ) ( 1)
(m k) (k 1)
(m 1)

(1.21)

42

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Our problem is now to know whether we can pin down the vector of expectation errors uniquely from that set of restrictions. Indeed, the vector t may
not be uniquely determined. This is the case for instance when the number
of expectation errors k exceeds the number of explosive components m. In
this case, equation (1.21) does not provide enough restrictions to determine
uniquely the vector t . In other words, it is possible to introduce expectation
errors which are not related with fundamental uncertainty the socalled
sunspot variables.
Sims [2000] shows that a necessary and sufficient condition for a stable solution
to exist is that the column space of Q2 B be contained in the column space of
Q2 C:
span(Q2 B) span(Q2 C)
Otherwise stated, we can reexpress Q2 B as a linear function of Q2 C (Q2 B =
Q2 C), implying that k > m. This is actually a generalization of the socalled
Blanchard and Khan condition that states that the number of explosive eigenvalues should be equal to the number of jump variables in the system. Lubik
and Schorfheide [2003] complement this statement by the following lemma.
Lemma 1 Statements (i) and (ii) are equivalent
(i) For every t R , there exists an t Rk such that Q2 Bt + Q2 Ct = 0.
(ii) There exists a (real) k matrix such that Q2 B = Q2 C
Endowed with this lemma, we can compute the set of all solutions (fully determinate and indeterminate solutions), reported in the following proposition.
Proposition 5 (Lubik and Schorfheide [2003]) Let t be a p 1 vector
of sunspot shocks, satisfying Et1 t = 0. Suppose that condition (i) of lemma
1 is satisfied. The full set of solutions for the forecast errors in the linear
rational expectations model is
1
t = (V1 D11
U1 Q2 B + V2 M1 )t + V2 M2 t

where M1 is a (k r) matrix and M2 is a (k r) p matrix.

1.5. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (II)

Proof: First of all, we have to find a solution to equation (1.21). The problem is that the rows of matrix Q2 C can be linearly dependent. Therefore,
we will use the Singular Value Decomposition of Q2 C
Q2 C =

U
|{z}
mm

which may be partitioned as


D11
Q2 C = (U1 U2 )
0

0
0

D
|{z}
mk

V
|{z}
kk



V1
V2

= U1 D11 V1

where D11 is a rr matrix, where r is the number of linearly independent

rows in Q2 C therefore the actual number of restrictions. Accordingly,
U1 is m r, and V1 is k r.
Given that we are looking for a solution that satisfies Q2 B = Q2 C,
equation (1.21) rewrites
U1 D11 (V1 t + V1 t ) = |{z}
0
| {z } |
{z
}
mr
r1
m1

We therefore now have r restrictions to identify the kdimensional vector

of expectation errors.
We guess that the solution implies that forecast errors are a linear function of (i) fundamental shocks and (ii) a p 1 vector of sunspot shocks
t , satisfying Et1 t = 0:
t = t + t
where is k and is k p.
Plugging this guess in the former equation, we get
U1 D11 (V1 + V1 )t + U1 D11 V1 t = 0
for all t and t . This triggers that we should have
U1 D11 (V1 + V1 ) = 0
U1 D11 V1 = 0

(1.22)
(1.23)

Let us first focus on equation (1.22). Since V is an orthonormal matrix,

it satisfies V V = I otherwise stated V1 V1 +V2 V2 = I and V V = I,
implying that V1 V2 = 0. A direct consequence of the first part of this
statement is that
e + V 2 M1
= V1 (V1 ) + V2 (V2 ) = V1

e V and M1 V . Since V V1 = I and V V2 = 0, (1.22)

with
1
2
1
1
therefore rewrites
e ) = 0
U1 D11 (V1 +

from which we get

e = V1

43

44

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

e and therefore . To do so,
We still need to identify to determine
we use the fact that Q2 B = Q2 C and Q2 C = U1 D11 V1 , to get
Q2 B = Q2 C = U1 D11 V1

Since U is orthonormal, we have U1 U1 = I, such that

1
V1 = U1 D11
Q2 B

e , we get
Therefore, plugging this result in the determination of
e = D1 U Q2 B

11 1

e + V2 M1 , we finally get
Since = V1

1
= V1 D11
U1 Q2 B + V2 M1

where M1 is left totally undetermined and therefore arbitrary.

We can now focus on (1.23) to determine . This is actually straightforward as it simply triggers that be orthogonal to V1 . But since
V1 V2 = 0, the orthogonal space of V1 is spanned by the columns of the
k (k r) matrix V2 . In other words, any linear combination of the
column of V2 would do the job. Hence
= V 2 M 2
where once again M2 is left totally undetermined and therefore arbitrary.
2

This last result tells us how to solve the model and under which condition
the system is determined or not. Indeed, let us recall that k is the number
of expectation errors, while r is the number of linearly independent expectation errors. According to this proposition, if k = r, all expectation errors
are linearly independent, and the system is therefore totally determinate. M1
and M2 are identically zeros. Conversely, if k > r expectation errors are not
linearly independent, meaning that the system does not provide enough restrictions to uniquely pin down the expectation errors. We therefore have to
introduce extrinsic uncertainty in the system the socalled sunspot variables. We will deal first with the determinate case, before considering the case
of indeterminate system.
Determinacy
This case occurs when the number of expectation errors exactly matches the
number of explosive components (k = m), or otherwise stated in the case

1.5. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (II)

45

where k = r. As shown in proposition 5, the expectation errors are then just

a linear combination of fundamental disturbances for all t since both M1 and
M2 reduce to nil matrices. Therefore, in this case, we have
1
t = V1 D11
U1 Q2 Bt

such that the overall effect of fundamental shocks on Wt is

1
(Q1 B Q1 CV1 D11
U1 Q2 B)t

while that of purely extrinsic expectation errors is nil. To get such an effect
.
in the first part of system (1.20), we shall premultiply by the matrix [I .. ]
1
where Q1 CV1 D11
U1 . Then, taking into account that W2t = 0, we have

T11 T12 T22

0
I



W1,t
W2,t



S11 S12 S22
W1,t1
=
0
0
W2,t1


Q1 Q2
Bt
+
0

Noting that the inverse of the matrix


is

T11 T12 T22

0
I

1
1
T11
T11
(T12 T22 )
0
I

we have

  1

  1

1
W1,t1
W1,t
T11 (Q1 Q2 )
T11 S11 T11
(S12 S22 )
Bt
+
=
W2,t1
W2,t
0
0
0
Now recall that Wt = Z Yt and that ZZ = I. Therefore, premultiplying the
last equation by Z, we end up with a solution of the form
Yt = My Yt1 + Me t

(1.24)

with
M =Z

1
1
T11
S11 T11
(S12 S22 )
0
0

Z and Me = Z

1
T11
(Q1 Q2 )
0

46

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Indeterminacy
This case arises as soon as the number of expectation errors is greater than the
number of explosive components (k > m), which translates into the fact that
k > r. As shown in proposition 5, the expectation errors are then not only
linear combinations of fundamental disturbances for all t but also of purely
extrinsic disturbances called sunspot variables. Then, the expectation errors
are shown to be of the form
1
t = (V1 D11
U1 Q2 B + V2 M1 )t + V2 M2 t

where both M1 and M2 can be freely chosen. This actually raises several questions. The first one is how to select M1 and M2 ? They are totally arbitrary
the only restriction we have to impose is that M1 is a (k r) matrix and M2
is a (k r) p matrix. A second one is then how to interpret these sunspots?
In order to partially circumvent these difficulties, it is useful to introduce the
notion of beliefs. For instance, this amounts to introduce new shocks the
sunspots beside the standard expectation error. In such a case, a variable
yt will be determined by its expectation at time t 1, a shock on the beliefs
that leads to a revision of forecasts, and the expectation error
yt = Et1 yt + t + t
where t is the shock on the belief, that satisfies Et1 t = 0, and t is the
expectation error. t is a k 1 vector. Then the system 1.19 rewrites
A0 Yt = A1 Yt1 + Bt + C(t + t )
which can be restated in the form
A0 Yt = A1 Yt1 + B

t
t

+ C t

where B = [B C]. Implicit in this rewriting of the system is the fact that the
belief shock be treated like a fundamental shock, therefore condition (1.21)
rewrites
Q2 B

t
t

+ Q2 C t = 0

which leads, according to proposition 5, to an expectation error of the form

1
1
t = (V1 D11
U1 Q2 B + V2 M1 )t + (V1 D11
U1 Q2 C + V2 M1 )t

1.5. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (II)

47

But, since Q2 C = U1 D11 V1 and V1 V1 + V2 V2 = I, this rewrites

1
t = (V1 D11
U1 Q2 B + V2 M1 )t + V2 (V2 + M1 )t

This shows that the expectation error is a function of both the fundamental
shocks and the beliefs.
If this latter formulation furnishes an economic interpretation to the sunspots,
it leaves unidentified the matrices M1 and M1 . From a practical point of view,
we can, arbitrarily, set these matrices to zeros and then proceed exactly as in
the determinate case, replacing B by B in the solution. This leads to


t
(1.25)
Yt = My Yt1 + Me
t
with
M =Z

1
1
(S12 S22 )
S11 T11
T11
0
0

Z and Me = Z

1
(Q1 Q2 )
T11
0

Note however, that even if we know the form of the solution, we know nothing
about the statistical properties of the t shocks. In particular, we do not know
their covariance matrix that can be set arbitrarily.

1.5.4

Using the model

In this section, we will show you how the solution may be used to study the
dynamic properties of the model from a quantitative point of view. We will
basically address two issues
1. Impulse response functions
2. Computation of moments
Impulse response functions
As we have already seen in the preceding chapter, the impulse response function of a variable to a shock gives us the expected response of the variable to
a shock at different horizons in other words this corresponds to the best
linear predictor of the variable if the economic environment remains the same
in the future. For instance, and just to remind you what it is, let us consider
the case of an AR(1) process:
xt = xt1 + (1 )x + t

48

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Assume for a while that no shocks occurred in the past, such that xt remained
steady at the level x from t = 0 to T . A unit positive shock of magnitude
occurs in T , xT is then given by
xT = x +

Figure 1.12: Impulse Response Function (AR(1))

6

xt

Time

In T + 1, no other shock feeds the process, such that xT +1 is given by

xT +1 = xT + (1 )x = x +
XT +2 is then given by
xT +2 = xT +1 + (1 )x = x + 2
therefore, as reported in figure 1.12, we have
xT +i = xT +i1 + (1 )x = x + i i > 1
In our system, obtaining impulse response functions is as simple as that, provided the solution has already been computed. Assume we want to compute

1.5. MULTIVARIATE RATIONAL EXPECTATIONS MODELS (II)

49

the response to one of the fundamental shocks (i,t Et ). On impact the

vector of endogenous variables (Yt )responds as

1
Yt = ME ei with eik =
0

if i = k
otherwise

The response as horizon j is then given by:

Yt+j = My Yt+j1 j > 0
Computation of moments
Let us focus on the computation of the moments for this economy. We will
describe two ways to do it. The first one uses a direct theoretical computation
of the moments, while the second one relies on MonteCarlo simulations.
The theoretical computation of moments can be achieved in a straightforward
way. Let us focus for a while on the covariance matrix of the state variables:
yy = E(Yt Yt )
Recall that in the most complicated case, we have
Yt = My Yt1 + ME t
with E(t t ) = ee .
Further, recall that we only consider stationary representations of the economy,
) whatever j. Hence, we have
such that SS = E(St+j St+j

yy = My yy My + My E(Yt1 t )ME + +Me E(t Yt1

)My + Me ee ME

Since both t are innovations, they are orthogonal to Yt , such that the previous
equation reduces to
yy = My yy My + Me ee ME
Solving this equation for SS can be achieved remembering that vec(ABC) =
(A C )vec(B), hence
vec(yy ) = (I My My )1 vec(ee )

50

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

The computation of covariances at leads and lags proceeds the same way. For
). From
instance, assume we want to compute jSS = E(St Stj

Yt = My Yt1 + Me t
we know that
Yt =

Myj Ytj

+ Me

j
X

Myi ti

i=0

Therefore,

E(Yt Ytj
)

Myj E(Ytj Ytj

)

+ Me

j
X

Myi E(ti Ytj

)

i=0

Since are innovations, they are orthogonal to any past value of Y , such that

0
if i < j

E(ti Ytj ) =

ee Me if i = j
Then, the previous equation reduces to

E(Yt Ytj
) = Myj yy + Me Myj ee Me

The MonteCarlo simulation is as simple as computing Impulse Response

Functions, as it just amounts to simulate a process for , impose an initial
condition for Y0 and to iterate on
Yt = My Yt1 + Me t for t = 0, . . . , T
Then moments can be computed and stored in a matrix. The experiment is
conducted N times, as N one can compute the asymptotic distribution
of the moments.

1.6

Economic examples

This section intends to provide you with some economic applications of the set
of tools we have described up to now. We will consider three examples, two of
which may be thought of as micro examples. In the first one a firm decides on
its labor demand, the second one is a macro model and endogenous growth
model `
a la Romer [1986] which allows to show that even a nonlinear model
may be expressed in linear terms and therefore may be solved in a very simple
way. The last one deals with the socalled Lucas critique which has strong
implications on the econometric side.

1.6. ECONOMIC EXAMPLES

1.6.1

51

Labor demand

We consider the case of a firm that has to decide on its level of employment.
The firm is infinitely lived and produces a good relying on a decreasing returns
to scale technology that essentially uses labor another way to think of it
would be to assume that physical capital is a fixedfactor. This technology is
represented by the production function
Yt = f0 nt

f1 2
n with f0 , f1 > 0.
2 t

Using labor incurs two sources of cost

1. The standard payment for labor services: wt nt where wt is the real wage,
which positive sequence {wt }
t=0 is taken as given by the firm
2. A cost of adjusting labor which may be justified either by appealing to
reorganization costs, training costs, and that takes the form

(nt nt1 )2 with > 0

2
Labor is then determined by maximizing the expected intertemporal profit

s 

X
1

f1 2
2
max Et
f0 nt+s nt+s wt+s nt+s (nt+s nt+s1 )
1+r
2
2
{n }
=0
s=0

First order conditions:

Finding the first order conditions associated to

this dynamic optimization problem may be achieved in various ways. Here,

we will follow Sargent [1987] and will adopt the Lagrangean approach. Let
us fix s for a while and make some accountancy in order to find all the terms
involving nt+s
in s i, i = 2, . . .
in s 1
in s
in s + 1
in s + i, i = 2, . . .

none
none


s 
1

f1 2
2
Et
f0 nt+s nt+s wt+s nt+s (nt+s nt+s1 )
1+r
2
2
s+1 



1
(nt+s+1 nt+s )2
Et
1+r
2
none

52

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Hence, finding the optimality condition associated to nt+s reduces to maxi-

mizing



s 

1
1

f1 2
2
2
Et
(nt+s+1 nt+s )
f0 nt+s nt+s wt+s nt+s (nt+s nt+s1 )
1+r
2
2
1+r 2
which yields the following first order condition

s 



1
1
Et
f0 f1 nt+s wt+s (nt+s nt+s1 ) +
(nt+s+1 nt+s ) = 0
1+r
1+r
since r is a constant this reduces to




1
(nt+s+1 nt+s ) = 0
Et f0 f1 nt+s wt+s (nt+s nt+s1 ) +
1+r
Now remark that this relationship holds whatever s, such that we may restrict
ourselves to the case s = 0 which then yields noting that nti , i > 0 belongs
to the information set
f0 f1 nt wt (nt nt1 ) +

(Et nt+1 nt ) = 0
1+r

rearranging terms


1+r
f1 (1 + r)
nt + (1 + r)nt1 +
(f0 wt ) = 0
Et nt+1 2 + r +

Finally we have the transversality condition

lim (1 + r)T (nT nT 1 )nT = 0

T +

Solving the model:

In this example, we will apply all three methods that

we have described previously. Let us first start with factorization.

The preceding equation may be rewritten using the forward operator as




f1 (1 + r)
1+r
2
P (F )nt1 F 2 + r +
(wt f 0)
F + 1 + r nt1 =

P (F ) may be factorized as
P (F ) = (F 1 )(F 2 )
Let us compute the discriminant of this second order polynomial




f1 (1 + r) 2
f1
f1
2+r+
(1 + r) + 2(2 + r) > 0
4(1 + r) = (1 + r)

1.6. ECONOMIC EXAMPLES

53

Hence, since > 0, we know that the two roots are real. Further
f1 (1 + r)
<0

f1
P (1) = (1 + r) + 2(2 + r) > 0

P (0) = 1 + r > 0


f1 (1 + r)
1

2+r+
>1
P (x) = 0 x =
2

P (1) =

P (0) being greater than 0 and since P(1) is negative, one root lies between
0 and 1, and the other one is therefore greater than 1 since lim P (x) = .
x

The system therefore satisfies the saddle path property.

Let us assume then that 1 < 1 and 2 > 1. The expectational equation
rewrites
(F 1 )(F 2 )nt1 = wt f0 (F 1 )nt1 =

1 + r wt f 0
F 2

or
nt = 1 nt1 +

1 + r f0 wt
1 + r X i
2 Et (f0 wt+i )
=

n
+
1 t1
2 1 1
2
2 F
i=0

Since 1 2 = (1 + r), this rewrites

1 X i
nt = 1 nt1 +
2 Et (f0 wt+i )

i=0

or developing the series

f0 (1 + r)
1 X i
2 Et wt+i
nt =
+ 1 nt1
(2 1)

i=0

For practical purposes let us assume that wt follows an AR(1) process of the
form
wt = wt1 + (1 )w + t
we have
Et wt+i = i wt + (1 i )w
such that nt rewrites
nt =

f0 (1 + r)
1+r
(1 + r)(1 )
w + 1 nt1

wt
(2 1) (2 1)(2 )
( 2 )

54

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

We now consider the problem from the method of undetermined coefficients

point of view, and guess that the solution takes the form
nt = 0 + 1 nt1 +

i Et wt+i

i=0

Plugging the guess in t + 1 in the Euler equation, we get

!
!

X
X
i Et+1 wt+i+1
i Et wt+i +
Et 0 + 1 0 + 1 nt1 +
i=0

i=0

!



X
f1 (1 + r)
i Et wt+i
0 + 1 nt1 +
2+r+

i=0

1+r
+(1 + r)nt1 +
(f0 wt ) = 0

which rewrites
0 (1 + 1 ) +

12 nt1

+ 1

i Et wt+i +

i=0



f1 (1 + r)
0 + 1 nt1 +
2+r+

+(1 + r)nt1 +

1+r
(f0 wt ) = 0

i Et wt+i+1

i=0

i Et wt+i

i=0

Identifying term by term, we get the system



f1 (1+r)

(1
+

2
+
r
+
0 + 1+r

0
1
f0 = 0




f
(1+r)
2 2 + r + 1
1 + (1 + r) = 0
1




f
(1+r)
1

0 1 2 + r +
1+r

=0





f
(1+r)
1
i 1 2 + r +
+ i1 = 0

The second equation of the system exactly corresponds to the second order
polynomial we solved in the factorization method. The system therefore exhibits the saddle path property so that 1 (0, 1) and 2 (1, ). Let us
recall that 1 + 2 = 2 + r + f1 (1 + r)/, such that the system for 0 and i
rewrites

0 (1 + 1 ) 2 + r +
0 2 1+r
=0

1
i = 2 i1

f1 (1+r)

0 +

1+r
f0

=0

1.6. ECONOMIC EXAMPLES

55

Therefore, we have
0 =

1+r
1
=
2

and i = i
2 0 . Finally, we have
0 =

f0 (1 + r)
(2 1)

We then find the previous solution

1 X i
f0 (1 + r)
2 Et wt+i
+ 1 nt1
nt =
(2 1)

i=0

As a final exercise, let us adopt the matricial approach to the problem. To

do so, and because this approach is essentially numerical, we need to assume a
particular process for the real wage. We will assume that it takes the preceding
AR(1) form. Further, we do not need to deal with levels in this approach such
that we will express the model in terms of deviation from its steady state. We
thus first compute this quantity, which is defined by


f1 (1 + r)
f0 w
1+r

n 2+r+
(f0 w) = 0 n =
n + (1 + r)n +

f1
Denoting n
bt = nt n and w
bt = wt w, and introducing the technical

variable zbt+1 = n
bt , the Labor demand reexpresses as


1+r
f1 (1 + r)
w
bt = 0
n
bt + (1 + r)b
zt
Et n
bt+1 2 + r +

We define the vector Yt = {b

zt+1 , n
bt , w
bt , Et n
bt+1 }. Remembering that n
bt =

Et1 n
bt + t , the system expresses as

1
0
0
1

1
1
0

2+r+

f1 (1+r)

0
0
1

0
0
0

1+r

zbt+1
n
bt
w
bt
Et n
bt+1

0
0
0
(1 + r)

0
0
1
0

0
0
0
0

0
0

t +

0
1
0
0
0
1
0
0

zbt
n
bt1
w
bt1
Et1 n
bt

We now provide you with an example of the type of dynamics this model
may generate. Figure 1.13 reports the impulse response function of labor to

56

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Table 1.1: Parameterization: labor demand
r
0.01

f0
1

f1
0.2

0.001/1

w
0.6

0.95

a positive shock on the real wage (table 1.1 reports the parameterization).
As expected, labor demand shifts downward instantaneously, but depending
on the size of the adjustment cost, the magnitude of the impact effect differs. When adjustment costs are low, the firm drastically cuts employment,
which goes back steadily to its initial level as the effects of the shock vanish. Conversely, when adjustment costs are high, the firm does not respond as
Figure 1.13: Impulse Response to a Wage Shock
Small adjustment costs ( = 0.001)
Real Wage

Labor Demand

0.8

0.6

0.4

0.2
0

10
Time

15

5
0

20

10
Time

15

20

15

20

High adjustment costs ( = 1)

Real Wage

Labor Demand

1.5
2

0.8

2.5
0.6
3
0.4
0.2
0

3.5
5

10
Time

15

20

4
0

10
Time

much as before since it wants to avoid paying the cost. Nevertheless, it remains
optimal to cut employment, so in order to minimize the cost, the firm spreads
it intertemporally by smoothing the employment profile, therefore generating
a hump shaped response of employment.

1.6. ECONOMIC EXAMPLES

Matlab Code: Labor Demand
%
% Labor demand
%
clear all
%
% Structural Parameters
%
r
= 0.02;
f0
= 1;
wb
= 0.6;
f1
= 0.2;
phi
= 1;
rho
= 0.95;
nb
= (f0-wb)/f1;
A0=[
1 -1 0 0
0 1 0 0
0 0 1 0
0 -(2+r+f1*(1+r)/phi) -(1+r)/phi 1
];
A1=[
0 0 0 0
0 0 0 1
0 0 rho 0
-(1+r) 0 0 0
];
B=[0;0;1;0];
C=[0;1;0;0];
% Call Sims Routine
[MY,ME] = sims_solve(A0,A1,B,C);
%
% IRF
%
nrep
= 20;
SHOCK = 1;
YS
= zeros(4,nrep);
YS(:,1)= ME*SHOCK;
for i = 2:nrep;
YS(:,i)=MY*YS(:,i-1);
end
T=1:nrep;
subplot(221);plot(T,Y(3,:));
subplot(222);plot(T,Y(2,:));

57

58

1.6.2

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

The Real Business Cycle Model

We consider an economy that consists of a large number of dynastic households and a large number of firms. Firms are producing a homogeneous final
product that can be either consumed or invested by means of capital and labor
services. Firms own their capital stock and hire labor supplied by the households. Households own the firms. In each and every period three perfectly
competitive markets open the markets for consumption goods, labor services, and financial capital in the form of firms shares. Household preferences
are characterized by the lifetime utility function:
Et

X
s=0

h1+
log(ct+s ) t+s
1+
s

where 0 < < 1 is a constant discount factor, ct is consumption in period

t, ht is the fraction of total available time devoted to productive activity in
period t, > 0 and > 0. We assume that there exists a central planner
that determines hours, consumption and capital accumulation maximizing the
households utility function subject to the following budget constraint
ct + it = yt

(1.26)

where it is investment, and yt is output. Investment is used to form physical

capital, which accumulates in the standard form as:
kt+1 = it + (1 )kt with 0 6 6 1

(1.27)

where is the constant physical depreciation rate.

Output is produced by means of capital and labor services, relying on a constant returns to scale technology represented by the following Cobb Douglas
production function:
yt = at kt h1
with 0 < < 1
t

(1.28)

at represents a stochastic shock to technology or Solow residual, which evolves

according to:
log(at ) = log(at1 ) + (1 ) log(a) + t

(1.29)

1.6. ECONOMIC EXAMPLES

59

The unconditional mean of at is a, || < 1 and t is a gaussian white noise

with standard deviation of . Therefore, the central planner solves
max

{ct+s ,kt+1+s }
s=0

Et

s log(ct+s )

s=0

s.t.

h1+
t+s
1+

kt+1 =yt = at kt h1
ct + (1 )kt
t
log(at ) = log(at1 ) + (1 ) log(a) + t
The set of conditions characterizing the equilibrium is given by
yt
h
t ct =(1 )
ht
yt =at kt h1
t
yt =ct + it
kt+1 =it + (1 )kt



yt+1
ct

+1
1 =Et
ct+1
kt+1

(1.30)
(1.31)
(1.32)
(1.33)
(1.34)

and the transversality condition

kt+1+s
=0
ct+s
The problem with this dynamic system is that it is fundamentally nonlinear
lim s

and therefore the methods we have developed so far are not designed to handle
it. The usual way to deal with this type of system is then to take a linear
or loglinear approximation of each equation about the deterministic steady
state. Therefore, the first step is to find the deterministic steady state.
Deterministic steady state

Recall that the steady state value of a vari-

able, x, is the value x such that xt = x for all t. Therefore, the steady state
of the RBC model is characterized by the set of equations:
y
h c =(1 )
h
y =ak h 1

(1.35)
(1.36)

y t =c + i

(1.37)

k =i + (1 )k


y
1 =Et
+1
k

(1.38)
(1.39)

60

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

From equation (1.38), we get

i = k

i
k
=

y
y

Then, equation (1.39) implies

1 (1 )
y
=
k

such that from the previous equation and (1.37)

i

c
=
=
s
=
= 1 si
c
y
1 (1 )
y

si

Then, from (1.35), we obtain

h =

1
sc

1
1+

Finally, it follows from the production function and the definition of y /k

that

y =a

1 (1 )

h , c = sc y , i = y c .

We are now in position to loglinearize the dynamic system.

Loglinearization:

A common practice in the macro literature is to take a

loglinear approximation to the equilibrium. Such an approximation is usually

taken because it delivers a natural interpretation of the coefficients in front of
the variables: these can be interpreted as elasticities. Indeed, lets consider
the following onedimensional function f (x) and lets assume that we want to
take a loglinear approximation of f around x. This would amount to have,
as deviation, a logdeviation rather than a simple deviation, such that we can
define
x
b = log(x) log(x )

Then, a restatement of the problem is in order, as we are to take an approximation with respect to log(x):
f (x) f (exp(log(x)))
which leads to the following first order Taylor expansion
f (x) f (x ) + f (exp(log(x )))exp(log(x ))b
x = f (x ) + f (x )x x
b

1.6. ECONOMIC EXAMPLES

61

Now, remember that by definition of the deterministic steady state, we have

f (x ) = 0, such that the latter equation reduces to
f (x) f (x )x x
b

Applying this technic to the system (1.30)(1.34), we end up with the system
(1 + )b
ht + b
ct ybt

(1.40)

ybt (1 )b
ht b
ht b
at = 0

(1.41)

ybt sc b
ct sibit = 0

(1.42)

b
kt+1 bit (1 )b
kt = 0

(1.43)

Et b
ct+1 b
ct (1 (1 ))(Et ybt+1 Et b
kt+1 )

(1.44)

b
at b
at1 bt

(1.45)

Note that only the last three equations of the system involve dynamics, but
they depend on variables that are defined in the first three equations. Either
we solve the first three equations in terms of the state and costate variables,
or we adapt a little bit the method. We choose the second solution.
Let us define Yt = {b
kt+1 , b
at , Et b
ct+1 } and Xt = {b
yt , b
ct , bit , b
ht }. The system can

be rewritten as a set of two equations. The first one gathers static equations
x Xt = y Yt1 + t + t

where t is the vector of expectation errors, which actually reduces to that

attached on b
ct , and

1
0
0 1

1
0
0
0
x =
=
1 sc si
0 y 0
1 1
0
1
0

0
0
0

0
1

1
0
=
0 0
0
0

The second one gathers the dynamic equations

0y Yt + 0x Et Xt+1 = 1y Yt1 + 1x + t + t

= 1

0
0

62

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

with

0
0 0 0
1
0 0
0
0 0 0
0
1 0 0x =
0y =
(1 (1 )) 0 0 0
1 (1 ) 0 1
0 0 0
1 0 0
0
1x = 0 0 0 0
1y = 0
0
0 0

0 1 0 0
0
0

1
0
=
=
0
0

From the first equation, we obtain

Xt = y Yt1 + t + t
where j = 1
x j , j = {y, , }. Furthermore, remembering that Et t+1 =
Et t+1 = 0, we have Et Xt+1 = y Yt . Hence, plugging this result and the first
equation in the second equation we get
A0 Yt = A1 Yt+1 + Bt + Ct
where A0 = 0y +0x y , A1 = 1y +1x y , B = +0x and C = +0x .
We then just use the algorithm as described previously.
Then, we make use of the result in proposition 5, to get t . Since it turns
out that the model is determinate, the expectation error is a function of the
fundamental shock t
1
t = V1 D11
U1 Q2 Bt

Plugging this result in the equation governing static equations, we end up with
1
Xt = y Yt1 + (e V1 D11
U1 Q2 B)t

Figure 1.14 then reports the impulse response function to a 1% technology

shock. These IRFs are obtained using the set of parameters reported in table
1.2.
Matlab Code: The RBC Model
clear all
% Clear memory
%
% Structural parameters
%
alpha
= 0.4;

1.6. ECONOMIC EXAMPLES

63

Table 1.2: The Real Business Cycle Model: parameters

0.4

0.988

0.025

delta
= 0.025;
rho
= 0.95;
beta
= 0.988;
%
% Deterministic Steady state
%
ysk = (1-beta*(1-delta))/(alpha*beta);
ksy = 1/ysk;
si = delta/ysk;
sc = 1-si;
% Define:
%
% Y=[k(t+1) a(t+1) E_tc(t+1)]
%
% X=[y,c,i,h]
%
ny = 3; % # of variables in vector Y
nx = 4; % # of variables in vector X
ne = 1; % # of fundamental shocks
nn = 1; % # of expectation errors
%
% Initialize the Upsilon matrices
%
UX=zeros(nx,nx);
UY=zeros(nx,ny);
UE=zeros(nx,ne);
UN=zeros(nx,nn);
G0Y=zeros(ny,ny);
G1Y=zeros(ny,ny);
G0X=zeros(ny,nx);
G1X=zeros(ny,nx);
GE=zeros(ny,ne);
GN=zeros(ny,nn);
%
% Production function
%
UX(1,1)=1;
UX(1,4)=alpha-1;
UY(1,1)=alpha;
UY(1,2)=rho;
UE(1)=1;
%
% Consumption c(t)=E(c(t)|t-1)+eta(t)
%

0.95

64

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

UX(2,2)=1;
UY(2,3)=1;
UN(2)=1;
%
% Resource constraint
%
UX(3,1)=1;
UX(3,2)=-sc;
UX(3,3)=-si;
%
% Consumption-leisure arbitrage
%
UX(4,1)=-1;
UX(4,2)=1;
UX(4,4)=1;
%
% Accumulation of capital
%
G0Y(1,1)=1;
G1Y(1,1)=1-delta;
G1X(1,3)=delta;
%
% Productivity shock
%
G0Y(2,2)=1;
G1Y(2,2)=rho;
GE(2)=1;
%
% Euler equation
%
G0Y(3,1)=1-beta*(1-delta);
G0Y(3,3)=1;
G0X(3,1)=-(1-beta*(1-delta));
G1X(3,2)=1;
%
% Solution
%
% Step 1: solve the first set of equations
%
PIY = inv(UX)*UY;
PIE = inv(UX)*UE;
PIN = inv(UX)*UN;
%
% Step 2: build the standard System
%
A0 = G0Y+G0X*PIY;
A1 = G1Y+G1X*PIY;
B
= GE+G1X*PIE;
C
= GN+G1X*PIN;
%
% Step 3: Call Sims routine
%
[MY,ME,ETA,MU_]=sims_solve(A0,A1,B,C);
%

1.6. ECONOMIC EXAMPLES

65

% Step 4: Recover the impact function

%
PIE=PIE-PIN*ETA;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
%
%
Impulse Response Functions
%
%
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
nrep
= 20;
% horizon of responses
YS
= zeros(3,nrep);
XS
= zeros(4,nrep);
Shock
= 1;
YS(:,1) = ME*Shock;
XS(:,1) = PIE;
for t=2:nrep;
YS(:,t) = MY*YS(:,t-1);
XS(:,t) = PIY*YS(:,t-1);
end
subplot(221);plot(XS(1,:));title(Output);xlabel(Time)
subplot(222);plot(XS(2,:));title(Consumption);xlabel(Time)
subplot(223);plot(XS(3,:));title(Investment);xlabel(Time)
subplot(224);plot(XS(4,:));title(Hours worked);xlabel(Time)

1.6.3

A model with indeterminacy

Let us consider the simplest new keynesian model, with the following IS curve
yt = Et yt+1 (it Et t+1 ) + gt
where yt denotes output, t is the inflation rate, it is the nominal interest rate
and gt is a stochastic shock that follows an AR(1) process of the form
gt = g gt1 + gt
the model also includes a Phillips curve that relates positively inflation to the
output gap
t = yt + Et t+1 + ut
where ut is a supply shock that obeys
ut = u ut1 + ut
For stationarity purposes, we have |g | < 1 and |u | < 1.
The model is closed by a simple Taylor rule of the form
it = t + y yt

66

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Figure 1.14: IRF to a technology shock

Output

Consumption

0.8

1.8
0.7
1.6
1.4

0.6

1.2
0.5
1
0.8
0

10
Time

15

20

0.4
0

Investment

10
Time

15

20

15

20

Hours worked

1.5

5
4

3
2

0.5

1
0
0

10
Time

15

20

0
0

10
Time

1.6. ECONOMIC EXAMPLES

67

Plugging this rule in the first equation, and remembering the definition of
expectation errors, the system rewrites
yt =Et1 yt + ty
t =Et1 t + t
gt =g gt1 + gt
ut =u ut1 + ut
(1 + y )yt =Et yt+1 t + Et t+1 + gt
t =yt + Et t+1 + ut
Defining Yt = {yt , t , gt , ut , Et yt+1 , Et t+1 } and t = {ty , t }, the system
rewrites

1
0
0
0
0
0

0
1
0
0
0
0

0
0
1
0
0
0

0
0
0
1
0
0

1 + y + 1 0 1

1
0 1 0

0
0

Yt = 0
0

0 0 0 1
0 0 0 0
0 g 0 0
0 0 u 0
0 0 0 0
0 0 0 0

0 0

0 0

1 0
t +

0 1

0 0
0 0

0
1
0
0
0
0
1
0
0
0
0
0

Yt1

0
1
0
0
0
0

The set of parameter used in the numerical experiment is reported in table

1.3. As predicted by theory of Taylor rules, a coefficient below 1 yields
indeterminacy.
Table 1.3: New Keynesian model: parameters

0.4

0.9

g
0.9

u
0.9

y
0.25

1.5/0.5

Matlab Code: A Model with Real Indeterminacy

clear all
% Clear memory
%
% Structural parameters

68

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

%
alpha
= 0.4;
gy
= 0.25;
gp
= 0.5;
rho_g
= 0.9;
rho_u
= 0.95;
lambda = 1;
beta
= 0.9;
% Define:
%
% Y=[y(t),pi(t),g(t),u(t),E_t y(t+1),E_t pi(t+1)]
%
ny = 6; % # of variables in vector Y
ne = 2; % # of fundamental shocks
nn = 2; % # of expectation errors
%
% Initialize the matrices
%
A0 = zeros(ny,ny);
A1 = zeros(ny,ny);
B
= zeros(ny,ne);
C
= zeros(ny,nn);
%
% Output
%
A0(1,1) = 1;
A1(1,5) = 1;
C(1,1) = 1;
%
% Inflation
%
A0(2,2) = 1;
A1(2,6) = 1;
C(2,2) = 1;
%
% IS shock
%
A0(3,3) = 1;
A1(3,3) = rho_g;
B(3,1) = 1;
%
% Supply shock
%
A0(4,4) = 1;
A1(4,4) = rho_u;
B(4,2) = 1;
%
% IS curve
%
A0(5,1) = 1+alpha*gy;
A0(5,2) = alpha*gp;
A0(5,3) = -1;
A0(5,5) = -1;
A0(5,6) = -alpha;

1.6. ECONOMIC EXAMPLES

69

%
% Phillips Curve
%
A0(6,1) = -lambda;
A0(6,2) = 1;
A0(6,4) = -1;
A0(6,6) = -beta;
%
% Call Sims routine
%
[MY,ME,ETA,MU_]=sims_solve(A0,A1,B,C);

1.6.4

AK growth model

Up to now, we have considered quadratic objective function in order to get

linear expectational difference equations. This may seem to be very restrictive.
However, there is a number of situations, where the dynamics generated by the
model is characterized by a linear expectational difference equation, despite
the objective function is not quadratic. We provide you with such an example
in this section.
We consider an endogenous growth model a` la Romer [1986] extended to a
stochastic environment. The economy consists of a large number of dynastic
households and a large number of firms. Firms are producing a homogeneous
final product that can be either consumed or invested by means of capital, but
contrary to the standard optimal growth model, returns to factors that can
be accumulated (namely capital) are exactly constant.
Household decides on consumption, Ct , and capital accumulation (or savings),
Kt+1 , maximizing her lifetime expected utility
max Et

s log(Ct+s )

s=0

subject to the resource constraint in the economy

Yt = Ct + It
and the law of motion of capital
Kt+1 = It + (1 )Kt with [0; 1]
It is investment, Yt denotes output, which is produced using a linear technology
of the form Yt = At Kt . At is a stochastic shock that we leave unspecified for

70

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

the moment. We may think of it as a shift on the technology, such that it

represents a technology shock.
First order conditions:

We now present the derivation of the optimal be-

havior of the consumer. The first order condition associated to the consumption/savings decisions may be obtained forming the following Lagrangean,
where t is the multiplier associated to the resource constraint
Lt = Et

s log(Ct+s ) + t (At Kt + (1 )Kt Ct Kt+1 )

s=0

Terms involving Ct :
max Et (log(Ct ) t Ct ) = max (log(Ct ) t Ct )
{Ct }

{Ct }

Therefore, the FOC associated to consumption writes

1
= t
Ct
Likewise for the saving decision, terms involving Kt+1 :
max t Kt+1 + Et [t+1 (At+1 Kt+1 + (1 )Kt+1 )]

{Kt+1 }

such that the FOC is given by

t = Et [t+1 (At+1 + 1 )]
Finally, we impose the socalled transversality condition


KT +1
T
lim Et
=0
T
CT
Solving the dynamic system:

Plugging the first order condition on con-

sumption in the Euler equation, we get



1
1
= Et
(At+1 + 1 )
Ct
Ct+1
This system seems to be nonlinear, but we can make it linear very easily.
Indeed, let us multiply both sides of the Euler equation by Kt+1 , we get


Kt+1
Kt+1
= Et
(At+1 + 1 )
Ct
Ct+1

1.6. ECONOMIC EXAMPLES

71

But the resource constraint states that

Kt+1 + Ct = Kt (At + 1 ) Kt+2 + Ct+1 = Kt+1 (At+1 + 1 )
such that the Euler equation rewrites




Kt+2 + Ct+1
Kt+2
Kt+1
= Et 1 +
= Et
Ct
Ct+1
Ct+1
Let us denote Xt = Kt+1 /Ct , the latter equation rewrites
Xt = Et (1 + Xt+1 )
which has the same form as (1.2). As we have already seen, the solution for
such an equation can be easily obtained iterating forward. We then get
Xt = lim

T
X
k=0

k + lim T Et (XT +1 )
T

The second term in the right hand side of the latter equation corresponds
precisely to the transversality condition. Hence, Xt reduces to
Xt =

Kt+1 =
Ct
1
1

Plugging this relation in the resource constraint, we get

Kt+1 = (At + 1 )Kt
and
Ct = (1 )(At + 1 )Kt
Time series properties

Let us consider the solution for capital accumula-

tion. Taking logs, we get

log(Kt+1 ) = log(Kt ) + log((At + 1 ))
since At is an exogenous stochastic process, we immediately see that the process may be rewritten as
log(Kt+1 ) = log(Kt ) + t
where t log((At + 1 )). Such that we see that capital is an non
stationary process (an I(1) process) more precisely a random walk. Since

72

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

consumption, output and investment are just a linear function of capital, the
nonstationarity of capital translates into the non stationarity of these variables. Nevertheless, as can be seen from the law of motion of consumption, for
example, log(Ct ) log(Kt ) is a stationary process. Kt and Ct are then said
to be cointegrated with a cointegrating vector (1, 1).
This has extremely important economic implications, that may be analyzed
in the light of the impulse response functions, reported in figure 1.15. In fact,
figure 1.15 reports two balanced growth paths for each variable: The first one
corresponds to the path without any shock, the second one corresponds to
the path that includes a non expected positive shock on technology in period
10. As can be seen, this shock yields a permanent increase in all variables.
Therefore, this model can account for the fact that countries may not converge.
Why is that so? The answer to this question is actually simple and may be
Figure 1.15: Impulse response functions
Output

Consumption

0.052

0.0135

0.05

0.013

0.048

0.0125

0.046

0.012

0.044

0.0115

0.042

0.011

0.04

0.0105

0.038

10

20

30

40

50

0.01

10

20

30

Time

Time

Investment

Capital

40

50

40

50

0.038
1.3

0.036
0.034

1.2

0.032
1.1
0.03
0.028

10

20

30
Time

40

50

10

20

30
Time

understood if we go back to the simplest Solow growth model. Assume there

is a similar shock in the Solow growth model, output increases on impact
and since income increases so does investment yielding higher accumulation.
Because the technology displays decreasing returns to capital in the solow
growth model, the marginal efficiency of capital decreases reducing incentives

1.6. ECONOMIC EXAMPLES

73

to investment so that capital accumulation slows down. The economy then

goes back to its steady state. Things are different in this model: Following
a shock, income increases. This triggers faster accumulation, but since the
marginal productivity of capital is totally determined by the exogenous shock,
there is no endogenous force that can drive the economy back to its steady
state. Therefore, each additional capital is kept forever.
This implies that following shocks, the economy will enter an ever growing
regime. This may be seen from figure 1.16 which reports a simulated path for
each variable. These simulated data may be used to generate time moments on
Figure 1.16: Simulated data
Output

Consumption

0.09

0.024

0.08

0.022
0.02

0.07

0.018
0.06
0.016
0.05

0.014

0.04
0.03

0.012
0

50

100
Time

150

200

0.01

50

Investment

100
Time

150

200

150

200

Capital

0.07

2.5

0.06
2
0.05
0.04
1.5
0.03
0.02

50

100
Time

150

200

50

100
Time

the rate of growth of each variable, which estimates are reported in table (1.4)
and which distributions are represented in figures 1.171.20. It is interesting
to note that all variables exhibit when taken in loglevels a spurious
correlation with output that just reflects the existence of a common trend due
to the balanced growth path hypothesis.

74

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Table 1.4: MonteCarlo Simulations

Corr(.,Y )

Corr(.,Y )

Y
0.40
0.79
1.00
-0.01
Y
0.99

C
0.40
0.09
0.30
0.93
C
0.99

I
0.40
1.06
0.99
-0.02
I
0.99

K
0.40
0.09
-0.08
0.93
K
0.99

Figure 1.17: Rates of growth: distribution of mean

Output

Consumption

200

200

150

150

100

100

50

50

4
Time

4
Time

x 10

Investment
200

150

150

100

100

50

50

4
Time

6
3

x 10

Capital

200

6
3

x 10

0
2.5

3.5

4
Time

4.5

5.5
3

x 10

1.6. ECONOMIC EXAMPLES

75

Figure 1.18: Rates of growth: distribution of standard deviation

Output

Consumption

200

200

150

150

100

100

50

50

8
Time

10
x 10

0.5

1
Time

Investment
200

150

150

100

100

50

50

0.01

2
x 10

Capital

200

0
0.008 0.009

1.5

0.011 0.012 0.013 0.014

Time

0.5

1
Time

1.5

2
x 10

Figure 1.19: Rates of growth: distribution of correlation with Y

Output

Consumption

5000

250

4000

200

3000

150

2000

100

1000

50

0
60

40

20

0
Time

20

40

60

0
0.2

0.25

Investment

0.3

0.35
Time

0.4

0.45

Capital

250

200

200

150

150
100
100
50

50
0
0.998

0.9985

0.999 0.9995
Time

1.0005

0
0.4

0.2

0
Time

0.2

0.4

76

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Figure 1.20: Rates of growth: distribution of first order autocorrelation

Output

Consumption

200

250
200

150

150
100
100
50

0
0.4

50

0.2

0
Time

0.2

0.4

0
0.7

0.75

Investment
200

150

150

100

100

50

50

0.2

0
Time

0.85
Time

0.9

0.95

Capital

200

0
0.4

0.8

0.2

0.4

0
0.4

0.2

0
Time

0.2

0.4

1.6. ECONOMIC EXAMPLES

Matlab Code: AK growth model
%
% AK growth model
%
long
= 200;
nsim
= 5000;
nrep
= 50;
%
% Structural parameters
%
gx
= 1.004;
beta
= 0.99;
delta
= 0.025;
rho
= 0.95;
se
= 0.0079;
ab
= (gx-beta*(1-delta))/beta;
K0
= 1;
%
% IRF
%
K1(1)
= K0;
K2(1)
= K0;
a2
= zeros(nrep,1);
K1
= zeros(nrep,1);
K2
= zeros(nrep,1);
K1(1)
= K0;
K2(1)
= K0;
a2(1)
= log(ab);
e
= zeros(nrep,1);
e(11)
= 10*se;
T=[1:nrep];
for i
= 2:nrep;
a2(i)= rho*a2(i-1)+(1-rho)*log(ab)+e(i);
K1(i)= beta*(ab+1-delta)*K1(i-1);
K2(i)= beta*(exp(a2(i-1))+1-delta)*K2(i-1);
end;
C1 = (1-beta)*(ab+1-delta).*K1;
Y1 = ab*K1;
I1 = Y1-C1;
C2 = (1-beta)*(exp(a2)+1-delta).*K2;
Y2 = exp(a2).*K2;
I2 = Y2-C2;
Y=[Y1(:) Y2(:)];
C=[C1(:) C2(:)];
K=[K1(:) K2(:)];
I=[I1(:) I2(:)];
%
% Simulations
%
cx=zeros(nsim,4);
mx=zeros(nsim,4);
sx=zeros(nsim,4);
rx=zeros(nsim,4);

77

78

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

for s
= 1:nsim;
disp(s)
randn(state,s);
e
= randn(long,1)*se;
a
= zeros(long,1);
K
= zeros(long,1);
a(1) = log(ab)+e(1);
K(1) = K0;
for
i
= 2:long;
a(i)= rho*a(i-1)+(1-rho)*log(ab)+e(i);
K(i)= beta*(exp(a(i-1))+1-delta)*K(i-1);
end;
C
= (1-beta)*(exp(a)+1-delta).*K;
Y
= exp(a).*K;
I
= Y-C;
X
= [Y C I K];
dx
= diff(log(X));
mx(s,:)
= mean(dx);
sx(s,:)
= std(dx);
tmp
= corrcoef(dx);cx(s,:)=tmp(1,:);
tmp
= corrcoef(dx(2:end,1),dx(1:end-1,1));ry=tmp(1,2);
tmp
= corrcoef(dx(2:end,2),dx(1:end-1,2));rc=tmp(1,2);
tmp
= corrcoef(dx(2:end,3),dx(1:end-1,3));ri=tmp(1,2);
tmp
= corrcoef(dx(2:end,4),dx(1:end-1,4));rk=tmp(1,2);
rx(s,:)
= [ry rc ri rk];
end;
disp(mean(mx))
disp(mean(sx))
disp(mean(cx))
disp(mean(rx))

1.6.5

Announcements

In the last two examples, we will help you to give an answer to this crucial
question:
Why do these two guys annoy us with rational expectations?
In this example we will show you how different may the impulse response to a
shock be different depending on the fact that the shock is announced or not.
To illustrate this issue, let us go back to the problem of asset pricing. Let pt
be the price of a stock, dt be the dividend which will be taken as exogenous
and r be the rate of return on a riskless asset, assumed to be held constant
over time. As we have seen earlier, standard theory of finance states that when
agents are risk neutral, the asset pricing equation is given by:
Et pt+1 pt dt
+
=r
pt
pt

1.6. ECONOMIC EXAMPLES

79

or equivalently

1
1
Et pt+1 +
dt
1+r
1+r
Let us now consider that the dividend policy of the firm is such that from
pt =

period 0 on, the firm serves a dividend equal to d0 . The price of the asset is
therefore given by
i

1 X
1
d0
pt =
Et dt+i =
1+r
1+r
r
i=0

If, in period T , the firm unexpectedly decides to serve a dividend of d1 > d0 ,

the price of the asset will be given by
d1
t > T
pt =
r
In other words, the price of the asset shifts upward to its new level, as shown
in the upperleft panel of figure 1.21. Let us now assume that the firm anFigure 1.21: Asset pricing behavior
t0=10, T=40

Unexpected shock
2.5

2.5

1.5

1.5

0.5

10

20

30
Time

40

50

60

0.5

10

20

t0=20, T=40
2.5

1.5

1.5

10

20

30
Time

40

50

60

40

50

60

t0=30, T=40

2.5

0.5

30
Time

40

50

60

0.5

10

20

30
Time

nounces in t0 < T that it will raise its dividend from d0 to d1 in period T .

This dramatically changes the behavior of the asset price, as the structure of
information is totally modified. Indeed, before the shock is announced by the
firm, the level of the asset price establishes at
d0
pt =
r

80

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

as before. In period t0 things change as the individuals now know that in

T t0 period the price will be different, this information is now included in
the information set they use to formulate expectations. Hence, from period
t0 to T , they take this information into account in their calculation, and the
asset price is now given by
pt =
=
=

it
it
T 1 

1 X
1
1 X
1
d0 +
d1
1+r
1+r
1+r
1+r
i=t
i=T
it
it
T
1 

X
X
1
1
1
1
d0 +
(d1 d0 + d0 )
1+r
1+r
1+r
1+r
i=t
i=T
it
it


1
1
1 X
1 X
d0 +
(d1 d0 )
1+r
1+r
1+r
1+r
i=t

i=T

Denoting j = i t in the first sum and = i T in the second, we have

j
+T t


1
1
1 X
1 X
d0 +
(d1 d0 )
pt =
1+r
1+r
1+r
1+r
j=0
=0

T t 

d0
1
d1 d0
=
+
r
1+r
r
Finally, from T on, the shock has taken place, such that the value of the asset
is given by

d1
r
Hence, the dynamics of the asset price is given by
d0
for t < t0

r
T t 


d1 d0
d0
1
pt =
for t0 6 t 6 T
+ 1+r
r

dr1
for t > T
r
pt =

Hence, compared to the earlier situation, there is now a transition phase that
takes place as soon as the individuals has learnt the news and exploits this
additional piece of information when formulating her expectations. This dynamics is depicted in figure 1.21 for different dates of announcement.

1.6.6

The Lucas critique

As a last example, we now have a look at the socalled Lucas critique. One
typical answer to the question raised in the previous section may be found

1.6. ECONOMIC EXAMPLES

81

in the socalled Lucas critique (see e.g. Lucas [1976]) , or the econometric
policy evaluation critique, which asserts that because the apparently (for old
fashioned econometricians) structural parameters of a model may change when
policy changes, standard econometrics may not be used to study alternative
regimes. In order to illustrate this point, let us go back to the simplest example
we were dealing with:
yt = aEt yt+1 + bxt
xt = xt1 + t
which solution is given by

b
xt
1 a
Now let us assume for a while that yt denotes output and xt is money, which is
yt =

discretionary provided by a central bank. An econometrician that has access

to data on output and money would estimate the reduced form of the model
yt = xt
where
b should converge to b/(1 a). Now the central banker would like to
evaluate the implications of a new monetary policy from t = T on
xt = xt1 + t with >

What should be done then? The oldfashioned econometrician would do the

following:
1. Take the estimated reduced form: yt =
bxt
2. Simulate paths for the new xt process

3. Analyse the properties of the time series

Stated like that all seems OK. But such an approach is totally false. Indeed,
underlying the rational expectations hypothesis is the fact that the agents
know the overall structure of the model, therefore, the agents know that from
t = T on the new monetary policy is
xt = xt1 + t with >

82

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Figure 1.22: The Lucas Critique
1.8
Misspecified
Correct
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0

10

15

20
Time

25

30

35

40

the model needs then to be solved again to yield

yt =

b
xt
1 a

Therefore the econometrician should reestimate the reduced form to get

b t
yt = x

Keeping the old reduced form implies a systematic bias of

ab( )
(1 a)(1 a)
To give you an idea of the type of mistake one may do, we report in figure
1.22 the impulse response to a monetary shock in the second monetary rule
when the old reduced form (misspecified) and the new one (correct) are used.
As it should be clear to you using the wrong rule leads to a systematic bias
in policy evaluation since it biases in this case the impact effect of the
policy. Why is that so? Because the rational expectations hypothesis implies
that the expectation function is part of the solution of the model.Keep in mind
that solving a RE model amounts to find the expectation function.
Hence, from an econometric point of view, the rational expectations hypothesis
has extremely important implications since they condition the way we should

1.6. ECONOMIC EXAMPLES

83

think of the model, solve the model and therefore evaluate and test the model.
This will be studied in the next chapter.

84

CHAPTER 1. EXPECTATIONS AND ECONOMIC DYNAMICS

Bibliography
Blanchard, O. and C. Kahn, The Solution of Linear Difference Models under
Rational Expectations, Econometrica, 1980, 48 (5), 13051311.
Blanchard, O.J. and S. Fisher, Lectures on Macroeconomics, Cambridge: MIT
Press, 1989.
Lubik, T.A. and F. Schorfheide, Computing Sunspot Equilibria in Linear Rational Expectations Models, Journal of Economic Dynamics and Control,
2003, 28, 273285.
Lucas, R., Econometric policy Evaluation : a Critique, in K. Brunner and
A.H. Meltzer, editors, The Phillips Curve and Labor Markets, Amsterdam: NorthHolland, 1976.
Muth, J.F., Optimal Properties of Exponentially Weighted Forecasts, Journal
of the American Statistical Association, 1960, 55.
, Rational Expections and the Theory of Price Movements, Econometrica, 1961, 29, 315335.
Romer, P., Increasing Returns and Long Run Growth, Journal of Political
Economy, 1986, 94, 10021037.
Sargent, T., Macroeconomic Theory, MIT Press, 1979.
Sargent, T.J., Dynamic Macroeconomic Theory, Londres: Harvard University
Press, 1987.
Sims, C., Solving Linear Rational Expectations Models, manuscript, Princeton
University 2000.

85

86

BIBLIOGRAPHY

Contents
1 Expectations and Economic Dynamics

1.1

The rational expectations hypothesis . . . . . . . . . . . . . . .

1.2

A prototypical model of rational expectations . . . . . . . . . .

1.2.1

Sketching up the model . . . . . . . . . . . . . . . . . .

1.2.2

Forward looking solutions: |a| < 1 . . . . . . . . . . . .

1.2.3

Backward looking solutions: |a| > 1 . . . . . . . . . . .

15

1.2.4

One step backward: bubbles . . . . . . . . . . . . . . . .

18

A step toward multivariate Models . . . . . . . . . . . . . . . .

23

1.3.1

The method of undetermined coefficients

. . . . . . . .

24

1.3.2

Factorization . . . . . . . . . . . . . . . . . . . . . . . .

28

1.3.3

A matricial approach . . . . . . . . . . . . . . . . . . . .

29

Multivariate Rational Expectations Models (The simple case) .

33

1.4.1

Representation . . . . . . . . . . . . . . . . . . . . . . .

33

1.4.2

Solving the system . . . . . . . . . . . . . . . . . . . . .

35

Multivariate Rational Expectations Models (II) . . . . . . . . .

38

1.5.1

Preliminary Linear Algebra . . . . . . . . . . . . . . . .

38

1.5.2

Representation . . . . . . . . . . . . . . . . . . . . . . .

39

1.5.3

Solving the system . . . . . . . . . . . . . . . . . . . . .

40

1.5.4

Using the model . . . . . . . . . . . . . . . . . . . . . .

47

Economic examples . . . . . . . . . . . . . . . . . . . . . . . . .

50

1.6.1

Labor demand . . . . . . . . . . . . . . . . . . . . . . .

51

1.6.2

The Real Business Cycle Model . . . . . . . . . . . . . .

58

1.6.3

A model with indeterminacy . . . . . . . . . . . . . . .

65

1.6.4

AK growth model . . . . . . . . . . . . . . . . . . . . .

69

1.6.5

Announcements . . . . . . . . . . . . . . . . . . . . . . .

78

1.3

1.4

1.5

1.6

87

88

CONTENTS
1.6.6

The Lucas critique . . . . . . . . . . . . . . . . . . . . .

80

List of Figures
1.1

The regular case . . . . . . . . . . . . . . . . . . . . . . . . . .

10

1.2

Forward Solution . . . . . . . . . . . . . . . . . . . . . . . . . .

12

1.3

The irregular case . . . . . . . . . . . . . . . . . . . . . . . . .

16

1.4

Backward Solution . . . . . . . . . . . . . . . . . . . . . . . . .

17

1.5

Deterministic Bubble . . . . . . . . . . . . . . . . . . . . . . . .

20

1.6

Bursting Bubble . . . . . . . . . . . . . . . . . . . . . . . . . .

22

1.7

Backwardforward solution . . . . . . . . . . . . . . . . . . . .

27

1.8

Geometrical interpretation of eigenvalues/eigenvectors . . . . .

30

1.9

A source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31

1.10 A sink: indeterminacy . . . . . . . . . . . . . . . . . . . . . . .

32

1.11 The saddle path

. . . . . . . . . . . . . . . . . . . . . . . . . .

33

1.12 Impulse Response Function (AR(1)) . . . . . . . . . . . . . . .

48

1.13 Impulse Response to a Wage Shock . . . . . . . . . . . . . . . .

56

1.14 IRF to a technology shock . . . . . . . . . . . . . . . . . . . . .

66

1.15 Impulse response functions

. . . . . . . . . . . . . . . . . . . .

72

1.16 Simulated data . . . . . . . . . . . . . . . . . . . . . . . . . . .

73

1.17 Rates of growth: distribution of mean . . . . . . . . . . . . . .

74

1.18 Rates of growth: distribution of standard deviation . . . . . . .

75

1.19 Rates of growth: distribution of correlation with Y . . . . . .

75

1.20 Rates of growth: distribution of first order autocorrelation . . .

76

1.21 Asset pricing behavior . . . . . . . . . . . . . . . . . . . . . . .

79

1.22 The Lucas Critique . . . . . . . . . . . . . . . . . . . . . . . . .

82

89

90

LIST OF FIGURES

List of Tables
1.1

Parameterization: labor demand . . . . . . . . . . . . . . . . .

56

1.2

The Real Business Cycle Model: parameters . . . . . . . . . . .

63

1.3

New Keynesian model: parameters . . . . . . . . . . . . . . . .

67

1.4

MonteCarlo Simulations . . . . . . . . . . . . . . . . . . . . .

74

91