1 Cointegration and Error Correction Model

This part discusses a new theory for a regression with nonstationary unit

root variables. In general, this should require a di¤erent treatment from a

conventional regression with stationary variables, which has been covered so

far. In particular, we focus on a class of the linear combination of the unit

root processes known as cointegrated process.

1.1 Stylized Facts about Economic Time Series

Casual inspection of most economic time series data such as GNP and prices

reveals that these series are non stationary. We can characterize some of the

key feature of the various series as follows:

1. Most of the series contain a clear trend. In general, it is hard to dis-

tinguish between trend stationary and di¤erence stationary processes.

2. Some series seem to meander. For example, the pound/dollar exchange

rate shows no particular tendency to increase or decrease. The pound

seems to go through sustained periods of appreciation and then depre-

ciation with no tendency to revert to a long-run mean. This type of

random walk behavior is typical of unit root series.

1
 3. Any shock to a series displays a high degree of persistence. For ex-

ample, the UK industrial production plummeted in the late 1970s and

not returning to its previous level until mid 80s. Overall the general

consensus is at least empirically that most macro economic time series

follow a unit root process.

4 Some series share co-movements with other series. For example, short-

and long-term interest rate, though meandering individually, track each

other quite closely maybe due to the underlying common economic

forces. This phenomenon is called cointegration.

A note on notations: It is widely used that the unit root process is called

an integrated of order 1 or for short I(1) process. On the other hand, a

stationary process is called an I(0) process.1

1
In this regard we can de…ne I(d) process, and d is a number of di¤erencing to render
the series stationary.

2
1.2 Spurious Regression

Suppose that two I(1) processes, yt and xt , are independently distributed.

We now consider the following simple regression:

yt = xt + error:

Clearly, there should be no systematic relationship between y and x, and

therefore, we should expect that an OLS estimate of should be close to zero,

or insigni…cantly di¤erent from zero, at least as the sample size increases.

But, as will be shown below, this is not the case. This phenomenon originated

from Yule (1926) was called “a nonsense correlation.”

Example 1 There are some famous examples for spurious correlation. One

is that of Yule (1926, Journal of the Royal Statistical Society), reporting a

correlation of 0.95 between the proportion of Church of England marriages to

all marriages and the mortality rate over the period 1866-1911. Yet another

is the econometric example reported by Hendry (1980, Economica) between

the price level and the cumulative rainfall in the UK.2

2
This relation proved resilient to many econometric diagnostic tests and was humorously
advanced as a new theory of in‡ation.

3
 As we have come to understand in recent years, it is commonality of

(stochastic) trending mechanisms in data that often leads to these spurious

relations. What makes the phenomenon dramatic is that it occurs even when

the data are otherwise independent.

In a prototypical spurious regression the …tted coe¢ cients are statistically

signi…cant when there is no true relationship between the dependent variable

and the regressors. Using Monte Carlo simulations Granger and Newbold

(1974, Journal of Econometrics) showed this phenomenon. Phillips (1986,

Journal of Econometrics) derived an analytic proof. These results are sum-

marized in the following theorem:

Theorem 1 (Spurious Regression) Suppose that y and x are independent

I(1) variables generated respectively by

yt = yt 1 + "t ;

xt = xt 1 + et ;

2 2
where "t iid (0; ") and et iid (0; e ), and "t and et are independent of

4
each other. Consider the regression,

y t = x t + ut : (1)

Then, as T ! 1,

(a) The OLS estimator of obtained from (1), denoted ^ , does not converge

to (true value of) zero.

(b) The t-statistic testing = 0 in (1) diverges to in…nity.

In sum, in the case of spurious regression, ^ takes any value randomly, and

its t-statistic always indicates signi…cance of the estimate. Though a formal

testing procedure will be needed to detect evidence of the spurious regression

or cointegration (see below), one useful guideline is that we are likely to face

with the spurious relation when we …nd a highly signi…cant t-ratio combined

with a rather low value of R2 and a low value of the Durbin-Watson statistic.

1.3 Cointegration

Economic theory often suggests that certain subset of variables should be

linked by a long-run equilibrium relationship. Although the variables under

5
consideration may drift away from equilibrium for a while, economic forces

or government actions may be expected to restore equilibrium.

Example 2 Consider the market for tomatoes in two parts of a country, the

north and the south with prices pnt and pst respectively. If these prices are

equal the market will be in equilibrium. So pnt = pst is called an attractor. If

the prices are unequal it will be possible to make a pro…t by buying tomatoes

in one region and selling them in the other. This trading mechanism will

be inclined to equate prices again, raising prices in the buying region and

lowering them in selling region.

When the concept of equilibrium is applied to I(1) variables, cointegra-

tion occurs; that is, cointegration is de…ned as a certain stationary linear

combination of multiple I(1) variables.

Example 3 Consider the consumption spending model. Although both con-

sumption and income exhibit a unit root, over the long run consumption tends

to be a roughly constant proportion of income, so that the di¤erence between

the log of consumption and log of income appears to be a stationary process.

Example 4 Another well-known example is the theory of Purchasing Power

Parity (PPP). This theory holds that apart from transportation costs, goods

6
should sell for the same e¤ective price in two countries. Let Pt denote the

index of the price level in US (in dollars per good), Pt denote the price

index for UK (in pounds per good), and St the rate of exchange between the

currencies (in dollars per pound). Then the PPP holds that Pt = St Pt ;

taking logarithm, pt = st + pt ; where pt = ln(Pt ); st = ln(St ); pt = ln(Pt ): In

practice, errors in measuring prices, transportation costs and di¤erences in

quality prevent PPP from holding exactly at every date t: A weaker version of

the hypothesis is that the variable zt de…ned by zt = pt st pt is stationary,

even though the individual elements (pt ; st ; pt ) are all I(1).

Cointegration brings with it two obvious econometric questions. The …rst is

how to estimate the cointegrating parameters and the second is how to test

whether two or more variables are cointegrated or spurious.

We …rst examine estimation of the cointegrating regression. Consider the

simple time series regression,3

yt = x t + ut ; (2)

3
The deterministic regressors such as intercept and a linear time trend can be easily
accommodated in the regression without changing the results in what follows.

7
where xt is an I(1) variable given by

xt = xt 1 + et : (3)

Since xt is I(1), it follows that yt is I(1). But, for yt and xt to be cointegrated,

their linear combination, ut = yt xt must be stationary. Thus, we assume:

2
Assumption 2.1. ut is iid process with mean zero and variance u.

Assumption 2.2. et ’s are stationary and independently distributed of ut .

Assumption 2.1 ensures that there exists a stationary cointegrating rela-

tionship between y and x. Assumption 2.2 implies that xt is exogenous, i.e.,

E (xt ut ) = 0, which has been one of the standard assumption.1

Theorem 2 (Cointegrating Regression) Consider the OLS estimator of

obtained from (2). Under Assumptions 2.1 and 2.2, as T ! 1,

(a) The OLS estimator ^ is consistent; that is, ^ !p , and T ^ has

an asymptotic normal distribution.

(b) The t-statistic testing = 0 in (2) converges to a standard normal

random variable.

8
 This theorem clearly indicates that most estimation and inference results

as obtained for the regression with stationary variables can be extended to

the cointegrating regression.

The result in theorem can be readily extended to a multiple cointegrating

regression with k regressors:

0
yt = xt + ut ;

where xt = (x1t ; x2t ; : : : ; xkt )0 is a k-dimensional I(1) regressors given by

xt = xt 1 + et ;

where et = (e1t ; e2t ; : : : ; ekt )0 are k-dimensional stationary disturbances, and

independently distributed of ut . In this case we need one additional condition:

Assumption 2.3. xt ’s not cointegrated among themselves.

Violation of Assumption 2.3 means that there may be more than one

cointegrating relations. Such a case cannot be covered in a single equation

approach, and will be covered in the system VAR approach to cointegra-

tion. Assumption 2.3 is similar to the multicollinearity assumption made in

9
a multiple regression with stationary regressors.

In this more general case the OLS estimator of is also consistent, and

has an asymptotic normal distribution such that multiple restrictions on

can be tested in a standard way using the Wald statistic, which is asymptot-

2
ically distributed.

1.3.1 Residual-based Test for Cointegration

We …nd that the fundamentally di¤erent conclusion is made between spurious

regression and cointegration. Therefore, the detection of cointegration is very

important in practice prior to estimation.

One of most popular tests for (a single) cointegration has been suggested

by Engle and Granger (1987, Econometrica). Consider the multiple regres-

sion:
0
yt = xt + ut ; t = 1; :::; T; (4)

where xt = (x1t ; x2t ; : : : ; xkt )0 is the k-dimensional I(1) regressors. Notice

that for yt and xt to be cointegrated, ut must be I(0). Otherwise it is

spurious. Thus, a basic idea behind is to test whether ut is I(0) or I(1).

The Engle and Granger cointegration test is carried out in two steps:4

4
Of course you need to pretest the individual variables for unit roots. By de…nition of

10
 1. Run the OLS regression of (4) and obtain the residuals by

u^t = yt ^ 0 xt ; t = 1; :::; T;

where ^ are the OLS estimate of .

2. Apply a unit root test to u^t by constructing an AR(1) regression for

u^t :

u^t = u^t 1 + "t : (5)

That is, do the DF t-test of H0 : = 1 against H1 : < 1 in (5).5

This is called the residual-based EGDF cointegration test. Strictly, it is

the test of no-cointegration, because the null of unit root in u^t implies that

there is no-cointegration between y and x. So if you reject H0 : = 1 in (5),

you may conclude that there is a cointegration and vice versa.

Notice that the asymptotic distribution of t-statistic for = 1 in (5) is

also non-standard, but more importantly di¤erent from that of the univariate

DF unit test. Main di¤erence stems from the fact that one needs to allow

cointegration you need to ascertain that all the variable involved are I(1):
5
Since u
^t is a zero-mean residual process, there is no need to include an intercept term
here.

11
for estimation uncertainty through ^ in the …rst step. The resulting test

distribution thus depends on the dimension of the regressors, k.

As in the case of the univariate unit root test, there are the three speci-

…cations with di¤erent deterministic components:

0
yt = xt + vt ; (6)

0
y t = a0 + xt + vt ; (7)

0
y t = a0 + a1 t + xt + vt ; (8)

where a0 is an intercept and a1 is the linear trend coe¢ cient. The three sets of

critical values of the EGDF tests have been provided by Engle and Yoo (1987,

Journal of Econometrics) and Phillips and Ouliaris (1990, Econometrica) for

di¤erent values of k.6

Since the serial correlation is also often a problem in practice, it is common

to use an augmented version of EGDF test; that is, extend (5) to

p 1
X
u^t = '^
ut 1 + i u^t 1 + "t ;
i=1

6
For example, Micro…t provides the corresponding critical values when estimating and
testing.

12
and do the t-test for ' = 0.

Notice that all the problems that a- ict the unit root tests also a- ict

the residual-based cointegration tests. In particular, the asymptotic critical

values may be seriously misleading in small samples. Unfortunately, the

cointegration tests are often severely lacking in power especially because of

the imprecision or uncertainty of estimating in the …rst step. Thus, failure

to reject the null of no-cointegration is common in application, which may

provide only weak evidence that two or more variables are not cointegrated.

1.4 Error Correction Model

The cointegrating regression so far considers only the long-run property of the

model, and does not deal with the short-run dynamics explicitly.7 Clearly,

a good time series modelling should describe both short-run dynamics and

the long-run equilibrium simultaneously. For this purpose we now develop

an error correction model (ECM). Although ECM has been popularized after

Engle and Granger, it has a long tradition in time series econometrics dating

back to Sargan (1964) or being embedded in the London School of Economics

7
Here the long-run relationship measures any relation between the level of the variables
under consideration while the short-run dynamics measure any dynamic adjustments be-
tween the …rst-di¤erence’s of the variables.

13
tradition.

To start, we de…ne the error correction term by

t = yt xt ;

where is a cointegrating coe¢ cient. In fact, t is the error from a regression

of yt on xt . Then an ECM is simply de…ned as

yt = t 1 + xt + ut ; (9)

where ut is iid. The ECM equation (9) simply says that yt can be explained

by the lagged t 1 and xt . Notice that t 1 can be thought of as an equi-

librium error (or disequilibrium term) occurred in the previous period. If it

is non-zero, the model is out of equilibrium and vice versa.

Example 5 Consider the simple case where xt = 0. Suppose that t 1 > 0,

which means that yt 1 is too high above its equilibrium value, so in order to

restore equilibrium, yt must be negative. This intuitively means that the

error correction coe¢ cient must be negative such that (9) is dynamically

stable. In other words, if yt 1 is above its equilibrium, then it will start falling

in the next period and the equilibrium error will be corrected in the model,

14
hence the term error correction model.

Notice that is called the long-run parameter, and and are called

short-run parameters. Thus the ECM has both long-run and short-run prop-

erties built in it. The former property is embedded in the error correction

term t 1 and the short-run behavior is partially but crucially captured by the

error correction coe¢ cient, . All the variables in the ECM are stationary,

and therefore, the ECM has no spurious regression problem.

Example 6 Express the (static) cointegrating model (2) in an ECM form.

First, notice that

yt (yt 1 yt 1 ) = xt + (xt 1 xt 1 ) + ut ;

yt = (yt 1 xt 1 ) + x t + ut :

Therefore, the associated error correction model becomes

yt = t 1 + xt + ut ;

where the error correction coe¢ cient is -1 by construction, meaning the per-

fect adjustment or error correction is made every period, which is unduly

15
restrictive and unlikely to happen in practice.

In general, the error correction term t 1 is unknown a priori, and needs

to be estimated. In the case of cointegration the following Engle and Granger

two-step procedure can be used:

1. Run a (cointegrating) regression of y on x and save the residuals, ^t =

yt ^ xt .

2. Run an ECM regression of y on ^t and x,

yt = ^t 1 + x t + ut :

Example 7 ECM model is also used in the Present Value (PV) model for

stocks. A PV model relates the price of a stock to the discounted sum of its

expected future dividends. It’s noted that 1. stock prices and dividends must

be cointegrated if dividends are I(1); 2. Persistent movements in dividends

have much larger e¤ects on price than transitory movements. Thus an ap-

proximate PV model with time-varying expected returns are introduced, an

ECM model. This is also called the dynamic Gordon model in …nance.

Remark 1 We only cover single equation modelling and testing for 1 coin-

16
tegration relationship. When there are more than 2 variables, there could

be more than 1 cointegration relations among them. Then a single equation

analysis cannot provide information about the number of cointegrating rela-

tions and cannot test several cointegration relations jointly. So we rely on a

system analysis which could solve the two problems. Namely, we use a VAR

(Vector Autoregressive) model and use Johansen’s test to test the reduced

rank restriction on the coe¢ cients of a VAR.

2.1 ARDL Modelling Approach to Cointegration Analy-

sis

In time series analysis the explanatory variable may in‡uence the dependent

variable with a time lag. This often necessiates the inclusion of lags of the

explanatory variable in the regression. Furthermore, the dependent variable

may be correlated with lags of itself, suggesting that lags of the dependent

variable should be included in the regression as well. These considerations

17
motivate the commonly used ARDL(p; q) model de…ned as follow:8

yt = 1 yt 1 + + p yt p + 0 xt + 1 xt 1 + + q xt q + ut ; (10)

2
where ut iid (0; ).

In the case where the variables of interest are trend stationary, the general

practice has been to de-trend the series and to model the de-trended series as

stationary distributed lag or autoregressive distributed lag (ARDL) models.

Estimation and inference concerning the properties of the model are then

carried out using standard asymptotic normal theory.

However, the analysis becomes more complicated when the variables are

di¤erence-stationary, or I(1).The recent literature on cointegration is con-

cerned with the analysis of the long run relations between I(1) variables,

and its basic premise is, at least implicitly, that in the presence of I(1) vari-

ables the traditional ARDL approach seems no longer applicable.9

Pesaran and Shin (1999) recently re-examined the traditional ARDL ap-

8
For convenience we do not include the deterministic regressors such as constant and
linear time trend.
9
Consequently, a large number of alternative estimation and hypothesis testing proce-
dures have been speci…cally developed for the analysis of I(1) variables. See Phillips and
Hansen (1990, Review of Economic Studies) and Phillips and Loretan (1991, Review of
Economic Studies).

18
proach for an analysis of a long run relationship when the underlying variables

are I(1), and …nd:

1. The ARDL-based estimators of the long-run coe¢ cients are also con-

sistent, and have an asymptotic normal distribution.

2. Valid inferences via the Wald on the cointegrating parameters can be

2
made using the standard asymptotic distribution.

We illustrate this approach using an ARDL(1,1) regression with an I(1)

regressor,

yt = yt 1 + 0 xt + 1 xt 1 + ut ; t = 1; :::; T; (11)

where , 0 and 1 are unknown parameters, and xt is an I(1) process gen-

erated by

xt = xt 1 + et ; (12)

We make the following assumptions :

2
A1. ut iid(0; u ).

A2. et is a general linear stationary process.

A3. ut and et are uncorrelated for all lags such that xt is strictly exogenous

with respect to ut .

19
A4. (Stability Condition) j j < 1, so that the model is dynamically stable.

Assumption A4 is similar to the stationarity condition for an AR(1)

process and implies that there exists a stable long-run relationship between

yt and xt .10

Suppose that a long-run equilibrium occurs at yt = y , xt = x and ut = 0

for all t. Then, (11) becomes

y = y + 0x + 1x ;

and therefore,

y = x;

0+ 1
where = 1
is called the long-run multiplier.

We derive an ECM directly from (11). First, transform (11) to

yt = (1 ) yt 1 +( 0 + 1 ) xt 1 + 0 xt + ut (13)

= yt 1 + xt 1 + xt + ut ;

where = (1 ), = 0 + 1 and = 0. Noting that < 0 under

10
If = 1, then there would be no long-run relationship.

20
Assumption A4, we further obtain

yt = (yt 1 xt 1 ) + xt + ut ; (14)

which is an ECM. can also be regarded as a long-run cointegrating parame-

ter. This shows that the ARDL speci…cation is indeed useful to characterize

both the long-run and short-run behavior. Notice that (11), (13) and (14)

are the same but represented di¤erently. We may call (13) and (14) as un-

restricted ECM and (restricted) ECM, respectively. From (13) the long-run

parameter is also obtained by

= : (15)

Theorem 3 (ARDL) Consider the unrestricted error correction model, (13).

Under assumptions A1 - A4,

(a) The ARDL-based estimators of the long-run parameter, given by ^ =

^=^ , converge to = = as T ! 1 and T ^ has the limiting

normal distribution.

(b) The t-statistic testing = 0 converges to the standard normal distribu-

tion.

21
 We also …nd that the OLS estimators of all of the short-run parameters,
p
, , are T -consistent and have the normal distribution.

In a more general case where ut and et in (13) are serially correlated, the

ARDL speci…cation needs to be augmented with adequate number of lagged

changes in the dependent and the independent variables before estimation

and inference is carried out: that is, we now consider the ARDL(p; q) or

unrestricted ECM model,

p 1 q 1
X X
yt = yt 1 + xt 1 + j yt j + j xt j + ut : (16)
j=1 j=0

To estimate an ECM based on an ARDL approach, Pesaran and Shin

(1999) suggested the following three-step procedure:

1. Use model selection criteria such as AIC and SBC, and choose the

orders of the ARDL model.

2. Run an ARDL(p; q) or unrestricted ECM regression (16),then save the

^
error correction term by ^t = yt ^ xt , where ^ =
^
and ^ and ^ are

the OLS estimators obtained from (16).

22
 3. Run an ECM regression of y on ^t , x, p 1 lagged y’s and q 1

lagged y’s,

p 1 q 1
X X
yt = ^t 1+ j yt j + j xt j + ut :
j=1 j=0

2.2 Vector Autoregressive Model

2 3 2 3
6 xt 7 6 ut 7
1. Given yt = 6
4
7 and et = 6
5 4
7
5 iidN (0; )
zt vt
2 3 2 32 3 2 3
6 xt 7 6'xx 'xz 7 6 xt 1 7 6 ut 7
yt = 'yt + e =) 6 7=6 76 7+6 7
1 t 4 5 4 54 5 4 5
zt 'zx 'zz zt 1 vt
In general, A(L)yt = B(L)et where

2 p
A(L) = 1 1L 2L ::::: pL

B(L) = 1 + 1L + 2 L2 + ::::: + pL
p
2 3
6'j;xx 'j;xz 7
j =6
4
7
5
'j;zx 'j;zz
Given an AR(1) representation,
2 32 3
6axx (L) axz (L)7 6 xt 7
A(L)yt = et =) yt = A(L) 1 et =) 6
4
76
54
7 =
5
azx (L) azz (L) zt

23
 2 3
6 ut 7
6 7
4 5
vt
2 3 2 32 3
6 xt 7 6 axx (L) axz (L)7 6 ut 7
=) 6
4
7 = [axx (L)azz (L) axz (L)azx (L)]
5
1 6
4
76
54
7
5
zt azx (L) azz (L) vt
Similarly a process like an ARM A(2; 1) such as:

y2t = '1 y3t 1 +

2 '2 yt 2 + u3
t+2 1 ut 13can2be represented
3 as:

6 yt 7 6'1 '2 1 7 6 yt 1 7 6 1 7
6 7 6 76 7 6 7
6 7 6 76 7 6 7
6 y 7 = 6 1 0 076 y 7 6 7
6 t 1 7 6 7 6 t 2 7 + 6 0 7 wt =) xt = Axt 1 + Cwt
6 7 6 76 7 6 7
4 5 4 54 5 4 5
ut 0 0 0 ut 1 1
It is sometimes convenient to de…ne the C matrix in such a way so the

variance-covariance
2 3 matrix of the shocks is the identity matrix. Then

6 e 7
6 7
6 7
C=6 0 7
6 =
7 and E(wt wt ) = I
6 7
4 5
e

2.2.1 Forecast

We start with a vector AR(1) representation which can have a M A(1) rep-

resentation as:
P
1
yt = Aj Cwt j
j=0

Et (yt+k ) = Ak yt

24
yt+1 Et (yt+1 ) = Cwt+1

Vt (yt+1 ) = CC =

yt+2 Et (yt+2 ) = Cwt+2 + ACwt+1

Vt (yt+2 ) = CC = + ACC = A=
kP1
Vt (yt+j ) = Aj CC = Aj=
j=0

Et (yt+k ) = AEt (yt+k 1 )

= =
Vt (yt+k ) = CC 2 + AV
3t (yt+k 1 )A
6 xt 7
Suppose, yt = 6
4
7 , then
5
zt

xt = 'xx1 xt 1 + 'xx2 xt 1 + ::: + 'xz1 zt 1 + 'xz2 xt 2 + ::: + ext (17)

zt = 'zx1 xt 1 + 'zx2 xt 1 + ::: + 'zz1 zt 1 + 'zz2 xt 2 + ::: + ezt

AR(1)
2 mapping
3 2 of the above equation gives:
32 3 2 3
6 xt 7 6 'xx1 'xz1 'xx2 'xz2 :: 7 6 xt 1 7 6 1 0 7
6 7 6 76 7 6 7
6 7 6 76 7 6 7
6 zt 7 6 :: 7 6 7 6 0 1 7 2 3
6 7 6 'zx1 'zz1 'zx2 'zz2 76 zt 1 7 6 7
6 7 6 76 7 6 7 e
6 7 6 76 7 6 7 6 xt 7
6xt 1 7 6 :: 7 6 7+6 0 0 7 6 7
6 7=6 1 0 0 : 76 xt 2 7 6 74 5
6 7 6 76 7 6 7 e
6 7 6 76 7 6 7 zt
6 zt 1 7 6 :: 7 6 7 6 7
6 7 6 0 1 0 : 76 zt 2 7 6 0 0 7
6 7 6 76 7 6 7
4 5 4 54 5 4 5
:: :: :: :: :: :: :: :: ::
=) yt = Ayt 1 + et

25
 We can also start with:

y2t = yt 1
13 +
2 2 yt 2 + :: + et3 2 3 2 3
6 yt 7 6 1 2 :: :: 7 6 yt 7 6 I 7
1
6 7 6 76 7 6 7
6 7 6 76 7 6 7
6 y 7 6 I 0 :: :: 7 6 yt 2 7 6 7
6 t 1 7 6 76 7 6 0 7
6 7=6 76 7 + 6 7 [et ]
6 7 6 76 7 6 7
6 y 7 6 0 I :: :: 7 6 yt 3 7 6 7
6 t 2 7 6 76 7 6 0 7
6 7 6 76 7 6 7
4 5 4 54 5 4 5
:: :: :: :: :: :: ::
Given the above AR(1) representation, we can forecats both x and z.

We can also add a small change in the above formulation by choosing the C

matrix such that the shocks are orthogonal to each other. This implies that

E(ee= ) = I.

2.2.2 Impulse Response

The impulse response is that path y follows if it is a¤ected by a single unit

shock in et , i.e., et j = 0; et = 1; et+j = 0. It is interesting as it allows us to

think about ‘cause’and ‘e¤ects’. For example, we may want to compute the

response of stock price to a shock in interest rate in Stock-Price - Interest

Rate VAR model and interpret the result as the e¤ect of interest rate on

stock price.
P
1
AR(1) Model: yt = 'yt 1 + et =) yt = 'j et j
j=0

et : 0 0 1 0 0 0 0

26
 yt : 0 0 1 ' '2 '3 '4
P
1
M A(1) Model: yt = j ut j
j=0

ut : 0 0 1 0 0 0 0

yt : 0 0 1 1 2 3 4

Given the above formulation, the vector process works in the same way.

If an vector M A(1) Model has the following representation:

yt = (L)ut where ut [uxt ; uzt ]= and (L) = 0 + 1L + ::::, then

f 0; 1 ; :::g de…ne
2 the impulse-response
3 functions.

6 xx (L) xz (L) 7
(L) = 6
4
7
5
zx (L) zz (L)

Therefore, xx (L) gives the response of xt+k to a unit shock in uxt , and

xz (L) gives the response of xt+k to a unit shock in uzt . In practice, however.

we map this to a vector AR(1) representation as it is more convenient to

calculate the imp[ulse response.

yt = Ayt 1 + Cet and therefore the impulse-respponse function is: C, AC,

A2 C,....Ak C,.....

2.2.3 Orthogonalization

27
 =
A(L)yt = et ; A(0) = I; E(et et ) = (18)

= 1
yt = B(L)et ; B(0) = I; E(et et ) = where B(L) = A(L)

Suppose we want to examine the response of yt to unit movements in ext

and ezt + 0:3ext . We call the new

2 shock 3
as 1t = ext and 2t = ezt + 0:3ext .

6 1 0 7
We de…ne t = Qet where Q = 6
4
7. Then we get the following:
5
0:3 1

yt = B(L)Q 1 Qet = C(L) t where C(L) = B(L)Q 1 ; C(L) gives the

response of yt to the new shocks t. Note that yt = B(L)et and yt =

C(L) t are observationally equivalent, since they produce the same series yt .

We, then, need to decide which linear combinations is the most interesting.

In order to do so, we state a set of assumptions, called orthogonalization

assumptions, which uniquely determine the linear combinations of shocks

that we …nd most interesting.

Orthogonal Shocks: The …rst assumption is that the shocks should be

orthogonal. Let us start with the example: ‘e¤ect of interest rate on stock

prices’. If the interest rate shock is correlated with the stock price shock,

then we we do not know if the response that we observe is the response

of stock price to interest rate, or to an announcement shock (in terms of

28
dividends) that happen to come at the same time as the interest rate shock

(it may be because the Reserve Bank of India sees the announcement shock

and accomodates it). Additionally, it is convenient to rescale the shocks

so that they have a unit variance. Thus we want to choose Q such that
= =
E( t t) = I. To do this we need a Q such that Q 1 Q 1=
= , E( t t) =
=
E(Qet et Q= ) = Q Q= = I. However, there may be many di¤erent Q’s that

can act as the square root matrices for . Given one such Q, we can have

another de…ned as Q = M Q where M is the orthogonal matrix M M = = I,

=
QQ = M QQ= M = = M M = = I. Since C(L) = B(L)Q 1 , we specify a

desired property of C(L), which help us to pin down Q.

Sims’Orthogonalization: Sims (1980) suggested that we specify prop-

erties of C(0) that gives the instantaneous response of each variable to each

orthogonalized
2 3 2 shock . Sims
3 2suggested
3 a lower triangular C(0):
6 xt 7 6 C0xx 0 7 6 xt 7
6 7=6 76 7 + C1 t 1 + :::
4 5 4 54 5
zt C0zx C0zz zt
The above spe…cation implies that 2t does not a¤ect the …rst variable,

xt , contemporaneously. Both shocks can a¤ect zt contemporaneously. In the

orginal VAR-model, A(0) = I implying that contemporaneous values of each

variable do not appear in the other varoable’s equation. A lower triangular

29
matrix C(0) implies that contemporaneous xt appears in the zt equation, but

zt does not appear in the xt equation.

Let us denote the orthogonalized autoregressive representation: D(L)yt =

t where D(L) = C(L) 1 . As C(0) is lower traingular D(0) is also lower

traingular.
2 32 3
6 D0xx 0 7 6 xt 7
6 76 7 + D1 yt + ::: =
4 54 5 1 t
D0zx D0zz zt
Assuming an AR(1) process we get:

D0xx xt = D1xx xt 1 D1xz zt 1 + xt (19)

D0zz zt = D0zx xt D1zx xt 1 D1zz zt 1 + zt

Note the above is numerically equivalent to estimate the system by OLS

with contemporaneous xt in the zt equation, but not the vice versa, and then

scaling each equation so that the eroor variance is one. Recall that OLS

estimates produce residuals that are uncorrelated with the RHS variables by

construction. Hence if we ran OLS as:

30
 xt = a1xx xt 1 + :: + a1xz zt 1 + xt (20)

zt = a0zx xt + a1zx xt 1 + :: + a1zz zt 1 + zt

The …rst OLS residual is xt = xt E(xt j xt 1 ; ::; zt 1 ) and hence a

linear combination of fxt ; xt 1 ; ::; zt 1 ; :::g. OLS residuals are orthogonal

to the RHS variables, so zt is orthogonal to any linear combination of

fxt ; xt 1 ; ::; zt 1 ; :::g by construction. Hence, xt and zt are uncorrelated

with each other. The term a0zx captures all of the contemporaneous correla-

tion of news in xt and that in xt .11

2.2.4 Variance Decomposition

The purpose is to compute an accounting of the forecast error variance in

the orthogonalized system: what percent of the k-step ahead forecast error

variance is due to which variable.

=
yt = C(L) t , E( t t) =I

The 1-step ahead forecast error variance is:

11
C(L) = B(L)Q 1
C(0) = B(0)Q 1
This implies that the Choleski decomposition produces a lower triangular Q.

31
 2 32 3
6 cxx;0 cxz;0 7 6 x;t+1 7
et = yt+1 Et (yt+1 ) = C(0) t+1 =6
4
76
54
7
5
czx;0 czz;0 z;t+1
where

C(L) = C(0) + C(1)L + C(2)L2 + ::: and the elements of C(L) as cxx;0 +

cxx;1 L + cxx;2 L2 + :::. Since, the s are uncorrelated and have unit variance:

Vt (xt+1 ) = c2xx;0 2
( x ) + c2xy;0 2
( z ) = c2xx;0 + c2xy;0 and similarly for z.

Thus c2xx;0 gives the amount of 1-step ahead forecast error variance of x due

to the x shock and c2xx;0 due to the z shock. One in general reports them
c2xx;0 c2
as fractions: , zz;0 .
c2xx;0 +c2xx;0 c2zx;0 +c2xx;0
Hence, we can write:
2 3 2 3
6 1 0 7 6 0 0 7
Vt (xt+1 ) = C0 C0 , I1 = 6 7 and I2 = 6 7
=
4 5 4 5
0 0 0 1
Thus, the part of the 1-step ahead forecast error variance due to the …rst
= =
(x) shock is C0 I1 C0 and the part due to the second (z) shock is C0 I2 C0 . We
= = =
also get: C0 C0 = C0 I1 C0 +C0 I2 C0 . If we think Im as a new covariance

matrix where all shocks but the m-th are turned o¤ then the total variance

of forecast errors must be equal to the part due to the m-th shock, given by
=
C0 Im C0 . Therefore,

yt+k Et (yt+k ) = C(0) t+k + C(1) t+k 1 + ::: + C(k 1) t+1

kP1
= = = =
Vt (yt+k ) = C0 C0 + C1 C1 + :::: + Ck 1 Ck 1 =) Vk;m = Cj Im Cj is the
j 0

32
variance of k-step ahead forecast due to the m-th shock, and the variance is
P
the sum of these components, Vt (yt+k ) = Vk;m . On the other hand, note
m

t = Qet , therefore we can get:

y2t = yt 1
13 +
2 2 yt 2 + :: + et3 2 3 2 3
1
6 yt 7 6 1 2 :: :: 7 6 yt 7 6 Q
1 7
6 7 6 76 7 6 7
6 7 6 76 7 6 7
6 y 7 6 I 0 :: :: 7 6 yt 2 7 6 7
6 t 1 7 6 76 7 6 0 7
6 7=6 76 7+6 7 [ t]
6 7 6 76 7 6 7
6 y 7 6 0 I :: :: 7 6 yt 3 7 6 7
6 t 2 7 6 76 7 6 0 7
6 7 6 76 7 6 7
4 5 4 54 5 4 5
:: :: :: :: :: :: ::
=
=) yt = Ayt 1 + C t , E( t t) =I

The impulse-response function is C, AC, A2 C,..... and can be found

recursively from:

IR0 = C, IRj = AIRj 1

1
If Q is lower diagonal, then only the 1st shock a¤ects the 1st variable:
kP1 kP1
Vt (yt+k ) = Aj CC = Aj= =) Vk;m = Aj CIm C = Aj=
j 0 j 0

Causality Y = X + e does not imply “x causes" y. The most natural

de…nition of cause: “cause should preced e¤ects. However, this may not be

true in Time-Series.

De…nition 1 : Granger Causality

33
 xt Granger causes yt if xt helps to forecast yt , given past yt . Consider

the following:

yt = a(L)yt 1 + b(L)xt 1 + ut

xt = c(L)yt 1 + d(L)xt 1 + vt

Our de…nition inplies that xt does not Granger cause yt if b(L) = 0, i.e.

the above system of equation s can be represented as:

yt = a(L)yt 1 + ut

xt = c(L)yt 1 + d(L)xt 1 + vt

Alternatively,
2 3 2 32 3 2 3
6 yt 7 6 a(L) b(L) 7 6 yt 1 7 6 ut 7
6 7=6 76 7+6 7
4 5 4 54 5 4 5
xt c(L) d(L) xt 1 vt
2 32 3 2 3
6 I La(L) Lb(L) 7 6 yt 7 6 ut 7
=) 6
4
76
54
7=6
5 4
7
5
Lc(L) I Ld(L) xt vt

34
 2 32 3 2 3
6 a (L) b (L) 7 6 yt 7 6 ut 7
=) 6
4
76
54
7=6
5 4
7
5
c (L) d (L) xt vt
Thus x does not Granger cause y i¤ b (L) = 0.

We can also have a M A representation to explain the concept of Granger

causality.
2 3 2 32 3
6 yt 7 6 d (L) b (L) 7 6 ut 7
6 7= 1 6 76 7
4 5 a (L)d (L) b (L)c (L) 4 54 5
xt c (L) a (L) vt
This implies that x does not Granger cause y if the Wold moving average

matrix lag polynomial is lower triangular. If x does not Granger cause y,

then y is a function of its own shocks only, and does not respond to x shocks.

x on the other hand, is a function of both y shocks and x shocks. This also

implies:

x does not Granger cause y if the y’s bivariate Wold representation is the

same as the univariate Wold representation. Therefore, x does not Granger

cause y if the projection of y on past y and x is the same as the projection

of y on past y alone.

Suppose,

35
 yt = e(L) t where t = yt P (yt jyt 1 ; yt 2 ; :::)

xt = f (L) t where t = xt P (xt jxt 1 ; xt 2 ; :::)

In the above formulation, xt does not Granger cause yt if E( t t+j ) =0

for all j > 0. This implies that the univariate innovations of xt are un-

correlated with the univariate innovations in yt . Our original idea that xt

Granger causes yt if its movements preced those of yt was true i¤ it applies

to innovations, not the level of the series. If xt does not Granger cause yt

then

xt = c(L)ut + d(L)vt

However, the univariate representation of xt is xt = f (L) t . This implies:

xt = c(L)ut + d(L)vt = f (L) t

36
 Thus, t is a linear combination of current and past ut and vt . Since ut

is the bivariate error:

E(ut jyt 1 ; yt 2 ; :::; xt 1 ; xt 2 ; :::) = E(ut jut 1 ; ut 2 ; :::; vt 1 ; vt 2 ; ::) = 0

This implies that ut is uncorrelated with lagged ut and vt and therefore

lagged t. If E( t t+j ) = 0, then past does not help to forecast , and

thus past does not help forecasting y given past y. Since xt = f (L) t , this

implies that past observations on x do not help in forecasting y given past y.

Facts

x does not Granger cause y if:

Past x do not help in forecasting y given past y. In a regression of y

on past y and past x, the coe¢ ceints of the past x are all zero.

The autoregressive representation is lower triangular.

Univariate innovations in x are not correlated with subsequent innova-

tions in y.

37