2.

2 Vector Autoregression (VAR)

Provided that in the previous example the series are not cointegrated an appropriate modeling approach is VAR for the dierences.
Suppose we have m time series yit, i = 1, . . . , m,
and t = 1, . . . , T (common length of the time
series). Then a vector autoregression model
is dened as

y1t
y2t
..
.
ymt

1
2
..
.
m

1t
2t
..
.
mt

(p)
11
(p)
21
..
.
(p)
m1

(1)
11
(1)
21
..
.
(1)
m1
(p)
12
(p)
22
..
.
(1)
m2

(1)
12
(1)
22
..
.
(1)
m2

...

...

(p)
1m
(p)
2m
..
.
(p)
mm

(1)
1m
(1)
2m
..
.
(1)
mm

y1,tp
y2,tp
..
.
ym,tp

(12)

10

y1,t1
y2,t1
..
.
ym,t1

In matrix notations
(13)yt = + 1yt1 + + pytp + t,
which can be further simplied by adopting
the matric form of a lag polynomial
(14)

(L) = I 1L . . . pLp.

Thus nally we get

(15)

(L)yt = t.

Note that in the above model each yit depends not only its own history but also on
other series history (cross dependencies). This
gives us several additional tools for analyzing
causal as well as feedback eects as we shall
see after a while.

11

A basic assumption in the above model is

that the residual vector follow a multivariate
white noise, i.e.
(16)

E(t) = 0


E(ts) =
0

if t = s
if t =
 s

The coecient matrices must satisfy certain

constraints in order that the VAR-model is
stationary. They are just analogies with the
univariate case, but in matrix terms. It is
required that roots of
(17) |I 1z 2z 2 pz p| = 0
lie outside the unit circle. Estimation can be
carried out by single equation least squares.

12

Example 2.3: Let us estimate a VAR(1) model for the

equity-bond data. First, however, test whether the
series are cointegrated. As is seen below, there is no
empirical evidence of cointegration (EViews results)
Sample(adjusted): 1965:06 1995:12
Included observations: 367 after adjusting end points
Trend assumption: Linear deterministic trend
Series: LFTA LDIV LR20 LTBILL
Lags interval (in first differences): 1 to 4
Unrestricted Cointegration Rank Test
===================================================================
Hypothesized
Trace
5 Percent
1 Percent
No. of CE(s) Eigenvalue Statistic Critical Value Critical Value
===================================================================
None
0.047131
46.02621
47.21
54.46
At most 1
0.042280
28.30808
29.68
35.65
At most 2
0.032521
12.45356
15.41
20.04
At most 3
0.000872
0.320012
3.76
6.65
===================================================================
*(**) denotes rejection of the hypothesis at the 5%(1%) level
Trace test indicates no cointegration at both 5% and 1% levels

13

VAR(2) Estimates:
Sample(adjusted): 1965:04 1995:12
Included observations: 369 after adjusting end points
Standard errors in ( ) & t-statistics in [ ]
===========================================================
DFTA
DDIV
DR20
DTBILL
===========================================================
DFTA(-1)
0.102018 -0.005389 -0.140021 -0.085696
(0.05407) (0.01280) (0.02838) (0.05338)
[1.88670][-0.42107] [-4.93432] [-1.60541]
DFTA(-2)
-0.170209 0.012231
0.014714
0.057226
(0.05564) (0.01317) (0.02920) (0.05493)
[-3.05895] [0.92869] [0.50389] [1.04180]
DDIV(-1)
-0.113741 0.035924
0.197934
0.280619
(0.22212) (0.05257) (0.11657) (0.21927)
[-0.51208] [0.68333] [1.69804] [1.27978]
DDIV(-2)
0.065178 0.103395
0.057329
0.165089
(0.22282) (0.05274) (0.11693) (0.21996)
[0.29252] [1.96055] [0.49026] [0.75053]
DR20(-1)
-0.359070 -0.003130
0.282760
0.373164
(0.11469) (0.02714) (0.06019) (0.11322)
[-3.13084] [-0.11530] [4.69797] [3.29596]
DR20(-2)
0.051323 -0.012058 -0.131182 -0.071333
(0.11295) (0.02673) (0.05928) (0.11151)
[0.45437][-0.45102] [-2.21300] [-0.63972]
DTBILL(-1)
0.068239 0.005752 -0.033665
0.232456
(0.06014) (0.01423) (0.03156) (0.05937)
[1.13472] [0.40412] [-1.06672] [3.91561]
DTBILL(-2)
-0.050220 0.023590
0.034734 -0.015863
(0.05902) (0.01397) (0.03098) (0.05827)
[-0.85082] [1.68858] [1.12132] [-0.27224]
C
0.892389 0.587148 -0.033749 -0.317976
(0.38128) (0.09024) (0.20010) (0.37640)
[2.34049] [6.50626] [-0.16867] [-0.84479]
===========================================================

14

Continues . . .
===========================================================
DFTA
DDIV
DR20
DTBILL
===========================================================
R-squared
0.057426
0.028885
0.156741 0.153126
Adj. R-squared
0.036480
0.007305
0.138002 0.134306
Sum sq. resids
13032.44
730.0689
3589.278 12700.62
S.E. equation
6.016746
1.424068
3.157565 5.939655
F-statistic
2.741619
1.338486
8.364390 8.136583
Log likelihood -1181.220 -649.4805
-943.3092 -1176.462
Akaike AIC
6.451058
3.569000
5.161567 6.425267
Schwarz SC
6.546443
3.664385
5.256953 6.520652
Mean dependent
0.788687
0.688433
0.052983 -0.013968
S.D. dependent
6.129588
1.429298
3.400942 6.383798
===========================================================
Determinant Residual Covariance
18711.41
Log Likelihood (d.f. adjusted)
-3909.259
Akaike Information Criteria
21.38352
Schwarz Criteria
21.76506
===========================================================

As is seen the number of estimated parameters grows rapidly very large.

15

Dening the order of a VAR-model

In the rst step it is assumed that all the series in the VAR model have equal lag lengths.
To determine the number of lags that should
be included, criterion functions can be utilized the same manner as in the univariate
case.

16

The underlying assumption is that the residuals follow a multivariate normal distribution,
i.e.
(18)

 Nm(0, ).

Akaikes criterion function (AIC) and Schwarzs

criterion (BIC) have gained popularity in determination of the number of lags in VAR.
In the original forms AIC and BIC are dened
as
(19)

AIC = 2 log L + 2s

(20)

BIC = 2 log L + s log T

where L stands for the Likelihood function,

and s denotes the number of estimated parameters.

17

The best tting model is the one that minimizes the criterion function.
For example in a VAR(j) model with m equations there are s = m(1 + jm) + m(m + 1)/2
estimated parameters.

18

Under the normality assumption BIC can be

simplied to
(21)

jm2 log T

,
BIC(j) = log ,j +

and AIC to
(22)

2jm2

AIC(j) = log ,j +
,

j = 0, . . . , p, where
(23)

T
1

1

j
jE

,j =

t,j
t,j = E
T t=j+1
T

with
t,j the OLS residual vector of the VAR(j)
model (i.e. VAR model estimated with j
lags), and
(24)

j =
E
j+1,j ,
j+2,j , ,
T,j

an m (T j) matrix.

19

The likelihood ratio (LR) test can be also

used in determining the order of a VAR. The
test is generally of the form
(25)

p|),
k | log |
LR = T (log |

k denotes the maximum likelihood

where
estimate of the residual covariance matrix of
p the estimate of VAR(p) (p >
VAR(k) and
k) residual covariance matrix. If VAR(k) (the
shorter model) is the true one, i.e., the null
hypothesis
(26)

H0 : k+1 = = p = 0

is true, then
(27)

LR 2
df ,

where the degrees of freedom, df , equals the

dierence of in the number of estimated parameters between the two models.

20

In an m variate VAR(k)-model each series has

p k lags less than those in VAR(p). Thus
the dierence in each equation is m(p k),
so that in total df = m2(p k).
Note that often, when T is small, a modied
LR
k | log |
p|)
LR = (T mg)(log |
is used to correct possible small sample bias,
where g = p k.
Yet another method is to use sequential LR
k+1|)
k | log |
LR(k) = (T m)(log |
which tests the null hypothesis
H0 : k+1 = 0
given that j = 0 for j = k + 2, k + 3, . . . , p.
EViews uses this method in View -> Lag Structure
-> Lag Length Criteria. . ..
21

Example 2.4: Let p = 12 then in the equity-bond

data dierent VAR models yield the following results.
Below are EViews results.
VAR Lag Order Selection Criteria
Endogenous variables: DFTA DDIV DR20 DTBILL
Exogenous variables: C
Sample: 1965:01 1995:12
Included observations: 359
=========================================================
Lag
LogL
LR
FPE
AIC
SC
HQ
--------------------------------------------------------0 -3860.59
NA
26324.1
21.530
21.573
21.547
1 -3810.15 99.473 21728.6* 21.338* 21.554* 21.424*
2 -3796.62 26.385 22030.2
21.352
21.741
21.506
3 -3786.22 20.052 22729.9
21.383
21.945
21.606
4 -3783.57 5.0395 24489.4
21.467
22.193
21.750
5 -3775.66 14.887 25625.4
21.502
22.411
21.864
6 -3762.32 24.831 26016.8
21.517
22.598
21.947
7 -3753.94 15.400 27159.4
21.560
22.814
22.059
8 -3739.07 27.018* 27348.2
21.566
22.994
22.134
9 -3731.30 13.933 28656.4
21.612
23.213
22.248
10 -3722.40 15.774 29843.7
21.651
23.425
22.357
11 -3715.54 12.004 31443.1
21.702
23.649
22.476
12 -3707.28 14.257 32880.6
21.745
23.865
22.588
=========================================================
* indicates lag order selected by the criterion
LR: sequential modified LR test statistic
(each test at 5% level)
FPE: Final prediction error
AIC: Akaike information criterion
SC: Schwarz information criterion
HQ: Hannan-Quinn information criterion

Criterion function minima are all at VAR(1) (SC or

BIC just borderline). LR-tests suggest VAR(8). Let
us look next at the residual autocorrelations.
22

In order to investigate whether the VAR residuals are White Noise, the hypothesis to be
tested is
(28)

H0 : 1 = = h = 0

where k = (ij (k)) is the autocorrelation

matrix (see later in the notes) of the residual
series with ij (k) the cross autocorrelation
of order k of the residuals series i and j. A
general purpose (portmanteau) test is the Qstatistic
(29)

Qh = T

k
1
k
1),
tr(
0
0

k=1

k = (
ij (k)) are the estimated (residwhere
0 the contempoual) autocorrelations, and
raneous correlations of the residuals.

See

e.g. L
utkepohl, Helmut (1993). Introduction to
Multiple Time Series, 2nd Ed., Ch. 4.4
23

Alternatively (especially in small samples) a

modied statistic is used

(30)

Qh = T 2

1
1
k
(T k)1 tr(
0 k 0 ).

k=1

24

The asymptotic distribution is 2 with df =

p2(h k). Note that in computer printouts
h is running from 1, 2, . . . h with h specied
by the user.
VAR(1) Residual Portmanteau Tests for Autocorrelations
H0: no residual autocorrelations up to lag h
Sample: 1966:02 1995:12
Included observations: 359
================================================
Lags
Q-Stat
Prob.
Adj Q-Stat Prob.
df
-----------------------------------------------1
1.847020
NA*
1.852179
NA*
NA*
2
27.66930
0.0346
27.81912
0.0332
16
3
44.05285
0.0761
44.34073
0.0721
32
4
53.46222
0.2725
53.85613
0.2603
48
5
72.35623
0.2215
73.01700
0.2059
64
6
96.87555
0.0964
97.95308
0.0843
80
7
110.2442
0.1518
111.5876
0.1320
96
8
137.0931
0.0538
139.0485
0.0424 112
9
152.9130
0.0659
155.2751
0.0507 128
10
168.4887
0.0797
171.2972
0.0599 144
11
179.3347
0.1407
182.4860
0.1076 160
12
189.0256
0.2379
192.5120
0.1869 176
================================================
*The test is valid only for lags larger than the
VAR lag order. df is degrees of freedom for
(approximate) chi-square distribution

There is still left some autocorrelation into

the VAR(1) residuals. Let us next check the
residuals of the VAR(2) model
25

VAR Residual Portmanteau Tests for Autocorrelations

H0: no residual autocorrelations up to lag h
Sample: 1965:01 1995:12
Included observations: 369
==============================================
Lags
Q-Stat
Prob.
Adj Q-Stat Prob. df
---------------------------------------------1
0.438464
NA*
0.439655
NA*
NA*
2
1.623778
NA*
1.631428
NA*
NA*
3
17.13353
0.3770 17.26832 0.3684 16
4
27.07272
0.7143 27.31642 0.7027 32
5
44.01332
0.6369 44.48973 0.6175 48
6
66.24485
0.3994 67.08872 0.3717 64
7
80.51861
0.4627 81.63849 0.4281 80
8
104.3903
0.2622 106.0392 0.2271 96
9
121.8202
0.2476 123.9049 0.2081 112
10
136.8909
0.2794 139.3953 0.2316 128
11
147.3028
0.4081 150.1271 0.3463 144
12
157.4354
0.5425 160.6003 0.4718 160
==============================================
*The test is valid only for lags larger than
the VAR lag order.
df is degrees of freedom for (approximate)
chi-square distribution

Now the residuals pass the white noise test.

On the basis of these residual analyses we can
select VAR(2) as the specication for further
analysis. Mills (1999) nds VAR(6) as the
most appropriate one. Note that there ordinary dierences (opposed to log-dierences)
are analyzed. Here, however, log transformations are preferred.
26

Vector ARMA (VARMA)

Similarly as is done in the univariate case
one can extend the VAR model to the vector
ARMA model
(31)yt = +

i=1

iyti + t

j tj

j=1

or
(32)

(L)yt = + (L)t,

where yt, , and t are m 1 vectors, and

is and j s are m m matrices, and
(33)

(L) = I 1L . . . pLp
(L) = I 1L . . . q Lq .

Provided that (L) is invertible, we always

can write the VARMA(p, q)-model as a VAR()
model with (L) = 1(L)(L). The presence of a vector MA component, however,
complicates the analysis somewhat.

27

Autocorrelation and Autocovariance Matrices

The kth cross autocorrelation of the ith and
jth time series, yit and yjt is dened as
(34)

ij (k) = E(yitk i)(yjt j ).

Although for the usual autocovariance k =

k , the same is not true for the cross autocovariance, but ij (k) = ij (k). The corresponding cross autocorrelations are
(35)

i,j (k) =

ij (k)

i(0)j (0)

The kth autocorrelation matrix is then

(36)

1 (k)
1,2 (k) . . .
2,1(k) 2(k)
...
k =
.
...
..
m,1 (k) m,2 (k) . . .

1,m (k)
2,m (k)
.
...

m (k)

28

Remark 2.4: k = k .

The diagonal elements, j (k), j = 1, . . . , m of

are the usual autocorrelations.

29

Example 2.5: Cross autocorrelations of the equitybond data.

1 =

Divt1
Ftat1
R20t1
Tblt1

Divt
0.0483
0.0160
0.0056
0.0536

Ftat
0.0099
0.1225
0.1403
0.0266

R20t
0.0566
0.2968
0.2889
0.1056

Tblt
0.0779
0.1620
0.3113
0.3275

Note: Correlations with absolute value exceeding 2 std err =

2 / 371 0.1038 are statistically signicant at the 5% level

(indicated by ).

30

Example 2.6: Consider individual security returns and

portfolio returns. For example French and Roll have
found that daily returns of individual securities are
slightly negatively correlated. The tables below, however suggest that daily returns of portfolios tend to
be positively correlated!

French,

K. and R. Ross (1986). Stock return variances: The arrival of information and reaction of
traders. Journal of Financial Economics, 17, 526.
31

One explanation could be that the cross autocorrelations are positive, which can be partially proved as follows: Let
(37)

m
1

1
rit = rt
rpt =
m i=1
m

denote the return of an equal weighted index,

where = (1, . . . , 1) is a vector of ones, and
rt = (r1t, . . . , rmt) is the vector of returns of
the securities. Then


(38)


 rt1  rt
 1
,
cov(rpt1 , rpt) = cov
=
,
m
m
m2

where 1 is the rst order autocovariance

matrix.

32

Therefore


m cov(rpt1, rpt ) = 1 = 1 tr(1 ) + tr(1 ),

2

(39)

where tr() is the trace operator which sums

the diagonal elements of a square matrix.
Consequently, because the right hand side
tends to be positive and tr(1) tends to be
negative, 1 tr(1), which contains only
cross autocovariances, must be positive, and
larger than the absolute value of tr(1), the
autocovariances of individual stock returns.

33