Vector error correction model, VECM Cointegrated VAR

Chapter 4

Financial Econometrics
Michael Hauser WS12/13

1 / 58

Content

Motivation: plausible economic relations Model with I(1) variables: spurious regression, bivariate cointegration Cointegration Examples: unstable VAR(1), cointegrated VAR(1) VECM, vector error correction model Cointegrated VAR models, model structure, estimation, testing, forecasting (Johansen) Bivariate cointegration

2 / 58

Motivation

3 / 58

Paths of Dow JC and DAX: 10/2009 - 10/2010

We observe a parallel development. Remarkably this pattern can be observed for single years at least since 1998, though both are assumed to be geometric random walks. They are non stationary, the log-series are I(1). If a linear combination of I(1) series is stationary, i.e. I(0), the series are called cointegrated. If there are 2 processes xt and yt are both I(1) and yt xt = t with t trend-stationary or simply I(0), then xt and yt are called cointegrated.

4 / 58

Cointegration in economics
This concept origins in macroeconomics where series often seen as I(1) are regressed onto, like private consumption, C, and disposable income, Y d . Despite I(1), Y d and C cannot diverge too much in either direction: C > Y d or C Y d Or, according to the theory of competitive markets the prot rate of rms (prots/invested capital) (appox I(1)) should converge to the market average over time. This means that prots should be proportional to the invested capital in the long run.

5 / 58

Common stochastic trend

The idea of cointegration is that there is a common stochastic trend, an I(1) process Z , underlying two (or more) processes X and Y . E.g. Xt = 0 + 1 Zt + t Yt = 0 + 1 Zt + t t and t are stationary, I(0), with mean 0. They may be serially correlated. Though Xt and Yt are both I(1), there exists a linear combination of them which is stationary: 1 Xt 1 Yt + const I(0)

6 / 58

Models with I(1) variables

7 / 58

Spurious regression
The spurious regression problem arises if arbitrarily trending or nonstationary series are regressed on each other. In case of trending the spuriously found relationship is due to the trend (growing over time) governing both series instead to economic reasons. t-statistic and R 2 are implausibly large. In case of nonstationarity (of I(1) type) the series - even without drifts - tend to show local trends, which tend to comove along for relative long periods.

8 / 58

Spurious regression: independent I(1)s

We simulate paths of 2 RWs without drift with independently generated standard normal white noises, t , t . Xt = Xt1 + t , Yt = Yt1 + t , t = 0, 1, 2, . . .

Then we estimate by LS the model Yt = + Xt + t In the population = 0 and = 0, since Xt and Yt are independent. Replications for increasing sample sizes shows that the DW-statistics are close to 0. R 2 is too large. t is I(1), nonstationary. the estimates are inconsistent. the t -statistic diverges with rate T.
9 / 58

Spurious regression: independence

As both X and Y are independent I(1)s, the relation can be checked consistently using rst differences. Yt = Xt + t Here we nd that has the expected distribution around zero, the t -values are t-distributed, the error t is WN.

10 / 58

Bivariate cointegration
However, if we observe two I(1) processes X and Y , so that the linear combination Yt = + Xt + t is stationary, i.e. t is stationary, then Xt and Yt are cointegrated. When we estimate this model with LS, the estimator is not only consistent, but superconsistent. It converges with the rate T , instead of T . However, the t -statistic is asy normal only if t is not serially correlated.

11 / 58

Bivariate cointegration: discussion

The Johansen procedure (which allows for correction for serial correlation easily) (see below) is to be preferred to single equation procedures. If the model is extended to 3 or more variables, more than one relation with stationary errors may exist, then when estimating a multiple regression, it is not clear what we get.

12 / 58

Cointegration

13 / 58

Denition: Cointegration
Denition: Given a set of I(1) variables {x1t , . . . , xkt }. If there exists a linear combination consisting of all vars with a vector so that 1 x1t + . . . + k xkt = x t . . . trend-stationary j = 0, j = 1, . . . , k . Then the xs are cointegrated of order CI(1,1). x t is a (trend-)stationary variable. The denition is symmetric in the vars. There is no interpretation of endogenous or exogenous vars. A simultaneous relationship is described.
Denition: Trend-stationarity means that after subtracting a deterministic trend the process is I(0).
14 / 58

Denition: Cointegration (cont)

is dened only up to a scale. If x t is trend-stationary, then also c( x t ) with c = 0. Moreover, any linear combination of cointegrated variables is stationary. More generally we could consider x I(d) and x I(d b) with b > 0. Then the xs are CI(d, b). We will deal only with the standard case of CI(1,1).

15 / 58

An unstable VAR(1), an example

16 / 58

An unstable VAR(1):

x t = 1 x t1 + t

We analyze in the following the properties of [ ] [ ][ ] [ ] x1t 0.5 1. x1,t1 1t = + x2t .25 0.5 x2,t1 2t t are weakly stationary and serially uncorrelated. We know a VAR(1) is stable, if the eigenvalues of 1 are less 1 in modulus. The eigenvalues of 1 are 1,2 = 0, 1. The roots of the characteristic function |I 1 z| = 0 should be outside the unit circle for stationarity. Actually, the roots are z = (1/) with = 0. z = 1. 1 has a root on the unit circle. So process x t is not stable. Remark: 1 is singular; its rank is 1.
17 / 58

Common trend
For all 1 there exists an invertible matrix L so that L1 L1 = is (for simplicity) diagonal containing the eigenvalues of 1 . We dene new variables y t = Lx t and t = Lt . Left multiplication of the VAR(1) with L gives Lx t = L1 x t1 + Lt (Lx t ) = L1 L1 (Lx t1 ) + (Lt ) y t = y t1 + t

18 / 58

Common trend: xs are I(1)

In our case L and are [ ] 1.0 2.0 L= , 0.5 1.0 Then [ ] = [ ][ [ = ] + ]

1 0 0 0 [

y1t y2t

1 0 0 0

y1,t1 y2,t1

] 1t 2t

t = Lt : 1t and 2t are linear combinations of stationary processes. So they are stationary. So also y2t is stationary. y1t is obviously integrated of order 1, I(1).

19 / 58

Common trend y1t , xs as function of y1t

y t = Lx t with L invertible, so we can express x t in y t . Left multiplication by L1 gives L1 y t = L1 y t1 + L1 t x t = (L1 )y t1 + t L1 = . . . x1t = (1/2)y1,t1 + 1t x2t = (1/4)y1,t1 + 2t Both x1t and x2t are I(1), since y1t is I(1). y1t is called the common trend. It is the common nonstationary component in both x1t and x2t .

20 / 58

Cointegrating relation

Now we eliminate y1,t1 in the system above by multiplying the 2nd equation by 2 and adding to the rst. x1t + 2x2t = (1,t + 22,t ) This gives a stationary process, which is called the cointegrating relation. This is the only linear combination (apart from a factor) of both nonstationary processes, which is stationary.

21 / 58

A cointegrated VAR(1), an example

22 / 58

A cointegrated VAR(1)
We go back to the system and proceed directly. x t = 1 x t1 + t and subtract x t1 on both sides (cp. the Dickey-Fuller statistic). [ ] [ ][ ] [ ] x1t .5 1. x1,t1 1t = + x2t .25 .5 x2,t1 2t The coefcient matrix , = (I 1 ), in x t = x t1 + t has only rank 1. It is singular. Then can be factorized as = (2 2) = (2 1)(1 2)
23 / 58

A cointegrated VAR(1)
k the number of endogenous variables, here k = 2. m = Rank() = 1, is the number of cointegrating relations. A solution for = is [ ] ( ) ( ) ( ) ( ) .5 1. .5 1 .5 = = 1 2 .25 .5 .25 2 .25 Substituted in the model ] ) [ ] [ ] ( [ ( ) x 1t .5 x1t 1,t1 + = 1 2 2t .25 x2,t1 x2t

24 / 58

A cointegrated VAR(1)

Multiplying out [ ] ( ) [ ] ( ) x1t .5 1t = x1,t1 + 2x2,t1 + x2t .25 2t The component (x1,t1 + 2x2,t1 ) appears in both equations. As the lhs variables and the errors are stationary, this linear combination is stationary. This component is our cointegrating relation from above.

25 / 58

Vector error correction, VEC

26 / 58

VECM, vector error correction model

Given a VAR(p) of I(1) xs (ignoring consts and determ trends) x t = 1 x t1 + . . . + p x tp + t There always exists an error correction representation of the form (trick x t = x t1 + x t ) x t = x t1 +
p1 i=1

x ti + t i

where and the are functions of the s. Specically, j =

p i=j+1

i ,

j = 1, . . . , p 1

= (I 1 . . . p ) = (1) The characteristic polynomial is I 1 z . . . p z p = (z).

27 / 58

Interpretation
If = 0, then there is no cointegration. Nonstationarity of I(1) type vanishes by taking differences. If has full rank, k , then the xs cannot be I(1) but are stationary. (1 x t = x t1 + 1 t ) The interesting case is, Rank() = m, 0 < m < k , as this is the case of cointegration. We write = (k k ) = (k m)[(k m) ] where the columns of contain the m cointegrating vectors, and the columns of the m adjustment vectors. Rank() = min[ Rank(), Rank() ]
28 / 58

Long term relationship

There is an adjustment to the equilibrium x or long term relation described by the cointegrating relation. Setting x = 0 we obtain the long run relation, i.e. x = 0 This may be wirtten as x = ( x ) = 0 In the case 0 < Rank() = Rank() = m < k the number of solutions of this system of linear equations which are different from zero is m. x = 0mk
29 / 58

Long term relationship

The long run relation does not hold perfectly in (t 1). There will be some deviation, an error, x t1 = t1 = 0 The adjustment coefcients in multiplied by the errors x t1 induce adjustment. They determine x t , so that the xs move in the correct direction in order to bring the system back to equilibrium.

30 / 58

Adjustment to deviations from the long run

The long run relation is in the example above x1,t1 + 2x2,t1 = t1 t is the stationary error. The adjustment of x1,t in t to t1 , the deviation from the long run in (t 1), is x1,t = (.5)t1 and x1,t = x1,t + x1,t1

If t1 > 0, the error is positive, i.e. x1,t1 is too large c.p., then x1,t , the change in x1 , is negative. x1 decreases to guarantee convergence back to the long run path. Similar for x2,t .
31 / 58

Cointegrated VAR models, CIVAR

32 / 58

Model
We consider a VAR(p) with x t I(1), (unit root) nonstationary. x t = + 1 x t1 + . . . + p x tp + t Then x t is I(0). = (1) is singular, i.e. |(1)| = 0 (For weakly stationarity, I(0): |(z)| = 0 only for |z| > 1.) The VEC representation reads with = x t = + x t1 +
p1 i=1

x ti + t i

x t1 is called the error-correction term.

33 / 58

3 cases

We distinguish 3 cases for I. m = 0 : II. 0 < m < k : III. m = k : =0 = ,

Rank() = m: (km) , ( )(mk )

|| = | (1)| = 0!

34 / 58

I. Rank() = 0, m = 0 :
In case of Rank() = 0, i.e. m = 0, it follows = 0, the null matrix. There does not exist a linear combination of the I(1) vars, which is stationary. The xs are not cointegrated. The EC form reduces to a stationary VAR(p 1) in differences. x t = +
p1 i=1

x ti + t i

has m = 0 eigenvalues, which are not different from 0.

35 / 58

II. Rank() = m, 0 < m < k :

The rank of is m, m < k . We factorize in two rank m matrices and . Rank() = Rank() = m. Both and are (k m). = = 0 The VEC form is then x t = + x t1 +
p1 i=1

x ti + t i

The xs are integrated, I(1). There are m eigenvalues () = 0. The xs are cointegrated. There are m linear combinations, which are stationary.
36 / 58

II. Rank() = m, 0 < m < k :

There are m linear independent cointegrating (column) vectors in . The m stationary linear combinations are x t . x t has (k m) unit roots, so (k m) common stochastic trends. There are k I(1) variables, m cointegrating relations (eigenvalues of different from 0), and (k m) stochastic trends. k = m + (k m)
37 / 58

III. Rank() = m, m = k :

Full rank of implies that || = | (1)| = 0. x t has no unit root. That is x t is I(0). There are (k m) = 0 stochastic trends. As consequence we model the relationship of the xs in levels, not in differences. There is no need to refer to the error correction representation.

38 / 58

II. Rank() = m, 0 < m < k : (cont) common trends

A general way to obtain the (k m) common trends is to use the orthogonal complement matrix of . = 0 {k (k m)} {k m} = {(k m) m} If the ECM is left multiplied by the error correction term vanishes, = ( ) = 0(k m)k with x t = ( x t ) ( x t ) = ( ) +
p1 i=1

( x ti ) + ( t ) i

39 / 58

II. Rank() = m, 0 < m < k : (cont) common trends

The resulting system is a (k m) dimensional system of rst differences, corresponding to (k m) independent RWs x t which are the common trends. Example (from above): = (1, .5) then = (1, 2) .

40 / 58

Non uniqueness of , in

For any orthogonal matrix mm , = I, = = ()() = ( ) where both and are of rank m. Usually the structure
= [I mm , (1 )m(km) ]

is imposed. The rst m variables belong only to one equation and their coeffs are 1. Economic interpretation is helpful when structuring . Also, a reordering of the vars might be necessary.
41 / 58

Inclusion of deterministic functions

There are several possibilities to specify the deterministic part, , in the model. 1 = 0: All components of x t are I(1) without drift. The stationary series w t = x t has a zero mean. 2 = (0 )k 1 = km c 0,m1 : This is the special case of a restricted constant. The ECM is x t = ( x t1 + c 0 ) + . . . w t = x t has a mean of (c 0 ). There is only a constant in the cointegrating relation, but the xs are I(1) without a drift. 3 = 0 = 0: The xs are I(1) with drift. The coint rel may have a nonzero mean. Intercept 0 may be spilt in a drift component and a const vector in the coint eqs.
42 / 58

Inclusion of deterministic functions

4 = t = 0 + (c 1 )t: Analogous, 0 enters the drift of the xs. c 1 becomes the trend in the coint rel. x t = 0 + ( x t1 + c 1 t) + . . . 5 = t = 0 + 1 t: Both constant and slope of the trend are unrestricted. The trending behavior in the xs is determined both by a drift and a quadratic trend. The coint rel may have a linear trend. Case 3, = 0 , is relevant for asset prices. Remark: The assignment of the const to either intercept or coint rel is not unique.
43 / 58

ML estimation: Johansen (1)

Estimation is a 3-step procedure: 1st step: We start with the VEC representation and extract the effects of the lagged x tj from the lhs x t and from the rhs x t1 . (Cp. Frisch-Waugh). This gives the residuals u t for x t and v t for x t1 , and the model u t = v t + t 2nd step: All variables in the cointegration relation are dealt with symmetrically. There are no endogenous and no exogeneous variables. We view this system as ()1 u t = v t where and are (k k ). The solution is obtained by canonical correlation.
44 / 58

Johansen (2): canonical correlation

We determine vectors j , j so that the linear combinations u t and j v t j correlate
maximal for j = 1, maximal subjcet to orthogonality wrt the solution for j = 1 ( j = 2), etc.

For the largest correlation we get a largest eigenvalue, 1 , for the second largest a smaller one, 2 < 1 , etc. The eigenvalues are the squared (canonical) correlation coefcients. The columns of are the associated normalized eigenvectors. The s are not the eigenvalues of , but have the same zero/nonzero properties.
45 / 58

Johansen (2)
Actually we solve a generalized eigenvalue problem |S 11 S 10 S 1 S 01 | = 0 00 with the sample covariance matrices S00 = 1 t ut u , T p S11 = S01 = 1 t ut v T p

1 t vtv T p

The number of eigenvalues larger 0 determines the rank of , resp. , and so the number of cointegrating relations: 1 > . . . > m > 0 = . . . = 0 = k

46 / 58

Johansen (3)
3rd step: In this nal step the adjustment parameters and the s are estimated. x t = + x t1 +
p1 i=1

x ti + t i

The maximized likelihood function based on m cointegrating vectors is m 2/T Lmax |S00 | (1 i )
i=1

Under Gaussian innovations and the model is true, the estimates of the matrices are asy normal and asy j efcient.
Remark: S00 depends only on x t and x tj , j = 1, . . . , p.
47 / 58

Test for cointegration: rank test

Given the specication of the deterministic term we test for the rank m of . There are 2 sequential tests the rank test, and the maximum eigenvalue test. rank test: H0 : Rank() = m against HA : Rank() > m

The likelihood ratio statistic is LKtr (m) = (T p)

k i=m+1

ln(1 i )

We start with m = 0 that is Rank() = 0, there is no cointegration against m 1, that there is at least one coint rel. Etc.
48 / 58

Test for cointegration: rank test

LKtr (m) takes large values (i.e. H0 is rejected) when the sum of the remaining eigenvalues m+1 m+2 . . . k is large. If is large (say 1), then small (say 0), then ln(1 i ) is large. ln(1 i ) 0.

49 / 58

Test for cointegration: max eigenvalue statistic

maximum eigenvalue test: H0 : Rank() = m The statistic is LKmax (m) = (T p) ln(1 m+1 ) We start with m = 0 that is Rank() = 0, there is no cointegration against m = 1, that there is one coint rel. Etc. In case we reject m = k 1 coint rel, we should have to conclude that there are m = k coint rel. But this would not t to the assumption of I(1) vars. The critical values of both test statistics are nonstandard and are obtained via Monte Carlo simulation.
50 / 58

against

HA : Rank() = m + 1

Forecasting, summary
The tted ECM can be used for forecasting x t+ . The forecasts of x t+ ( -step ahead) are obtained recursively. x t+ = x t+ + x t+ 1 A summary: If all vars are stationary / the VAR is stable, the adequate model is a VAR in levels. If the vars are integrated of order 1 but not cointegrated, the adequate model is a VAR in rst differences (no level components included). If the vars are integrated and cointegrated, the adequate model is a cointegrated VAR. It is estimated in the rst differences with the cointegrating relations (the levels) as explanatory vars.
51 / 58

Bivariate cointegration

52 / 58

Estimation and testing: Engle and Granger

Engle-Granger: x t , yt I(1) yt = + x + ut t MacKinnon has tabulated critical values for the test of the LS residuals ut under the null of no cointegration (of a unit root), similar to the augmented Dickey-Fuller test. H0 : ut I(1), no coint HA : ut I(0), coint

The test distribution depends on the inclusion of an intercept or a trend. Additional lagged differences may be used. If u is stationary, xs and y are cointegrated.
53 / 58

Phillips-Ouliaris test
Phillips-Ouliaris: Two residuals are compared. ut from the Engle-Granger test and t from z t = z t1 + t estimated via LS, where z t = (yt , x ) . t t is stationary, ut only if the vars are cointegrated. 2 2 Intuitively the ratio (s /su ) is small under no coint and large under coint. H0 : no coint HA : coint

Two test statisticis Pu and Pz are available. (ca.po {urca} in R.)

Remark: If zt a RW, then zt = 1zt1 + t and t stationary.
54 / 58

Exercises and references

55 / 58

Exercises
Choose 2 of (1G, 2) and 1 out of (3G, 4G). 1G Use Ex4_SpurReg_R.txt to generate and comment the small sample distribution of the t-statistic and R 2 for (a) the spurious regression problem, (b) for the model in Xt and Yt . 2 Given a VAR(2) in standard form. Derive the VEC representation. Show the equivalence of both representations. Use x t = x t1 + x t . 3G Investigate the cointegration properties of stock indices. (a) For 2 stock exchanges, (b) for 3 or more stock exchanges. Use Ex4_R.txt.

56 / 58

Exercises

4G Investigate the price series of black and white pepper, PepperPrices from the R library(AER) wrt cointegration and give the VECM.

57 / 58

References

Tsay 8.5-6 Geyer 3.1 Johnson and Wichern: Multivariate Analysis, for canonical correlation Phillips and Ouliaris(1990): Asymptotic properties of residual based tests for cointegration, Econometrica 58, 165-193

58 / 58