1 Vector Error Correction Models and Cointegration It is quite possible for random walks to be related to each other so that

a regression of one random walk on the other has a stationary error term. For example let ∆xt = ε
∆yt = u

and let

be stationary. The simplest example is that yt = −xt + v. That is, let one random walk be the negative of the other – allowing for some error. Then the sum is simply a random error with no unit root or autocorrelation. If the combination of unit root variables is not unit root then there must be some relationship between them. This is an if and only if statement. (Green, p. 856) If you find cointegration then a relationship exists, if not it does not. Therefore if you are interested in establishing that a relationship exists between unit root variables, this is equivalent to establishing cointegration. That relationship is called the cointegrating vector, which for our example is (1,1) since the sum is stationary. There is a way to write a system that captures all the relationships and avoids unit roots. Consider ∆xt = α1 ( β1 yt −1 + β2 xt −1 ) + εt + vt ∆yt = α2 ( β1 yt −1 + β2 xt −1 ) + ut + vt This is called a vector error correction model. The error correction comes from the cointegrating relationship. The betas contain the cointegrating equation and the alphas the speeds of adjustment. If y and x are far from their equilibrium relationship, either y or x or both must change, the alphas let the data choose. The vector part of the name does not apply to the model above, but it will if the error terms are autocorrelated. Let εt = ρ ε −1 + w so that t xt = xt −1 + ρ ε −1 + w t xt − ρxt −1 = xt −1 − ρ( xt −2 + εt −1 ) + ρ ε −1 + w t xt −1 = xt −2 + εt −1 ) (Note that xt − xt −1 = ρ( xt −1 − xt −2 ) + w ∆xt = ρ∆xt −1 + w The third building block of vector error correction models consists of stationary variables. The variable x may be “a random walk with drift” signifying that ∆xt = d + εt where d is a constant. This would create a trend. The change in x could also be a function of a time variable, or its square. Indeed, we could insert any stationary variable. Application: PPP We will apply all of these ideas in a purchasing power parity model. Let p, p* and s be the logs of home prices, foreign prices and the spot rate (home currency per unit of foreign). We specify a model that is not particularly realistic but will allow all three separate building blocks to be traced easily:

yt + xt

2 ∆pt = d + ut ∆pt * = vt where vt = ρvt −1 + εt ∴∆pt* = ρ∆pt*−1 + εt ∆st = α( st −1 − pt −1 + pt*−1 ) + w Now use the help facility to print out p. Or. perhaps the exchange rate is fixed and the prices adjust. We will create the VEC that corresponds to the model above. which is a contradiction. Toward this end we rewrite our model in level form: pt = pt −1 + d + ut pt* = (1 + ρ) pt*−1 − ρpt*−2 + εt st = (1 +α) st −1 −αpt −1 +αpt*−1 + wt The VAR model is 0 0   pt −1  0 0  pt   1  p*  =  0 1 + ρ 0   pt*−1  + 0 − ρ  t      st   − α α 1 + α   st −1  0 0       The VEC is  ∆pt   0 0 0   pt −1  0 0  ∆p *  =  0 0 0   p *  +  0 − ρ  t    t −1    ∆st   − α α α   st −1  0 0       0  pt − 2  1  ut    p *  0  d +  ε  0  t − 2  +    t  wt  0  st − 2  0       0  ∆pt −1  1  ut    ∆p *  +  0 d +  ε  0   t −1     t  wt  0  ∆st −1  0       It is worthwhile checking that the 3x3 matrices in the VEC are related to the 3x3’s in the VAR as described. The example we constructed assumed cointegration and Π is clearly not of full rank. but it is instructive to construct the VAR model first. There can not be three cointegrating equations. Note that the matrix Π is the 3x3 in the VEC that premultiplies the vector of levels (lagged once). we assumed only one cointegrating vector that affected only one equation so: 0 0  0   0  0 Π= 0 0  =  0 [ −1 1 1] = αβ'    −α α α  α3      where the α3 is simply a reminder that we could have placed the cointegrating equation in each row of the system rather than imposing the assumption that the spot rate alone adjusts to any deviation from inequality. This matrix will always exist but it may be of full rank. It is not necessary.) If there were two cointegrating equations.) The Johansen Cointegration test checks the rank of Π . In our case. perhaps one of the countries is small and follows the lead of the large country. If there are then Π is of full rank. Π=α ' where α is the vector of speeds of adjustment and β ’ is the transpose of the cointegrating vector(s).0. (The cointegrating equation could be in both price equations. . α would be 3x2 and β ’ 2x3. (The determinant is zero. 523-524 of the users guide for Eviews 4. If Π is less than full rank it can β be decomposed in an interesting way.